From ffa90e04d31e96c577abe7e3327f72b463957d04 Mon Sep 17 00:00:00 2001 From: comar Date: Mon, 10 May 2021 16:20:19 +0900 Subject: [PATCH 01/25] fix: the exception not using default topk argument --- inference.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/inference.py b/inference.py index 89efb1fb..445dc4e5 100755 --- a/inference.py +++ b/inference.py @@ -119,8 +119,8 @@ def main(): with open(os.path.join(args.output_dir, './topk_ids.csv'), 'w') as out_file: filenames = loader.dataset.filenames(basename=True) for filename, label in zip(filenames, topk_ids): - out_file.write('{0},{1},{2},{3},{4},{5}\n'.format( - filename, label[0], label[1], label[2], label[3], label[4])) + out_file.write('{0},{1}\n'.format( + filename, ','.join([ str(v) for v in label]))) if __name__ == '__main__': From d7e1e7144a62f0368787c081eb999d27cb97d159 Mon Sep 17 00:00:00 2001 From: comar Date: Mon, 10 May 2021 16:52:40 +0900 Subject: [PATCH 02/25] fix: the exeption when topk is 1 --- inference.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/inference.py b/inference.py index 445dc4e5..5fcf1e60 100755 --- a/inference.py +++ b/inference.py @@ -114,7 +114,7 @@ def main(): _logger.info('Predict: [{0}/{1}] Time {batch_time.val:.3f} ({batch_time.avg:.3f})'.format( batch_idx, len(loader), batch_time=batch_time)) - topk_ids = np.concatenate(topk_ids, axis=0).squeeze() + topk_ids = np.concatenate(topk_ids, axis=0) with open(os.path.join(args.output_dir, './topk_ids.csv'), 'w') as out_file: filenames = loader.dataset.filenames(basename=True) From 2a72d38ba2544f9a73b2d142d035b19f8515845c Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 14 May 2021 23:06:49 -0700 Subject: [PATCH 03/25] Update README.md --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 4cd599f1..cbcfd00f 100644 --- a/README.md +++ b/README.md @@ -28,7 +28,7 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor * 1k trained variants: `tf_efficientnetv2_s/m/l` * 21k trained variants: `tf_efficientnetv2_s/m/l_21k` * 21k pretrained -> 1k fine-tuned: `tf_efficientnetv2_s/m/l_21ft1k` - * v2 models w/ v1 scaling: `tf_efficientnet_v2_b0` through `b3` + * v2 models w/ v1 scaling: `tf_efficientnetv2_b0` through `b3` * Rename my prev V2 guess `efficientnet_v2s` -> `efficientnetv2_rw_s` * Some blank `efficientnetv2_*` models in-place for future native PyTorch training From 7077f16c6a72960dd5a2ca042ed79c8e72eeb765 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Sat, 15 May 2021 12:42:26 -0700 Subject: [PATCH 04/25] Change 21k model naming from _21k to _in21k for consistency with existing 21k models. --- README.md | 4 ++-- timm/models/efficientnet.py | 36 ++++++++++++++++++------------------ 2 files changed, 20 insertions(+), 20 deletions(-) diff --git a/README.md b/README.md index cbcfd00f..ca283605 100644 --- a/README.md +++ b/README.md @@ -26,8 +26,8 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor ### May 14, 2021 * Add EfficientNet-V2 official model defs w/ ported weights from official [Tensorflow/Keras](https://github.com/google/automl/tree/master/efficientnetv2) impl. * 1k trained variants: `tf_efficientnetv2_s/m/l` - * 21k trained variants: `tf_efficientnetv2_s/m/l_21k` - * 21k pretrained -> 1k fine-tuned: `tf_efficientnetv2_s/m/l_21ft1k` + * 21k trained variants: `tf_efficientnetv2_s/m/l_in21k` + * 21k pretrained -> 1k fine-tuned: `tf_efficientnetv2_s/m/l_in21ft1k` * v2 models w/ v1 scaling: `tf_efficientnetv2_b0` through `b3` * Rename my prev V2 guess `efficientnet_v2s` -> `efficientnetv2_rw_s` * Some blank `efficientnetv2_*` models in-place for future native PyTorch training diff --git a/timm/models/efficientnet.py b/timm/models/efficientnet.py index 1716e92d..0c0464b5 100644 --- a/timm/models/efficientnet.py +++ b/timm/models/efficientnet.py @@ -332,28 +332,28 @@ default_cfgs = { mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), - 'tf_efficientnetv2_s_21ft1k': _cfg( + 'tf_efficientnetv2_s_in21ft1k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_s_21ft1k-d7dafa41.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), input_size=(3, 300, 300), test_input_size=(3, 384, 384), pool_size=(10, 10), crop_pct=1.0), - 'tf_efficientnetv2_m_21ft1k': _cfg( + 'tf_efficientnetv2_m_in21ft1k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_m_21ft1k-bf41664a.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), - 'tf_efficientnetv2_l_21ft1k': _cfg( + 'tf_efficientnetv2_l_in21ft1k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_l_21ft1k-60127a9d.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), - 'tf_efficientnetv2_s_21k': _cfg( + 'tf_efficientnetv2_s_in21k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_s_21k-6337ad01.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), num_classes=21843, input_size=(3, 300, 300), test_input_size=(3, 384, 384), pool_size=(10, 10), crop_pct=1.0), - 'tf_efficientnetv2_m_21k': _cfg( + 'tf_efficientnetv2_m_in21k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_m_21k-361418a2.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), num_classes=21843, input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), - 'tf_efficientnetv2_l_21k': _cfg( + 'tf_efficientnetv2_l_in21k': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-effv2-weights/tf_efficientnetv2_l_21k-91a19ec9.pth', mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), num_classes=21843, input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), @@ -1929,62 +1929,62 @@ def tf_efficientnetv2_l(pretrained=False, **kwargs): @register_model -def tf_efficientnetv2_s_21ft1k(pretrained=False, **kwargs): +def tf_efficientnetv2_s_in21ft1k(pretrained=False, **kwargs): """ EfficientNet-V2 Small. Pretrained on ImageNet-21k, fine-tuned on 1k. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_s('tf_efficientnetv2_s_21ft1k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_s('tf_efficientnetv2_s_in21ft1k', pretrained=pretrained, **kwargs) return model @register_model -def tf_efficientnetv2_m_21ft1k(pretrained=False, **kwargs): +def tf_efficientnetv2_m_in21ft1k(pretrained=False, **kwargs): """ EfficientNet-V2 Medium. Pretrained on ImageNet-21k, fine-tuned on 1k. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_m('tf_efficientnetv2_m_21ft1k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_m('tf_efficientnetv2_m_in21ft1k', pretrained=pretrained, **kwargs) return model @register_model -def tf_efficientnetv2_l_21ft1k(pretrained=False, **kwargs): +def tf_efficientnetv2_l_in21ft1k(pretrained=False, **kwargs): """ EfficientNet-V2 Large. Pretrained on ImageNet-21k, fine-tuned on 1k. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_l('tf_efficientnetv2_l_21ft1k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_l('tf_efficientnetv2_l_in21ft1k', pretrained=pretrained, **kwargs) return model @register_model -def tf_efficientnetv2_s_21k(pretrained=False, **kwargs): +def tf_efficientnetv2_s_in21k(pretrained=False, **kwargs): """ EfficientNet-V2 Small w/ ImageNet-21k pretrained weights. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_s('tf_efficientnetv2_s_21k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_s('tf_efficientnetv2_s_in21k', pretrained=pretrained, **kwargs) return model @register_model -def tf_efficientnetv2_m_21k(pretrained=False, **kwargs): +def tf_efficientnetv2_m_in21k(pretrained=False, **kwargs): """ EfficientNet-V2 Medium w/ ImageNet-21k pretrained weights. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_m('tf_efficientnetv2_m_21k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_m('tf_efficientnetv2_m_in21k', pretrained=pretrained, **kwargs) return model @register_model -def tf_efficientnetv2_l_21k(pretrained=False, **kwargs): +def tf_efficientnetv2_l_in21k(pretrained=False, **kwargs): """ EfficientNet-V2 Large w/ ImageNet-21k pretrained weights. Tensorflow compatible variant """ kwargs['bn_eps'] = BN_EPS_TF_DEFAULT kwargs['pad_type'] = 'same' - model = _gen_efficientnetv2_l('tf_efficientnetv2_l_21k', pretrained=pretrained, **kwargs) + model = _gen_efficientnetv2_l('tf_efficientnetv2_l_in21k', pretrained=pretrained, **kwargs) return model From e7f0db866412b9ae61332c205270c9fc0ef5083c Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Sun, 16 May 2021 08:31:52 -0700 Subject: [PATCH 05/25] Fix drop/drop_path arg on MLP-Mixer model. Fix #641 --- timm/models/mlp_mixer.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/timm/models/mlp_mixer.py b/timm/models/mlp_mixer.py index c2c96e6c..248568fc 100644 --- a/timm/models/mlp_mixer.py +++ b/timm/models/mlp_mixer.py @@ -96,8 +96,8 @@ class MlpMixer(nn.Module): mlp_layer=Mlp, norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU, - drop=0., - drop_path=0., + drop_rate=0., + drop_path_rate=0., nlhb=False, ): super().__init__() @@ -108,7 +108,7 @@ class MlpMixer(nn.Module): self.blocks = nn.Sequential(*[ MixerBlock( hidden_dim, self.stem.num_patches, tokens_dim, channels_dim, - mlp_layer=mlp_layer, norm_layer=norm_layer, act_layer=act_layer, drop=drop, drop_path=drop_path) + mlp_layer=mlp_layer, norm_layer=norm_layer, act_layer=act_layer, drop=drop_rate, drop_path=drop_path_rate) for _ in range(num_blocks)]) self.norm = norm_layer(hidden_dim) self.head = nn.Linear(hidden_dim, self.num_classes) # zero init From 00548b8427739bf9954dd8ce522e2833de616baf Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E9=91=AB=E6=9D=B0?= Date: Tue, 18 May 2021 19:21:53 +0800 Subject: [PATCH 06/25] Add Twins --- timm/models/__init__.py | 1 + timm/models/twins.py | 625 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 626 insertions(+) create mode 100644 timm/models/twins.py diff --git a/timm/models/__init__.py b/timm/models/__init__.py index 46ea155f..293b459d 100644 --- a/timm/models/__init__.py +++ b/timm/models/__init__.py @@ -39,6 +39,7 @@ from .vision_transformer_hybrid import * from .vovnet import * from .xception import * from .xception_aligned import * +from .twins import * from .factory import create_model, split_model_name, safe_model_name from .helpers import load_checkpoint, resume_checkpoint, model_parameters diff --git a/timm/models/twins.py b/timm/models/twins.py new file mode 100644 index 00000000..27be4cba --- /dev/null +++ b/timm/models/twins.py @@ -0,0 +1,625 @@ +""" Twins +A PyTorch impl of : `Twins: Revisiting the Design of Spatial Attention in Vision Transformers` + - https://arxiv.org/pdf/2104.13840.pdf + +Code/weights from https://github.com/Meituan-AutoML/Twins, original copyright/license info below + +""" +# -------------------------------------------------------- +# Twins +# Copyright (c) 2021 Meituan +# Licensed under The Apache 2.0 License [see LICENSE for details] +# Written by Xinjie Li, Xiangxiang Chu +# -------------------------------------------------------- + +import logging +import math +from copy import deepcopy +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from functools import partial + +from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD +from .layers import Mlp, DropPath, to_2tuple, trunc_normal_ +from .registry import register_model +from .vision_transformer import _cfg +from .vision_transformer import Block as TimmBlock +from .vision_transformer import Attention as TimmAttention +from .helpers import build_model_with_cfg, overlay_external_default_cfg +from .vision_transformer import checkpoint_filter_fn, _init_vit_weights + +_logger = logging.getLogger(__name__) + +def _cfg(url='', **kwargs): + return { + 'url': url, + 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None, + 'crop_pct': .9, 'interpolation': 'bicubic', 'fixed_input_size': True, + 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, + 'first_conv': 'patch_embed.proj', 'classifier': 'head', + **kwargs + } + +default_cfgs = { + 'twins_pcpvt_small': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_small.pth', + ), + 'twins_pcpvt_base': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_base.pth', + ), + 'twins_pcpvt_large': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_large.pth', + ), + 'twins_svt_small': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_small.pth', + ), + 'twins_svt_base': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_base.pth', + ), + 'twins_svt_large': _cfg( + url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_large.pth', + ), +} + + + +class GroupAttention(nn.Module): + """ + LSA: self attention within a group + """ + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., ws=1): + assert ws != 1 + super(GroupAttention, self).__init__() + assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." + + self.dim = dim + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + self.ws = ws + + def forward(self, x, H, W): + """ + There are two implementations for this function, zero padding or mask. We don't observe obvious difference for + both. You can choose any one, we recommend forward_padding because it's neat. However, + the masking implementation is more reasonable and accurate. + Args: + x: + H: + W: + + Returns: + + """ + return self.forward_padding(x, H, W) + + def forward_mask(self, x, H, W): + B, N, C = x.shape + x = x.view(B, H, W, C) + pad_l = pad_t = 0 + pad_r = (self.ws - W % self.ws) % self.ws + pad_b = (self.ws - H % self.ws) % self.ws + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + _h, _w = Hp // self.ws, Wp // self.ws + mask = torch.zeros((1, Hp, Wp), device=x.device) + mask[:, -pad_b:, :].fill_(1) + mask[:, :, -pad_r:].fill_(1) + + x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) # B, _h, _w, ws, ws, C + mask = mask.reshape(1, _h, self.ws, _w, self.ws).transpose(2, 3).reshape(1, _h*_w, self.ws*self.ws) + attn_mask = mask.unsqueeze(2) - mask.unsqueeze(3) # 1, _h*_w, ws*ws, ws*ws + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-1000.0)).masked_fill(attn_mask == 0, float(0.0)) + qkv = self.qkv(x).reshape(B, _h * _w, self.ws * self.ws, 3, self.num_heads, + C // self.num_heads).permute(3, 0, 1, 4, 2, 5) # n_h, B, _w*_h, nhead, ws*ws, dim + q, k, v = qkv[0], qkv[1], qkv[2] # B, _h*_w, n_head, ws*ws, dim_head + attn = (q @ k.transpose(-2, -1)) * self.scale # B, _h*_w, n_head, ws*ws, ws*ws + attn = attn + attn_mask.unsqueeze(2) + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) # attn @v -> B, _h*_w, n_head, ws*ws, dim_head + attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C) + x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C) + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + x = x.reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + def forward_padding(self, x, H, W): + B, N, C = x.shape + x = x.view(B, H, W, C) + pad_l = pad_t = 0 + pad_r = (self.ws - W % self.ws) % self.ws + pad_b = (self.ws - H % self.ws) % self.ws + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + _h, _w = Hp // self.ws, Wp // self.ws + x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) + qkv = self.qkv(x).reshape(B, _h * _w, self.ws * self.ws, 3, self.num_heads, + C // self.num_heads).permute(3, 0, 1, 4, 2, 5) + q, k, v = qkv[0], qkv[1], qkv[2] + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C) + x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C) + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + x = x.reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class Attention(nn.Module): + """ + GSA: using a key to summarize the information for a group to be efficient. + """ + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1): + super().__init__() + assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." + + self.dim = dim + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.q = nn.Linear(dim, dim, bias=qkv_bias) + self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + self.sr_ratio = sr_ratio + if sr_ratio > 1: + self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio) + self.norm = nn.LayerNorm(dim) + + def forward(self, x, H, W): + B, N, C = x.shape + q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + + if self.sr_ratio > 1: + x_ = x.permute(0, 2, 1).reshape(B, C, H, W) + x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1) + x_ = self.norm(x_) + kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + else: + kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + k, v = kv[0], kv[1] + + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + + return x + + +class Block(nn.Module): + + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, + attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio) + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + def forward(self, x, H, W): + x = x + self.drop_path(self.attn(self.norm1(x), H, W)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + +class SBlock(TimmBlock): + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1): + super(SBlock, self).__init__(dim, num_heads, mlp_ratio, qkv_bias, qk_scale, drop, attn_drop, + drop_path, act_layer, norm_layer) + + def forward(self, x, H, W): + return super(SBlock, self).forward(x) + + +class GroupBlock(TimmBlock): + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1, ws=1): + super(GroupBlock, self).__init__(dim, num_heads, mlp_ratio, qkv_bias, qk_scale, drop, attn_drop, + drop_path, act_layer, norm_layer) + del self.attn + if ws == 1: + self.attn = Attention(dim, num_heads, qkv_bias, qk_scale, attn_drop, drop, sr_ratio) + else: + self.attn = GroupAttention(dim, num_heads, qkv_bias, qk_scale, attn_drop, drop, ws) + + def forward(self, x, H, W): + x = x + self.drop_path(self.attn(self.norm1(x), H, W)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + """ + + def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): + super().__init__() + # img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + + self.img_size = img_size + self.patch_size = patch_size + assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \ + f"img_size {img_size} should be divided by patch_size {patch_size}." + self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1] + self.num_patches = self.H * self.W + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + self.norm = nn.LayerNorm(embed_dim) + + def forward(self, x): + B, C, H, W = x.shape + + x = self.proj(x).flatten(2).transpose(1, 2) + x = self.norm(x) + H, W = H // self.patch_size[0], W // self.patch_size[1] + + return x, (H, W) + + +# borrow from PVT https://github.com/whai362/PVT.git +class PyramidVisionTransformer(nn.Module): + def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512], + num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., + attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, + depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], block_cls=Block): + super().__init__() + self.num_classes = num_classes + self.depths = depths + + # patch_embed + self.patch_embeds = nn.ModuleList() + self.pos_embeds = nn.ParameterList() + self.pos_drops = nn.ModuleList() + self.blocks = nn.ModuleList() + + for i in range(len(depths)): + if i == 0: + self.patch_embeds.append(PatchEmbed(img_size, patch_size, in_chans, embed_dims[i])) + else: + self.patch_embeds.append( + # PatchEmbed(img_size // patch_size // 2 ** (i - 1), 2, embed_dims[i - 1], embed_dims[i]) + PatchEmbed((img_size[0] // patch_size // 2**(i-1),img_size[1] // patch_size // 2**(i-1)), 2, embed_dims[i - 1], embed_dims[i]) + ) + patch_num = self.patch_embeds[-1].num_patches + 1 if i == len(embed_dims) - 1 else self.patch_embeds[ + -1].num_patches + self.pos_embeds.append(nn.Parameter(torch.zeros(1, patch_num, embed_dims[i]))) + self.pos_drops.append(nn.Dropout(p=drop_rate)) + + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + cur = 0 + for k in range(len(depths)): + _block = nn.ModuleList([block_cls( + dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, + sr_ratio=sr_ratios[k]) + for i in range(depths[k])]) + self.blocks.append(_block) + cur += depths[k] + + self.norm = norm_layer(embed_dims[-1]) + + # cls_token + self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dims[-1])) + + # classification head + self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity() + + # init weights + for pos_emb in self.pos_embeds: + trunc_normal_(pos_emb, std=.02) + self.apply(self._init_weights) + + def reset_drop_path(self, drop_path_rate): + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))] + cur = 0 + for k in range(len(self.depths)): + for i in range(self.depths[k]): + self.blocks[k][i].drop_path.drop_prob = dpr[cur + i] + cur += self.depths[k] + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'cls_token'} + + def get_classifier(self): + return self.head + + def reset_classifier(self, num_classes, global_pool=''): + self.num_classes = num_classes + self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity() + + def forward_features(self, x): + B = x.shape[0] + for i in range(len(self.depths)): + x, (H, W) = self.patch_embeds[i](x) + if i == len(self.depths) - 1: + cls_tokens = self.cls_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, x), dim=1) + x = x + self.pos_embeds[i] + x = self.pos_drops[i](x) + for blk in self.blocks[i]: + x = blk(x, H, W) + if i < len(self.depths) - 1: + x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() + + x = self.norm(x) + + return x[:, 0] + + def forward(self, x): + x = self.forward_features(x) + x = self.head(x) + + return x + + +# PEG from https://arxiv.org/abs/2102.10882 +class PosCNN(nn.Module): + def __init__(self, in_chans, embed_dim=768, s=1): + super(PosCNN, self).__init__() + self.proj = nn.Sequential(nn.Conv2d(in_chans, embed_dim, 3, s, 1, bias=True, groups=embed_dim), ) + self.s = s + + def forward(self, x, H, W): + B, N, C = x.shape + feat_token = x + cnn_feat = feat_token.transpose(1, 2).view(B, C, H, W) + if self.s == 1: + x = self.proj(cnn_feat) + cnn_feat + else: + x = self.proj(cnn_feat) + x = x.flatten(2).transpose(1, 2) + return x + + def no_weight_decay(self): + return ['proj.%d.weight' % i for i in range(4)] + + +class CPVTV2(PyramidVisionTransformer): + """ + Use useful results from CPVT. PEG and GAP. + Therefore, cls token is no longer required. + PEG is used to encode the absolute position on the fly, which greatly affects the performance when input resolution + changes during the training (such as segmentation, detection) + """ + def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512], + num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., + attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, + depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], block_cls=Block): + super(CPVTV2, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, mlp_ratios, + qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, norm_layer, depths, + sr_ratios, block_cls) + del self.pos_embeds + del self.cls_token + self.pos_block = nn.ModuleList( + [PosCNN(embed_dim, embed_dim) for embed_dim in embed_dims] + ) + self.apply(self._init_weights) + + def _init_weights(self, m): + import math + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + elif isinstance(m, nn.Conv2d): + fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + fan_out //= m.groups + m.weight.data.normal_(0, math.sqrt(2.0 / fan_out)) + if m.bias is not None: + m.bias.data.zero_() + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1.0) + m.bias.data.zero_() + + def no_weight_decay(self): + return set(['cls_token'] + ['pos_block.' + n for n, p in self.pos_block.named_parameters()]) + + def forward_features(self, x): + B = x.shape[0] + + for i in range(len(self.depths)): + x, (H, W) = self.patch_embeds[i](x) + x = self.pos_drops[i](x) + for j, blk in enumerate(self.blocks[i]): + x = blk(x, H, W) + if j == 0: + x = self.pos_block[i](x, H, W) # PEG here + if i < len(self.depths) - 1: + x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() + + x = self.norm(x) + + return x.mean(dim=1) # GAP here + + +class Twins_PCPVT(CPVTV2): + def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256], + num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., + attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, + depths=[4, 4, 4], sr_ratios=[4, 2, 1], block_cls=SBlock): + super(Twins_PCPVT, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, + mlp_ratios, qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, + norm_layer, depths, sr_ratios, block_cls) + + +class Twins_SVT(Twins_PCPVT): + """ + alias Twins-SVT + """ + def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256], + num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., + attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, + depths=[4, 4, 4], sr_ratios=[4, 2, 1], block_cls=GroupBlock, wss=[7, 7, 7]): + super(Twins_SVT, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, + mlp_ratios, qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, + norm_layer, depths, sr_ratios, block_cls) + del self.blocks + self.wss = wss + # transformer encoder + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + cur = 0 + self.blocks = nn.ModuleList() + for k in range(len(depths)): + _block = nn.ModuleList([block_cls( + dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, + sr_ratio=sr_ratios[k], ws=1 if i % 2 == 1 else wss[k]) for i in range(depths[k])]) + self.blocks.append(_block) + cur += depths[k] + self.apply(self._init_weights) + + +def _conv_filter(state_dict, patch_size=16): + """ convert patch embedding weight from manual patchify + linear proj to conv""" + out_dict = {} + for k, v in state_dict.items(): + if 'patch_embed.proj.weight' in k: + v = v.reshape((v.shape[0], 3, patch_size, patch_size)) + out_dict[k] = v + + return out_dict + +def _create_twins_svt(variant, pretrained=False, default_cfg=None, **kwargs): + if default_cfg is None: + default_cfg = deepcopy(default_cfgs[variant]) + overlay_external_default_cfg(default_cfg, kwargs) + default_num_classes = default_cfg['num_classes'] + default_img_size = default_cfg['input_size'][-2:] + + num_classes = kwargs.pop('num_classes', default_num_classes) + img_size = kwargs.pop('img_size', default_img_size) + if kwargs.get('features_only', None): + raise RuntimeError('features_only not implemented for Vision Transformer models.') + + model = build_model_with_cfg( + Twins_SVT, variant, pretrained, + default_cfg=default_cfg, + img_size=img_size, + num_classes=num_classes, + pretrained_filter_fn=checkpoint_filter_fn, + **kwargs) + + return model + +def _create_twins_pcpvt(variant, pretrained=False, default_cfg=None, **kwargs): + if default_cfg is None: + default_cfg = deepcopy(default_cfgs[variant]) + overlay_external_default_cfg(default_cfg, kwargs) + default_num_classes = default_cfg['num_classes'] + default_img_size = default_cfg['input_size'][-2:] + + num_classes = kwargs.pop('num_classes', default_num_classes) + img_size = kwargs.pop('img_size', default_img_size) + if kwargs.get('features_only', None): + raise RuntimeError('features_only not implemented for Vision Transformer models.') + + model = build_model_with_cfg( + CPVTV2, variant, pretrained, + default_cfg=default_cfg, + img_size=img_size, + num_classes=num_classes, + pretrained_filter_fn=checkpoint_filter_fn, + **kwargs) + + return model + + +@register_model +def twins_pcpvt_small(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], + **kwargs) + return _create_twins_pcpvt('twins_pcpvt_small', pretrained=pretrained, **model_kwargs) + + +@register_model +def twins_pcpvt_base(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1], + **kwargs) + return _create_twins_pcpvt('twins_pcpvt_base', pretrained=pretrained, **model_kwargs) + + +@register_model +def twins_pcpvt_large(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], + **kwargs) + return _create_twins_pcpvt('twins_pcpvt_large', pretrained=pretrained, **model_kwargs) + + +@register_model +def twins_svt_small(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[64, 128, 256, 512], num_heads=[2, 4, 8, 16], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 10, 4], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], + **kwargs) + return _create_twins_svt('twins_svt_small', pretrained=pretrained, **model_kwargs) + + +@register_model +def twins_svt_base(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[96, 192, 384, 768], num_heads=[3, 6, 12, 24], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], + **kwargs) + + return _create_twins_svt('twins_svt_base', pretrained=pretrained, **model_kwargs) + + +@register_model +def twins_svt_large(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=4, embed_dims=[128, 256, 512, 1024], num_heads=[4, 8, 16, 32], mlp_ratios=[4, 4, 4, 4], + qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], + **kwargs) + + return _create_twins_svt('twins_svt_large', pretrained=pretrained, **model_kwargs) From d5af752117e5b8a110dde57b191b7d3080a62fcd Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Wed, 19 May 2021 09:55:05 -0700 Subject: [PATCH 07/25] Add preliminary gMLP and ResMLP impl to Mlp-Mixer --- timm/models/layers/__init__.py | 2 +- timm/models/layers/mlp.py | 34 ++++- timm/models/mlp_mixer.py | 266 +++++++++++++++++++++++++++++---- 3 files changed, 267 insertions(+), 35 deletions(-) diff --git a/timm/models/layers/__init__.py b/timm/models/layers/__init__.py index 90241f5c..4aae99e3 100644 --- a/timm/models/layers/__init__.py +++ b/timm/models/layers/__init__.py @@ -20,7 +20,7 @@ from .helpers import to_ntuple, to_2tuple, to_3tuple, to_4tuple, make_divisible from .inplace_abn import InplaceAbn from .linear import Linear from .mixed_conv2d import MixedConv2d -from .mlp import Mlp, GluMlp +from .mlp import Mlp, GluMlp, GatedMlp from .norm import GroupNorm from .norm_act import BatchNormAct2d, GroupNormAct from .padding import get_padding, get_same_padding, pad_same diff --git a/timm/models/layers/mlp.py b/timm/models/layers/mlp.py index b65c8d07..b3f8de11 100644 --- a/timm/models/layers/mlp.py +++ b/timm/models/layers/mlp.py @@ -34,9 +34,10 @@ class GluMlp(nn.Module): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features * 2) + assert hidden_features % 2 == 0 + self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) + self.fc2 = nn.Linear(hidden_features // 2, out_features) self.drop = nn.Dropout(drop) def forward(self, x): @@ -47,3 +48,32 @@ class GluMlp(nn.Module): x = self.fc2(x) x = self.drop(x) return x + + +class GatedMlp(nn.Module): + """ MLP as used in gMLP + """ + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, + gate_layer=None, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + if gate_layer is not None: + assert hidden_features % 2 == 0 + self.gate = gate_layer(hidden_features) + hidden_features = hidden_features // 2 # FIXME base reduction on gate property? + else: + self.gate = nn.Identity() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.gate(x) + x = self.fc2(x) + x = self.drop(x) + return x diff --git a/timm/models/mlp_mixer.py b/timm/models/mlp_mixer.py index 248568fc..2241fe43 100644 --- a/timm/models/mlp_mixer.py +++ b/timm/models/mlp_mixer.py @@ -1,4 +1,6 @@ -""" MLP-Mixer in PyTorch +""" MLP-Mixer, ResMLP, and gMLP in PyTorch + +This impl originally based on MLP-Mixer paper. Official JAX impl: https://github.com/google-research/vision_transformer/blob/linen/vit_jax/models_mixer.py @@ -12,6 +14,25 @@ Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2 year={2021} } +Also supporting preliminary (not verified) implementations of ResMlp, gMLP, and possibly more... + +Paper: `ResMLP: Feedforward networks for image classification...` - https://arxiv.org/abs/2105.03404 +@misc{touvron2021resmlp, + title={ResMLP: Feedforward networks for image classification with data-efficient training}, + author={Hugo Touvron and Piotr Bojanowski and Mathilde Caron and Matthieu Cord and Alaaeldin El-Nouby and + Edouard Grave and Armand Joulin and Gabriel Synnaeve and Jakob Verbeek and Hervé Jégou}, + year={2021}, + eprint={2105.03404}, +} + +Paper: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 +@misc{liu2021pay, + title={Pay Attention to MLPs}, + author={Hanxiao Liu and Zihang Dai and David R. So and Quoc V. Le}, + year={2021}, + eprint={2105.08050}, +} + A thank you to paper authors for releasing code and weights. Hacked together by / Copyright 2021 Ross Wightman @@ -25,7 +46,7 @@ import torch.nn as nn from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .helpers import build_model_with_cfg, overlay_external_default_cfg -from .layers import PatchEmbed, Mlp, GluMlp, DropPath, lecun_normal_ +from .layers import PatchEmbed, Mlp, GluMlp, GatedMlp, DropPath, lecun_normal_, to_2tuple from .registry import register_model @@ -43,7 +64,6 @@ def _cfg(url='', **kwargs): default_cfgs = dict( mixer_s32_224=_cfg(), mixer_s16_224=_cfg(), - mixer_s16_glu_224=_cfg(), mixer_b32_224=_cfg(), mixer_b16_224=_cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_mixer_b16_224-76587d61.pth', @@ -60,15 +80,29 @@ default_cfgs = dict( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_mixer_l16_224_in21k-846aa33c.pth', num_classes=21843 ), + + gmixer_12_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), + gmixer_24_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), + + resmlp_12_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), + resmlp_24_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), + resmlp_36_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), + + gmlp_ti16_224=_cfg(), + gmlp_s16_224=_cfg(), + gmlp_b16_224=_cfg(), ) class MixerBlock(nn.Module): - + """ Residual Block w/ token mixing and channel MLPs + Based on: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 + """ def __init__( - self, dim, seq_len, tokens_dim, channels_dim, - mlp_layer=Mlp, norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU, drop=0., drop_path=0.): + self, dim, seq_len, mlp_ratio=(0.5, 4.0), mlp_layer=Mlp, + norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU, drop=0., drop_path=0.): super().__init__() + tokens_dim, channels_dim = [int(x * dim) for x in to_2tuple(mlp_ratio)] self.norm1 = norm_layer(dim) self.mlp_tokens = mlp_layer(seq_len, tokens_dim, act_layer=act_layer, drop=drop) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() @@ -81,6 +115,78 @@ class MixerBlock(nn.Module): return x +class Affine(nn.Module): + def __init__(self, dim): + super().__init__() + self.alpha = nn.Parameter(torch.ones((1, 1, dim))) + self.beta = nn.Parameter(torch.zeros((1, 1, dim))) + + def forward(self, x): + return torch.addcmul(self.beta, self.alpha, x) + + +class ResBlock(nn.Module): + """ Residual MLP block w/ LayerScale and Affine 'norm' + + Based on: `ResMLP: Feedforward networks for image classification...` - https://arxiv.org/abs/2105.03404 + """ + def __init__( + self, dim, seq_len, mlp_ratio=4, mlp_layer=Mlp, norm_layer=Affine, + act_layer=nn.GELU, init_values=1e-4, drop=0., drop_path=0.): + super().__init__() + channel_dim = int(dim * mlp_ratio) + self.norm1 = norm_layer(dim) + self.linear_tokens = nn.Linear(seq_len, seq_len) + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + self.mlp_channels = mlp_layer(dim, channel_dim, act_layer=act_layer, drop=drop) + self.ls1 = nn.Parameter(init_values * torch.ones(dim)) + self.ls2 = nn.Parameter(init_values * torch.ones(dim)) + + def forward(self, x): + x = x + self.drop_path(self.ls1 * self.linear_tokens(self.norm1(x).transpose(1, 2)).transpose(1, 2)) + x = x + self.drop_path(self.ls2 * self.mlp_channels(self.norm2(x))) + return x + + +class SpatialGatingUnit(nn.Module): + """ Spatial Gating Unit + + Based on: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 + """ + def __init__(self, dim, seq_len, norm_layer=nn.LayerNorm): + super().__init__() + gate_dim = dim // 2 + self.norm = norm_layer(gate_dim) + self.proj = nn.Linear(seq_len, seq_len) + + def forward(self, x): + u, v = x.chunk(2, dim=-1) + v = self.norm(v) + v = self.proj(v.transpose(-1, -2)) + return u * v.transpose(-1, -2) + + +class SpatialGatingBlock(nn.Module): + """ Residual Block w/ Spatial Gating + + Based on: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 + """ + def __init__( + self, dim, seq_len, mlp_ratio=4, mlp_layer=GatedMlp, + norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU, drop=0., drop_path=0.): + super().__init__() + channel_dim = int(dim * mlp_ratio) + self.norm = norm_layer(dim) + sgu = partial(SpatialGatingUnit, seq_len=seq_len) + self.mlp_channels = mlp_layer(dim, channel_dim, act_layer=act_layer, gate_layer=sgu, drop=drop) + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + + def forward(self, x): + x = x + self.drop_path(self.mlp_channels(self.norm(x))) + return x + + class MlpMixer(nn.Module): def __init__( @@ -91,24 +197,27 @@ class MlpMixer(nn.Module): patch_size=16, num_blocks=8, hidden_dim=512, - tokens_dim=256, - channels_dim=2048, + mlp_ratio=(0.5, 4.0), + block_layer=MixerBlock, mlp_layer=Mlp, norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU, drop_rate=0., drop_path_rate=0., nlhb=False, + stem_norm=False, ): super().__init__() self.num_classes = num_classes - self.stem = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=hidden_dim) - # FIXME drop_path (stochastic depth scaling rule?) + self.stem = PatchEmbed( + img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=hidden_dim, + norm_layer=norm_layer if stem_norm else None) + # FIXME drop_path (stochastic depth scaling rule or all the same?) self.blocks = nn.Sequential(*[ - MixerBlock( - hidden_dim, self.stem.num_patches, tokens_dim, channels_dim, - mlp_layer=mlp_layer, norm_layer=norm_layer, act_layer=act_layer, drop=drop_rate, drop_path=drop_path_rate) + block_layer( + hidden_dim, self.stem.num_patches, mlp_ratio, mlp_layer=mlp_layer, norm_layer=norm_layer, + act_layer=act_layer, drop=drop_rate, drop_path=drop_path_rate) for _ in range(num_blocks)]) self.norm = norm_layer(hidden_dim) self.head = nn.Linear(hidden_dim, self.num_classes) # zero init @@ -136,6 +245,9 @@ def _init_weights(m, n: str, head_bias: float = 0.): if n.startswith('head'): nn.init.zeros_(m.weight) nn.init.constant_(m.bias, head_bias) + elif n.endswith('gate.proj'): + nn.init.normal_(m.weight, std=1e-4) + nn.init.ones_(m.bias) else: nn.init.xavier_uniform_(m.weight) if m.bias is not None: @@ -177,8 +289,9 @@ def _create_mixer(variant, pretrained=False, default_cfg=None, **kwargs): @register_model def mixer_s32_224(pretrained=False, **kwargs): """ Mixer-S/32 224x224 + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=32, num_blocks=8, hidden_dim=512, tokens_dim=256, channels_dim=2048, **kwargs) + model_args = dict(patch_size=32, num_blocks=8, hidden_dim=512, **kwargs) model = _create_mixer('mixer_s32_224', pretrained=pretrained, **model_args) return model @@ -186,28 +299,19 @@ def mixer_s32_224(pretrained=False, **kwargs): @register_model def mixer_s16_224(pretrained=False, **kwargs): """ Mixer-S/16 224x224 + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=16, num_blocks=8, hidden_dim=512, tokens_dim=256, channels_dim=2048, **kwargs) + model_args = dict(patch_size=16, num_blocks=8, hidden_dim=512, **kwargs) model = _create_mixer('mixer_s16_224', pretrained=pretrained, **model_args) return model -@register_model -def mixer_s16_glu_224(pretrained=False, **kwargs): - """ Mixer-S/16 224x224 - """ - model_args = dict( - patch_size=16, num_blocks=8, hidden_dim=512, tokens_dim=256, channels_dim=1536, - mlp_layer=GluMlp, act_layer=nn.SiLU, **kwargs) - model = _create_mixer('mixer_s16_glu_224', pretrained=pretrained, **model_args) - return model - - @register_model def mixer_b32_224(pretrained=False, **kwargs): """ Mixer-B/32 224x224 + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=32, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) + model_args = dict(patch_size=32, num_blocks=12, hidden_dim=768, **kwargs) model = _create_mixer('mixer_b32_224', pretrained=pretrained, **model_args) return model @@ -215,8 +319,9 @@ def mixer_b32_224(pretrained=False, **kwargs): @register_model def mixer_b16_224(pretrained=False, **kwargs): """ Mixer-B/16 224x224. ImageNet-1k pretrained weights. + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, **kwargs) model = _create_mixer('mixer_b16_224', pretrained=pretrained, **model_args) return model @@ -224,8 +329,9 @@ def mixer_b16_224(pretrained=False, **kwargs): @register_model def mixer_b16_224_in21k(pretrained=False, **kwargs): """ Mixer-B/16 224x224. ImageNet-21k pretrained weights. + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, **kwargs) model = _create_mixer('mixer_b16_224_in21k', pretrained=pretrained, **model_args) return model @@ -233,8 +339,9 @@ def mixer_b16_224_in21k(pretrained=False, **kwargs): @register_model def mixer_l32_224(pretrained=False, **kwargs): """ Mixer-L/32 224x224. + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=32, num_blocks=24, hidden_dim=1024, tokens_dim=512, channels_dim=4096, **kwargs) + model_args = dict(patch_size=32, num_blocks=24, hidden_dim=1024, **kwargs) model = _create_mixer('mixer_l32_224', pretrained=pretrained, **model_args) return model @@ -242,8 +349,9 @@ def mixer_l32_224(pretrained=False, **kwargs): @register_model def mixer_l16_224(pretrained=False, **kwargs): """ Mixer-L/16 224x224. ImageNet-1k pretrained weights. + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, tokens_dim=512, channels_dim=4096, **kwargs) + model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, **kwargs) model = _create_mixer('mixer_l16_224', pretrained=pretrained, **model_args) return model @@ -251,7 +359,101 @@ def mixer_l16_224(pretrained=False, **kwargs): @register_model def mixer_l16_224_in21k(pretrained=False, **kwargs): """ Mixer-L/16 224x224. ImageNet-21k pretrained weights. + Paper: 'MLP-Mixer: An all-MLP Architecture for Vision' - https://arxiv.org/abs/2105.01601 """ - model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, tokens_dim=512, channels_dim=4096, **kwargs) + model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, **kwargs) model = _create_mixer('mixer_l16_224_in21k', pretrained=pretrained, **model_args) return model + + +@register_model +def gmixer_12_224(pretrained=False, **kwargs): + """ Glu-Mixer-12 224x224 (short & fat) + Experiment by Ross Wightman, adding (Si)GLU to MLP-Mixer + """ + model_args = dict( + patch_size=20, num_blocks=12, hidden_dim=512, mlp_ratio=(1.0, 6.0), + mlp_layer=GluMlp, act_layer=nn.SiLU, **kwargs) + model = _create_mixer('gmixer_12_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def gmixer_24_224(pretrained=False, **kwargs): + """ Glu-Mixer-24 224x224 (tall & slim) + Experiment by Ross Wightman, adding (Si)GLU to MLP-Mixer + """ + model_args = dict( + patch_size=20, num_blocks=24, hidden_dim=384, mlp_ratio=(1.0, 6.0), + mlp_layer=GluMlp, act_layer=nn.SiLU, **kwargs) + model = _create_mixer('gmixer_24_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def resmlp_12_224(pretrained=False, **kwargs): + """ ResMLP-12 + Paper: `ResMLP: Feedforward networks for image classification...` - https://arxiv.org/abs/2105.03404 + """ + model_args = dict( + patch_size=16, num_blocks=12, hidden_dim=384, mlp_ratio=4, block_layer=ResBlock, norm_layer=Affine, **kwargs) + model = _create_mixer('resmlp_12_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def resmlp_24_224(pretrained=False, **kwargs): + """ ResMLP-24 + Paper: `ResMLP: Feedforward networks for image classification...` - https://arxiv.org/abs/2105.03404 + """ + model_args = dict( + patch_size=16, num_blocks=24, hidden_dim=384, mlp_ratio=4, block_layer=ResBlock, norm_layer=Affine, **kwargs) + model = _create_mixer('resmlp_24_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def resmlp_36_224(pretrained=False, **kwargs): + """ ResMLP-36 + Paper: `ResMLP: Feedforward networks for image classification...` - https://arxiv.org/abs/2105.03404 + """ + model_args = dict( + patch_size=16, num_blocks=36, hidden_dim=384, mlp_ratio=4, block_layer=ResBlock, norm_layer=Affine, **kwargs) + model = _create_mixer('resmlp_36_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def gmlp_ti16_224(pretrained=False, **kwargs): + """ gMLP-Tiny + Paper: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 + """ + model_args = dict( + patch_size=16, num_blocks=30, hidden_dim=128, mlp_ratio=6, block_layer=SpatialGatingBlock, + mlp_layer=GatedMlp, **kwargs) + model = _create_mixer('gmlp_ti16_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def gmlp_s16_224(pretrained=False, **kwargs): + """ gMLP-Small + Paper: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 + """ + model_args = dict( + patch_size=16, num_blocks=30, hidden_dim=256, mlp_ratio=6, block_layer=SpatialGatingBlock, + mlp_layer=GatedMlp, **kwargs) + model = _create_mixer('gmlp_s16_224', pretrained=pretrained, **model_args) + return model + + +@register_model +def gmlp_b16_224(pretrained=False, **kwargs): + """ gMLP-Base + Paper: `Pay Attention to MLPs` - https://arxiv.org/abs/2105.08050 + """ + model_args = dict( + patch_size=16, num_blocks=30, hidden_dim=512, mlp_ratio=6, block_layer=SpatialGatingBlock, + mlp_layer=GatedMlp, **kwargs) + model = _create_mixer('gmlp_b16_224', pretrained=pretrained, **model_args) + return model From 5bcf686cb0aad39b7c9114931db2e7fc2bc4f24c Mon Sep 17 00:00:00 2001 From: talrid Date: Wed, 19 May 2021 20:51:10 +0300 Subject: [PATCH 08/25] mixer_b16_224_miil --- timm/models/mlp_mixer.py | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+) diff --git a/timm/models/mlp_mixer.py b/timm/models/mlp_mixer.py index 248568fc..87edbfd6 100644 --- a/timm/models/mlp_mixer.py +++ b/timm/models/mlp_mixer.py @@ -60,6 +60,15 @@ default_cfgs = dict( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_mixer_l16_224_in21k-846aa33c.pth', num_classes=21843 ), + # Mixer ImageNet-21K-P pretraining + mixer_b16_224_miil_in21k=_cfg( + url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil_in21k.pth', + mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', num_classes=11221, + ), + mixer_b16_224_miil=_cfg( + url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil.pth', + mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', + ), ) @@ -255,3 +264,21 @@ def mixer_l16_224_in21k(pretrained=False, **kwargs): model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, tokens_dim=512, channels_dim=4096, **kwargs) model = _create_mixer('mixer_l16_224_in21k', pretrained=pretrained, **model_args) return model + +@register_model +def mixer_b16_224_miil(pretrained=False, **kwargs): + """ Mixer-B/16 224x224. ImageNet-21k pretrained weights. + Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K + """ + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) + model = _create_mixer('mixer_b16_224_miil', pretrained=pretrained, **model_args) + return model + +@register_model +def mixer_b16_224_miil_in21k(pretrained=False, **kwargs): + """ Mixer-B/16 224x224. ImageNet-1k pretrained weights. + Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K + """ + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) + model = _create_mixer('mixer_b16_224_miil_in21k', pretrained=pretrained, **model_args) + return model \ No newline at end of file From 6d81374b88ba7390cb5045b512787c6cc5b728f1 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Wed, 19 May 2021 11:09:42 -0700 Subject: [PATCH 09/25] Update tests for new mlp models --- tests/test_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_models.py b/tests/test_models.py index ced2fd76..b77b29ff 100644 --- a/tests/test_models.py +++ b/tests/test_models.py @@ -15,7 +15,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'): torch._C._jit_set_profiling_mode(False) # transformer models don't support many of the spatial / feature based model functionalities -NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', 'mixer_*'] +NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*'] NUM_NON_STD = len(NON_STD_FILTERS) # exclude models that cause specific test failures From d046498e0bf8ee5a8fcc80d91452363c62c262f0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E9=91=AB=E6=9D=B0?= Date: Thu, 20 May 2021 11:20:39 +0800 Subject: [PATCH 10/25] update test_models.py --- tests/test_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_models.py b/tests/test_models.py index b77b29ff..3013d0b9 100644 --- a/tests/test_models.py +++ b/tests/test_models.py @@ -15,7 +15,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'): torch._C._jit_set_profiling_mode(False) # transformer models don't support many of the spatial / feature based model functionalities -NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*'] +NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*', 'twins_*'] NUM_NON_STD = len(NON_STD_FILTERS) # exclude models that cause specific test failures From 240e6677468392283835c372fe2addc72514cff9 Mon Sep 17 00:00:00 2001 From: talrid Date: Thu, 20 May 2021 10:23:07 +0300 Subject: [PATCH 11/25] Revert "mixer_b16_224_miil" --- timm/models/mlp_mixer.py | 27 --------------------------- 1 file changed, 27 deletions(-) diff --git a/timm/models/mlp_mixer.py b/timm/models/mlp_mixer.py index 87edbfd6..248568fc 100644 --- a/timm/models/mlp_mixer.py +++ b/timm/models/mlp_mixer.py @@ -60,15 +60,6 @@ default_cfgs = dict( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_mixer_l16_224_in21k-846aa33c.pth', num_classes=21843 ), - # Mixer ImageNet-21K-P pretraining - mixer_b16_224_miil_in21k=_cfg( - url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil_in21k.pth', - mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', num_classes=11221, - ), - mixer_b16_224_miil=_cfg( - url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil.pth', - mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', - ), ) @@ -264,21 +255,3 @@ def mixer_l16_224_in21k(pretrained=False, **kwargs): model_args = dict(patch_size=16, num_blocks=24, hidden_dim=1024, tokens_dim=512, channels_dim=4096, **kwargs) model = _create_mixer('mixer_l16_224_in21k', pretrained=pretrained, **model_args) return model - -@register_model -def mixer_b16_224_miil(pretrained=False, **kwargs): - """ Mixer-B/16 224x224. ImageNet-21k pretrained weights. - Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K - """ - model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) - model = _create_mixer('mixer_b16_224_miil', pretrained=pretrained, **model_args) - return model - -@register_model -def mixer_b16_224_miil_in21k(pretrained=False, **kwargs): - """ Mixer-B/16 224x224. ImageNet-1k pretrained weights. - Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K - """ - model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, tokens_dim=384, channels_dim=3072, **kwargs) - model = _create_mixer('mixer_b16_224_miil_in21k', pretrained=pretrained, **model_args) - return model \ No newline at end of file From dc1a4efd28b335ebd85e13d64edd78404f75aeb7 Mon Sep 17 00:00:00 2001 From: talrid Date: Thu, 20 May 2021 10:35:50 +0300 Subject: [PATCH 12/25] mixer_b16_224_miil, mixer_b16_224_miil_in21k models --- timm/models/mlp_mixer.py | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+) diff --git a/timm/models/mlp_mixer.py b/timm/models/mlp_mixer.py index 2241fe43..92ca115b 100644 --- a/timm/models/mlp_mixer.py +++ b/timm/models/mlp_mixer.py @@ -80,6 +80,15 @@ default_cfgs = dict( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_mixer_l16_224_in21k-846aa33c.pth', num_classes=21843 ), + # Mixer ImageNet-21K-P pretraining + mixer_b16_224_miil_in21k=_cfg( + url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil_in21k.pth', + mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', num_classes=11221, + ), + mixer_b16_224_miil=_cfg( + url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/ImageNet_21K_P/models/timm/mixer_b16_224_miil.pth', + mean=(0, 0, 0), std=(1, 1, 1), crop_pct=0.875, interpolation='bilinear', + ), gmixer_12_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), gmixer_24_224=_cfg(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD), @@ -365,6 +374,23 @@ def mixer_l16_224_in21k(pretrained=False, **kwargs): model = _create_mixer('mixer_l16_224_in21k', pretrained=pretrained, **model_args) return model +@register_model +def mixer_b16_224_miil(pretrained=False, **kwargs): + """ Mixer-B/16 224x224. ImageNet-21k pretrained weights. + Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K + """ + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, **kwargs) + model = _create_mixer('mixer_b16_224_miil', pretrained=pretrained, **model_args) + return model + +@register_model +def mixer_b16_224_miil_in21k(pretrained=False, **kwargs): + """ Mixer-B/16 224x224. ImageNet-1k pretrained weights. + Weights taken from: https://github.com/Alibaba-MIIL/ImageNet21K + """ + model_args = dict(patch_size=16, num_blocks=12, hidden_dim=768, **kwargs) + model = _create_mixer('mixer_b16_224_miil_in21k', pretrained=pretrained, **model_args) + return model @register_model def gmixer_12_224(pretrained=False, **kwargs): From 8086943b6f4cef1ad7b1f044eafcd8e138dd5cfd Mon Sep 17 00:00:00 2001 From: Alexander Soare Date: Thu, 20 May 2021 11:27:58 +0100 Subject: [PATCH 13/25] allow resize positional embeddings to non-square grid --- timm/models/vision_transformer.py | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py index cc7e0903..1acdd808 100644 --- a/timm/models/vision_transformer.py +++ b/timm/models/vision_transformer.py @@ -352,7 +352,7 @@ def _init_vit_weights(m, n: str = '', head_bias: float = 0., jax_impl: bool = Fa nn.init.ones_(m.weight) -def resize_pos_embed(posemb, posemb_new, num_tokens=1): +def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=[]): # Rescale the grid of position embeddings when loading from state_dict. Adapted from # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224 _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape) @@ -363,11 +363,12 @@ def resize_pos_embed(posemb, posemb_new, num_tokens=1): else: posemb_tok, posemb_grid = posemb[:, :0], posemb[0] gs_old = int(math.sqrt(len(posemb_grid))) - gs_new = int(math.sqrt(ntok_new)) - _logger.info('Position embedding grid-size from %s to %s', gs_old, gs_new) + if not len(gs_new): # backwards compatibility + gs_new = [int(math.sqrt(ntok_new))]*2 + _logger.info('Position embedding grid-size from %s to %s', [gs_old, gs_old], gs_new) posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) - posemb_grid = F.interpolate(posemb_grid, size=(gs_new, gs_new), mode='bilinear') - posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new * gs_new, -1) + posemb_grid = F.interpolate(posemb_grid, size=gs_new, mode='bilinear') + posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1) posemb = torch.cat([posemb_tok, posemb_grid], dim=1) return posemb @@ -385,7 +386,8 @@ def checkpoint_filter_fn(state_dict, model): v = v.reshape(O, -1, H, W) elif k == 'pos_embed' and v.shape != model.pos_embed.shape: # To resize pos embedding when using model at different size from pretrained weights - v = resize_pos_embed(v, model.pos_embed, getattr(model, 'num_tokens', 1)) + v = resize_pos_embed(v, model.pos_embed, getattr(model, 'num_tokens', 1), + model.patch_embed.grid_size) out_dict[k] = v return out_dict From 79760198640dbb7e63889a322c1c70c1b5113b97 Mon Sep 17 00:00:00 2001 From: Alexander Soare Date: Thu, 20 May 2021 11:55:48 +0100 Subject: [PATCH 14/25] extend positional embedding resizing functionality to tnt --- timm/models/tnt.py | 31 +++++++++++++++++++++++-------- 1 file changed, 23 insertions(+), 8 deletions(-) diff --git a/timm/models/tnt.py b/timm/models/tnt.py index cc732677..8e038718 100644 --- a/timm/models/tnt.py +++ b/timm/models/tnt.py @@ -14,7 +14,9 @@ from functools import partial from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from timm.models.helpers import load_pretrained from timm.models.layers import Mlp, DropPath, trunc_normal_ +from timm.models.layers.helpers import to_2tuple from timm.models.registry import register_model +from timm.models.vision_transformer import resize_pos_embed def _cfg(url='', **kwargs): @@ -118,11 +120,15 @@ class PixelEmbed(nn.Module): """ def __init__(self, img_size=224, patch_size=16, in_chans=3, in_dim=48, stride=4): super().__init__() - num_patches = (img_size // patch_size) ** 2 + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + # grid_size property necessary for resizing positional embedding + self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) + num_patches = (self.grid_size[0]) * (self.grid_size[1]) self.img_size = img_size self.num_patches = num_patches self.in_dim = in_dim - new_patch_size = math.ceil(patch_size / stride) + new_patch_size = [math.ceil(ps / stride) for ps in patch_size] self.new_patch_size = new_patch_size self.proj = nn.Conv2d(in_chans, self.in_dim, kernel_size=7, padding=3, stride=stride) @@ -130,11 +136,11 @@ class PixelEmbed(nn.Module): def forward(self, x, pixel_pos): B, C, H, W = x.shape - assert H == self.img_size and W == self.img_size, \ - f"Input image size ({H}*{W}) doesn't match model ({self.img_size}*{self.img_size})." + assert H == self.img_size[0] and W == self.img_size[1], \ + f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." x = self.proj(x) x = self.unfold(x) - x = x.transpose(1, 2).reshape(B * self.num_patches, self.in_dim, self.new_patch_size, self.new_patch_size) + x = x.transpose(1, 2).reshape(B * self.num_patches, self.in_dim, self.new_patch_size[0], self.new_patch_size[1]) x = x + pixel_pos x = x.reshape(B * self.num_patches, self.in_dim, -1).transpose(1, 2) return x @@ -155,7 +161,7 @@ class TNT(nn.Module): num_patches = self.pixel_embed.num_patches self.num_patches = num_patches new_patch_size = self.pixel_embed.new_patch_size - num_pixel = new_patch_size ** 2 + num_pixel = new_patch_size[0] * new_patch_size[1] self.norm1_proj = norm_layer(num_pixel * in_dim) self.proj = nn.Linear(num_pixel * in_dim, embed_dim) @@ -163,7 +169,7 @@ class TNT(nn.Module): self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.patch_pos = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) - self.pixel_pos = nn.Parameter(torch.zeros(1, in_dim, new_patch_size, new_patch_size)) + self.pixel_pos = nn.Parameter(torch.zeros(1, in_dim, new_patch_size[0], new_patch_size[1])) self.pos_drop = nn.Dropout(p=drop_rate) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule @@ -224,6 +230,14 @@ class TNT(nn.Module): return x +def checkpoint_filter_fn(state_dict, model): + """ convert patch embedding weight from manual patchify + linear proj to conv""" + if state_dict['patch_pos'].shape != model.patch_pos.shape: + state_dict['patch_pos'] = resize_pos_embed(state_dict['patch_pos'], + model.patch_pos, getattr(model, 'num_tokens', 1), model.pixel_embed.grid_size) + return state_dict + + @register_model def tnt_s_patch16_224(pretrained=False, **kwargs): model = TNT(patch_size=16, embed_dim=384, in_dim=24, depth=12, num_heads=6, in_num_head=4, @@ -231,7 +245,8 @@ def tnt_s_patch16_224(pretrained=False, **kwargs): model.default_cfg = default_cfgs['tnt_s_patch16_224'] if pretrained: load_pretrained( - model, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3)) + model, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3), + filter_fn=checkpoint_filter_fn) return model From 40c506ba1ecac691955a8e486b99036d294cb763 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Thu, 20 May 2021 23:17:28 +0000 Subject: [PATCH 15/25] Add ConViT --- timm/models/__init__.py | 1 + timm/models/convit.py | 445 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 446 insertions(+) create mode 100644 timm/models/convit.py diff --git a/timm/models/__init__.py b/timm/models/__init__.py index 46ea155f..4d1230bd 100644 --- a/timm/models/__init__.py +++ b/timm/models/__init__.py @@ -2,6 +2,7 @@ from .byoanet import * from .byobnet import * from .cait import * from .coat import * +from .convit import * from .cspnet import * from .densenet import * from .dla import * diff --git a/timm/models/convit.py b/timm/models/convit.py new file mode 100644 index 00000000..82a0d988 --- /dev/null +++ b/timm/models/convit.py @@ -0,0 +1,445 @@ +"""These modules are adapted from those of timm, see +https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py +""" + +import torch +import torch.nn as nn +from functools import partial +import torch.nn.functional as F + +from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD +from .helpers import build_model_with_cfg +from timm.models.helpers import load_pretrained +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ +from timm.models.registry import register_model + +import torch +import torch.nn as nn +import matplotlib.pyplot as plt + + +def _cfg(url='', **kwargs): + return { + 'url': url, + 'num_classes': 1000, 'input_size': (3, 224, 224), + 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, + **kwargs + } + + +default_cfgs = { + # ConViT + 'convit_tiny': _cfg( + url="https://dl.fbaipublicfiles.com/convit/convit_tiny.pth"), + 'convit_small': _cfg( + url="https://dl.fbaipublicfiles.com/convit/convit_small.pth"), + 'convit_base': _cfg( + url="https://dl.fbaipublicfiles.com/convit/convit_base.pth") +} + + +class Mlp(nn.Module): + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +class GPSA(nn.Module): + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., + locality_strength=1., use_local_init=True): + super().__init__() + self.num_heads = num_heads + self.dim = dim + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.qk = nn.Linear(dim, dim * 2, bias=qkv_bias) + self.v = nn.Linear(dim, dim, bias=qkv_bias) + + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.pos_proj = nn.Linear(3, num_heads) + self.proj_drop = nn.Dropout(proj_drop) + self.locality_strength = locality_strength + self.gating_param = nn.Parameter(torch.ones(self.num_heads)) + self.apply(self._init_weights) + if use_local_init: + self.local_init(locality_strength=locality_strength) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + def forward(self, x): + B, N, C = x.shape + if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: + self.get_rel_indices(N) + + attn = self.get_attention(x) + v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + x = (attn @ v).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + def get_attention(self, x): + B, N, C = x.shape + qk = self.qk(x).reshape(B, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k = qk[0], qk[1] + pos_score = self.rel_indices.expand(B, -1, -1,-1) + pos_score = self.pos_proj(pos_score).permute(0,3,1,2) + patch_score = (q @ k.transpose(-2, -1)) * self.scale + patch_score = patch_score.softmax(dim=-1) + pos_score = pos_score.softmax(dim=-1) + + gating = self.gating_param.view(1,-1,1,1) + attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score + attn /= attn.sum(dim=-1).unsqueeze(-1) + attn = self.attn_drop(attn) + return attn + + def get_attention_map(self, x, return_map = False): + + attn_map = self.get_attention(x).mean(0) # average over batch + distances = self.rel_indices.squeeze()[:,:,-1]**.5 + dist = torch.einsum('nm,hnm->h', (distances, attn_map)) + dist /= distances.size(0) + if return_map: + return dist, attn_map + else: + return dist + + def local_init(self, locality_strength=1.): + + self.v.weight.data.copy_(torch.eye(self.dim)) + locality_distance = 1 #max(1,1/locality_strength**.5) + + kernel_size = int(self.num_heads**.5) + center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 + for h1 in range(kernel_size): + for h2 in range(kernel_size): + position = h1+kernel_size*h2 + self.pos_proj.weight.data[position,2] = -1 + self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance + self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance + self.pos_proj.weight.data *= locality_strength + + def get_rel_indices(self, num_patches): + img_size = int(num_patches**.5) + rel_indices = torch.zeros(1, num_patches, num_patches, 3) + ind = torch.arange(img_size).view(1,-1) - torch.arange(img_size).view(-1, 1) + indx = ind.repeat(img_size,img_size) + indy = ind.repeat_interleave(img_size,dim=0).repeat_interleave(img_size,dim=1) + indd = indx**2 + indy**2 + rel_indices[:,:,:,2] = indd.unsqueeze(0) + rel_indices[:,:,:,1] = indy.unsqueeze(0) + rel_indices[:,:,:,0] = indx.unsqueeze(0) + device = self.qk.weight.device + self.rel_indices = rel_indices.to(device) + + +class MHSA(nn.Module): + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + def get_attention_map(self, x, return_map = False): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] + attn_map = (q @ k.transpose(-2, -1)) * self.scale + attn_map = attn_map.softmax(dim=-1).mean(0) + + img_size = int(N**.5) + ind = torch.arange(img_size).view(1,-1) - torch.arange(img_size).view(-1, 1) + indx = ind.repeat(img_size,img_size) + indy = ind.repeat_interleave(img_size,dim=0).repeat_interleave(img_size,dim=1) + indd = indx**2 + indy**2 + distances = indd**.5 + distances = distances.to('cuda') + + dist = torch.einsum('nm,hnm->h', (distances, attn_map)) + dist /= N + + if return_map: + return dist, attn_map + else: + return dist + + + def forward(self, x): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] + + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class Block(nn.Module): + + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): + super().__init__() + self.norm1 = norm_layer(dim) + self.use_gpsa = use_gpsa + if self.use_gpsa: + self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) + else: + self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + def forward(self, x): + x = x + self.drop_path(self.attn(self.norm1(x))) + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding, from timm + """ + def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) + self.img_size = img_size + self.patch_size = patch_size + self.num_patches = num_patches + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + self.apply(self._init_weights) + + def forward(self, x): + B, C, H, W = x.shape + assert H == self.img_size[0] and W == self.img_size[1], \ + f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." + x = self.proj(x).flatten(2).transpose(1, 2) + return x + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + +class HybridEmbed(nn.Module): + """ CNN Feature Map Embedding, from timm + """ + def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768): + super().__init__() + assert isinstance(backbone, nn.Module) + img_size = to_2tuple(img_size) + self.img_size = img_size + self.backbone = backbone + if feature_size is None: + with torch.no_grad(): + training = backbone.training + if training: + backbone.eval() + o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1] + feature_size = o.shape[-2:] + feature_dim = o.shape[1] + backbone.train(training) + else: + feature_size = to_2tuple(feature_size) + feature_dim = self.backbone.feature_info.channels()[-1] + self.num_patches = feature_size[0] * feature_size[1] + self.proj = nn.Linear(feature_dim, embed_dim) + self.apply(self._init_weights) + + def forward(self, x): + x = self.backbone(x)[-1] + x = x.flatten(2).transpose(1, 2) + x = self.proj(x) + return x + + +class ConViT(nn.Module): + """ Vision Transformer with support for patch or hybrid CNN input stage + """ + def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, + num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., + drop_path_rate=0., hybrid_backbone=None, norm_layer=nn.LayerNorm, global_pool=None, + local_up_to_layer=3, locality_strength=1., use_pos_embed=True): + super().__init__() + embed_dim *= num_heads + self.num_classes = num_classes + self.local_up_to_layer = local_up_to_layer + self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models + self.locality_strength = locality_strength + self.use_pos_embed = use_pos_embed + + if hybrid_backbone is not None: + self.patch_embed = HybridEmbed( + hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim) + else: + self.patch_embed = PatchEmbed( + img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) + num_patches = self.patch_embed.num_patches + self.num_patches = num_patches + + self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + self.pos_drop = nn.Dropout(p=drop_rate) + + if self.use_pos_embed: + self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) + trunc_normal_(self.pos_embed, std=.02) + + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule + self.blocks = nn.ModuleList([ + Block( + dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, + drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, + use_gpsa=True, + locality_strength=locality_strength) + if i 0 else nn.Identity() + + trunc_normal_(self.cls_token, std=.02) + self.head.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'pos_embed', 'cls_token'} + + def get_classifier(self): + return self.head + + def reset_classifier(self, num_classes, global_pool=''): + self.num_classes = num_classes + self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity() + + def forward_features(self, x): + B = x.shape[0] + x = self.patch_embed(x) + + cls_tokens = self.cls_token.expand(B, -1, -1) + + if self.use_pos_embed: + x = x + self.pos_embed + x = self.pos_drop(x) + + for u,blk in enumerate(self.blocks): + if u == self.local_up_to_layer : + x = torch.cat((cls_tokens, x), dim=1) + x = blk(x) + + x = self.norm(x) + return x[:, 0] + + def forward(self, x): + x = self.forward_features(x) + x = self.head(x) + return x + + +def _create_convit(variant, pretrained=False, **kwargs): + return build_model_with_cfg( + ConViT, variant, pretrained, + default_cfg=default_cfgs[variant], + **kwargs) + + +@register_model +def convit_tiny(pretrained=False, **kwargs): + model_args = dict( + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + num_heads=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) + model = _create_convit( + variant='convit_tiny', pretrained=pretrained, **model_args) + return model + +@register_model +def convit_small(pretrained=False, **kwargs): + model_args = dict( + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + num_heads=9, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) + model = _create_convit( + variant='convit_small', pretrained=pretrained, **model_args) + return model + +@register_model +def convit_base(pretrained=False, **kwargs): + model_args = dict( + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + num_heads=16, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) + model = _create_convit( + variant='convit_base', pretrained=pretrained, **model_args) + return model From 8b1f2e8e1f4c73f43d9e956c2162884ab319b1a3 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Thu, 20 May 2021 23:42:42 +0000 Subject: [PATCH 16/25] remote unused matplotlib import --- timm/models/convit.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/timm/models/convit.py b/timm/models/convit.py index 82a0d988..29970c76 100644 --- a/timm/models/convit.py +++ b/timm/models/convit.py @@ -9,13 +9,11 @@ import torch.nn.functional as F from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .helpers import build_model_with_cfg -from timm.models.helpers import load_pretrained from timm.models.layers import DropPath, to_2tuple, trunc_normal_ from timm.models.registry import register_model import torch import torch.nn as nn -import matplotlib.pyplot as plt def _cfg(url='', **kwargs): From 163331748935559923ffb6aa5ed1882b47a6a92a Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Fri, 21 May 2021 01:11:56 +0000 Subject: [PATCH 17/25] update tests and exclude convit_base --- tests/test_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_models.py b/tests/test_models.py index b77b29ff..1bf0d738 100644 --- a/tests/test_models.py +++ b/tests/test_models.py @@ -24,7 +24,7 @@ if 'GITHUB_ACTIONS' in os.environ: # and 'Linux' in platform.system(): EXCLUDE_FILTERS = [ '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm', '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', - '*resnetrs350*', '*resnetrs420*'] + NON_STD_FILTERS + '*resnetrs350*', '*resnetrs420*', 'convit_base'] + NON_STD_FILTERS else: EXCLUDE_FILTERS = NON_STD_FILTERS From 5db1eb6ba56f35fce8bc06e85c7339e7c714a4f4 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Fri, 21 May 2021 02:11:20 +0000 Subject: [PATCH 18/25] Add defaults --- timm/models/convit.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/timm/models/convit.py b/timm/models/convit.py index 29970c76..31c05df3 100644 --- a/timm/models/convit.py +++ b/timm/models/convit.py @@ -19,8 +19,9 @@ import torch.nn as nn def _cfg(url='', **kwargs): return { 'url': url, - 'num_classes': 1000, 'input_size': (3, 224, 224), + 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None, 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, + 'first_conv': 'patch_embed.proj', 'classifier': 'head', **kwargs } From 50d6aab0efb53b4072008780fb7ea3cc82e0236f Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Fri, 21 May 2021 03:46:47 +0000 Subject: [PATCH 19/25] Add convit to non-std filters as vit_ --- tests/test_models.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/test_models.py b/tests/test_models.py index 1bf0d738..f098fefd 100644 --- a/tests/test_models.py +++ b/tests/test_models.py @@ -15,7 +15,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'): torch._C._jit_set_profiling_mode(False) # transformer models don't support many of the spatial / feature based model functionalities -NON_STD_FILTERS = ['vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*'] +NON_STD_FILTERS = ['vit_*', 'tnt_*', 'convit_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*'] NUM_NON_STD = len(NON_STD_FILTERS) # exclude models that cause specific test failures @@ -24,7 +24,7 @@ if 'GITHUB_ACTIONS' in os.environ: # and 'Linux' in platform.system(): EXCLUDE_FILTERS = [ '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm', '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', - '*resnetrs350*', '*resnetrs420*', 'convit_base'] + NON_STD_FILTERS + '*resnetrs350*', '*resnetrs420*'] + NON_STD_FILTERS else: EXCLUDE_FILTERS = NON_STD_FILTERS From be99eef9c14fe63a2ebf3cdd2784d16140851004 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Thu, 20 May 2021 23:38:35 -0700 Subject: [PATCH 20/25] Remove redundant code, cleanup, fix torchscript. --- timm/models/twins.py | 495 +++++++++++++------------------------------ 1 file changed, 149 insertions(+), 346 deletions(-) diff --git a/timm/models/twins.py b/timm/models/twins.py index 27be4cba..ce51c497 100644 --- a/timm/models/twins.py +++ b/timm/models/twins.py @@ -11,11 +11,9 @@ Code/weights from https://github.com/Meituan-AutoML/Twins, original copyright/li # Licensed under The Apache 2.0 License [see LICENSE for details] # Written by Xinjie Li, Xiangxiang Chu # -------------------------------------------------------- - -import logging import math from copy import deepcopy -from typing import Optional +from typing import Optional, Tuple import torch import torch.nn as nn @@ -25,13 +23,9 @@ from functools import partial from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .layers import Mlp, DropPath, to_2tuple, trunc_normal_ from .registry import register_model -from .vision_transformer import _cfg -from .vision_transformer import Block as TimmBlock -from .vision_transformer import Attention as TimmAttention +from .vision_transformer import Attention from .helpers import build_model_with_cfg, overlay_external_default_cfg -from .vision_transformer import checkpoint_filter_fn, _init_vit_weights -_logger = logging.getLogger(__name__) def _cfg(url='', **kwargs): return { @@ -43,6 +37,7 @@ def _cfg(url='', **kwargs): **kwargs } + default_cfgs = { 'twins_pcpvt_small': _cfg( url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_small.pth', @@ -64,78 +59,34 @@ default_cfgs = { ), } +Size_ = Tuple[int, int] -class GroupAttention(nn.Module): - """ - LSA: self attention within a group +class LocallyGroupedAttn(nn.Module): + """ LSA: self attention within a group """ - def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., ws=1): + def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1): assert ws != 1 - super(GroupAttention, self).__init__() + super(LocallyGroupedAttn, self).__init__() assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." self.dim = dim self.num_heads = num_heads head_dim = dim // num_heads - self.scale = qk_scale or head_dim ** -0.5 + self.scale = head_dim ** -0.5 - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.qkv = nn.Linear(dim, dim * 3, bias=True) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) self.ws = ws - def forward(self, x, H, W): - """ - There are two implementations for this function, zero padding or mask. We don't observe obvious difference for - both. You can choose any one, we recommend forward_padding because it's neat. However, - the masking implementation is more reasonable and accurate. - Args: - x: - H: - W: - - Returns: - - """ - return self.forward_padding(x, H, W) - - def forward_mask(self, x, H, W): - B, N, C = x.shape - x = x.view(B, H, W, C) - pad_l = pad_t = 0 - pad_r = (self.ws - W % self.ws) % self.ws - pad_b = (self.ws - H % self.ws) % self.ws - x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) - _, Hp, Wp, _ = x.shape - _h, _w = Hp // self.ws, Wp // self.ws - mask = torch.zeros((1, Hp, Wp), device=x.device) - mask[:, -pad_b:, :].fill_(1) - mask[:, :, -pad_r:].fill_(1) - - x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) # B, _h, _w, ws, ws, C - mask = mask.reshape(1, _h, self.ws, _w, self.ws).transpose(2, 3).reshape(1, _h*_w, self.ws*self.ws) - attn_mask = mask.unsqueeze(2) - mask.unsqueeze(3) # 1, _h*_w, ws*ws, ws*ws - attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-1000.0)).masked_fill(attn_mask == 0, float(0.0)) - qkv = self.qkv(x).reshape(B, _h * _w, self.ws * self.ws, 3, self.num_heads, - C // self.num_heads).permute(3, 0, 1, 4, 2, 5) # n_h, B, _w*_h, nhead, ws*ws, dim - q, k, v = qkv[0], qkv[1], qkv[2] # B, _h*_w, n_head, ws*ws, dim_head - attn = (q @ k.transpose(-2, -1)) * self.scale # B, _h*_w, n_head, ws*ws, ws*ws - attn = attn + attn_mask.unsqueeze(2) - attn = attn.softmax(dim=-1) - attn = self.attn_drop(attn) # attn @v -> B, _h*_w, n_head, ws*ws, dim_head - attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C) - x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C) - if pad_r > 0 or pad_b > 0: - x = x[:, :H, :W, :].contiguous() - x = x.reshape(B, N, C) - x = self.proj(x) - x = self.proj_drop(x) - return x - - def forward_padding(self, x, H, W): + def forward(self, x, size: Size_): + # There are two implementations for this function, zero padding or mask. We don't observe obvious difference for + # both. You can choose any one, we recommend forward_padding because it's neat. However, + # the masking implementation is more reasonable and accurate. B, N, C = x.shape + H, W = size x = x.view(B, H, W, C) pad_l = pad_t = 0 pad_r = (self.ws - W % self.ws) % self.ws @@ -144,8 +95,8 @@ class GroupAttention(nn.Module): _, Hp, Wp, _ = x.shape _h, _w = Hp // self.ws, Wp // self.ws x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) - qkv = self.qkv(x).reshape(B, _h * _w, self.ws * self.ws, 3, self.num_heads, - C // self.num_heads).permute(3, 0, 1, 4, 2, 5) + qkv = self.qkv(x).reshape( + B, _h * _w, self.ws * self.ws, 3, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5) q, k, v = qkv[0], qkv[1], qkv[2] attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) @@ -159,22 +110,56 @@ class GroupAttention(nn.Module): x = self.proj_drop(x) return x - -class Attention(nn.Module): - """ - GSA: using a key to summarize the information for a group to be efficient. + # def forward_mask(self, x, size: Size_): + # B, N, C = x.shape + # H, W = size + # x = x.view(B, H, W, C) + # pad_l = pad_t = 0 + # pad_r = (self.ws - W % self.ws) % self.ws + # pad_b = (self.ws - H % self.ws) % self.ws + # x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + # _, Hp, Wp, _ = x.shape + # _h, _w = Hp // self.ws, Wp // self.ws + # mask = torch.zeros((1, Hp, Wp), device=x.device) + # mask[:, -pad_b:, :].fill_(1) + # mask[:, :, -pad_r:].fill_(1) + # + # x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) # B, _h, _w, ws, ws, C + # mask = mask.reshape(1, _h, self.ws, _w, self.ws).transpose(2, 3).reshape(1, _h * _w, self.ws * self.ws) + # attn_mask = mask.unsqueeze(2) - mask.unsqueeze(3) # 1, _h*_w, ws*ws, ws*ws + # attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-1000.0)).masked_fill(attn_mask == 0, float(0.0)) + # qkv = self.qkv(x).reshape( + # B, _h * _w, self.ws * self.ws, 3, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5) + # # n_h, B, _w*_h, nhead, ws*ws, dim + # q, k, v = qkv[0], qkv[1], qkv[2] # B, _h*_w, n_head, ws*ws, dim_head + # attn = (q @ k.transpose(-2, -1)) * self.scale # B, _h*_w, n_head, ws*ws, ws*ws + # attn = attn + attn_mask.unsqueeze(2) + # attn = attn.softmax(dim=-1) + # attn = self.attn_drop(attn) # attn @v -> B, _h*_w, n_head, ws*ws, dim_head + # attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C) + # x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C) + # if pad_r > 0 or pad_b > 0: + # x = x[:, :H, :W, :].contiguous() + # x = x.reshape(B, N, C) + # x = self.proj(x) + # x = self.proj_drop(x) + # return x + + +class GlobalSubSampleAttn(nn.Module): + """ GSA: using a key to summarize the information for a group to be efficient. """ - def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1): + def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1): super().__init__() assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." self.dim = dim self.num_heads = num_heads head_dim = dim // num_heads - self.scale = qk_scale or head_dim ** -0.5 + self.scale = head_dim ** -0.5 - self.q = nn.Linear(dim, dim, bias=qkv_bias) - self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias) + self.q = nn.Linear(dim, dim, bias=True) + self.kv = nn.Linear(dim, dim * 2, bias=True) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) @@ -183,18 +168,19 @@ class Attention(nn.Module): if sr_ratio > 1: self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio) self.norm = nn.LayerNorm(dim) + else: + self.sr = None + self.norm = None - def forward(self, x, H, W): + def forward(self, x, size: Size_): B, N, C = x.shape q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) - if self.sr_ratio > 1: - x_ = x.permute(0, 2, 1).reshape(B, C, H, W) - x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1) - x_ = self.norm(x_) - kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) - else: - kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + if self.sr is not None: + x = x.permute(0, 2, 1).reshape(B, C, *size) + x = self.sr(x).reshape(B, C, -1).permute(0, 2, 1) + x = self.norm(x) + kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) k, v = kv[0], kv[1] attn = (q @ k.transpose(-2, -1)) * self.scale @@ -210,52 +196,46 @@ class Attention(nn.Module): class Block(nn.Module): - def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., - drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1): + def __init__(self, dim, num_heads, mlp_ratio=4., drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1, ws=None): super().__init__() self.norm1 = norm_layer(dim) - self.attn = Attention( - dim, - num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, - attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio) + if ws is None: + self.attn = Attention(dim, num_heads, False, None, attn_drop, drop) + elif ws == 1: + self.attn = GlobalSubSampleAttn(dim, num_heads, attn_drop, drop, sr_ratio) + else: + self.attn = LocallyGroupedAttn(dim, num_heads, attn_drop, drop, ws) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) - def forward(self, x, H, W): - x = x + self.drop_path(self.attn(self.norm1(x), H, W)) + def forward(self, x, size: Size_): + x = x + self.drop_path(self.attn(self.norm1(x), size)) x = x + self.drop_path(self.mlp(self.norm2(x))) - return x -class SBlock(TimmBlock): - def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., - drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1): - super(SBlock, self).__init__(dim, num_heads, mlp_ratio, qkv_bias, qk_scale, drop, attn_drop, - drop_path, act_layer, norm_layer) - - def forward(self, x, H, W): - return super(SBlock, self).forward(x) - - -class GroupBlock(TimmBlock): - def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., - drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1, ws=1): - super(GroupBlock, self).__init__(dim, num_heads, mlp_ratio, qkv_bias, qk_scale, drop, attn_drop, - drop_path, act_layer, norm_layer) - del self.attn - if ws == 1: - self.attn = Attention(dim, num_heads, qkv_bias, qk_scale, attn_drop, drop, sr_ratio) - else: - self.attn = GroupAttention(dim, num_heads, qkv_bias, qk_scale, attn_drop, drop, ws) +class PosConv(nn.Module): + # PEG from https://arxiv.org/abs/2102.10882 + def __init__(self, in_chans, embed_dim=768, stride=1): + super(PosConv, self).__init__() + self.proj = nn.Sequential(nn.Conv2d(in_chans, embed_dim, 3, stride, 1, bias=True, groups=embed_dim), ) + self.stride = stride - def forward(self, x, H, W): - x = x + self.drop_path(self.attn(self.norm1(x), H, W)) - x = x + self.drop_path(self.mlp(self.norm2(x))) + def forward(self, x, size: Size_): + B, N, C = x.shape + cnn_feat_token = x.transpose(1, 2).view(B, C, *size) + x = self.proj(cnn_feat_token) + if self.stride == 1: + x += cnn_feat_token + x = x.flatten(2).transpose(1, 2) return x + def no_weight_decay(self): + return ['proj.%d.weight' % i for i in range(4)] + class PatchEmbed(nn.Module): """ Image to Patch Embedding @@ -263,7 +243,7 @@ class PatchEmbed(nn.Module): def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): super().__init__() - # img_size = to_2tuple(img_size) + img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) self.img_size = img_size @@ -275,90 +255,62 @@ class PatchEmbed(nn.Module): self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) self.norm = nn.LayerNorm(embed_dim) - def forward(self, x): + def forward(self, x) -> Tuple[torch.Tensor, Size_]: B, C, H, W = x.shape x = self.proj(x).flatten(2).transpose(1, 2) x = self.norm(x) - H, W = H // self.patch_size[0], W // self.patch_size[1] + out_size = (H // self.patch_size[0], W // self.patch_size[1]) - return x, (H, W) + return x, out_size -# borrow from PVT https://github.com/whai362/PVT.git -class PyramidVisionTransformer(nn.Module): - def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512], - num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., - attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, - depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], block_cls=Block): +class Twins(nn.Module): + # Adapted from PVT https://github.com/whai362/PVT.git + def __init__( + self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=(64, 128, 256, 512), + num_heads=(1, 2, 4, 8), mlp_ratios=(4, 4, 4, 4), drop_rate=0., attn_drop_rate=0., drop_path_rate=0., + norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=(3, 4, 6, 3), sr_ratios=(8, 4, 2, 1), wss=None, + block_cls=Block): super().__init__() self.num_classes = num_classes self.depths = depths - # patch_embed + img_size = to_2tuple(img_size) + prev_chs = in_chans self.patch_embeds = nn.ModuleList() - self.pos_embeds = nn.ParameterList() self.pos_drops = nn.ModuleList() - self.blocks = nn.ModuleList() - for i in range(len(depths)): - if i == 0: - self.patch_embeds.append(PatchEmbed(img_size, patch_size, in_chans, embed_dims[i])) - else: - self.patch_embeds.append( - # PatchEmbed(img_size // patch_size // 2 ** (i - 1), 2, embed_dims[i - 1], embed_dims[i]) - PatchEmbed((img_size[0] // patch_size // 2**(i-1),img_size[1] // patch_size // 2**(i-1)), 2, embed_dims[i - 1], embed_dims[i]) - ) - patch_num = self.patch_embeds[-1].num_patches + 1 if i == len(embed_dims) - 1 else self.patch_embeds[ - -1].num_patches - self.pos_embeds.append(nn.Parameter(torch.zeros(1, patch_num, embed_dims[i]))) + self.patch_embeds.append(PatchEmbed(img_size, patch_size, prev_chs, embed_dims[i])) self.pos_drops.append(nn.Dropout(p=drop_rate)) + prev_chs = embed_dims[i] + img_size = tuple(t // patch_size for t in img_size) + patch_size = 2 + self.blocks = nn.ModuleList() dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule cur = 0 for k in range(len(depths)): _block = nn.ModuleList([block_cls( - dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, - sr_ratio=sr_ratios[k]) - for i in range(depths[k])]) + dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], drop=drop_rate, + attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[k], + ws=1 if wss is None or i % 2 == 1 else wss[k]) for i in range(depths[k])]) self.blocks.append(_block) cur += depths[k] - self.norm = norm_layer(embed_dims[-1]) + self.pos_block = nn.ModuleList([PosConv(embed_dim, embed_dim) for embed_dim in embed_dims]) - # cls_token - self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dims[-1])) + self.norm = norm_layer(embed_dims[-1]) # classification head self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity() # init weights - for pos_emb in self.pos_embeds: - trunc_normal_(pos_emb, std=.02) self.apply(self._init_weights) - def reset_drop_path(self, drop_path_rate): - dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))] - cur = 0 - for k in range(len(self.depths)): - for i in range(self.depths[k]): - self.blocks[k][i].drop_path.drop_prob = dpr[cur + i] - cur += self.depths[k] - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - @torch.jit.ignore def no_weight_decay(self): - return {'cls_token'} + return set(['pos_block.' + n for n, p in self.pos_block.named_parameters()]) def get_classifier(self): return self.head @@ -367,76 +319,7 @@ class PyramidVisionTransformer(nn.Module): self.num_classes = num_classes self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity() - def forward_features(self, x): - B = x.shape[0] - for i in range(len(self.depths)): - x, (H, W) = self.patch_embeds[i](x) - if i == len(self.depths) - 1: - cls_tokens = self.cls_token.expand(B, -1, -1) - x = torch.cat((cls_tokens, x), dim=1) - x = x + self.pos_embeds[i] - x = self.pos_drops[i](x) - for blk in self.blocks[i]: - x = blk(x, H, W) - if i < len(self.depths) - 1: - x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() - - x = self.norm(x) - - return x[:, 0] - - def forward(self, x): - x = self.forward_features(x) - x = self.head(x) - - return x - - -# PEG from https://arxiv.org/abs/2102.10882 -class PosCNN(nn.Module): - def __init__(self, in_chans, embed_dim=768, s=1): - super(PosCNN, self).__init__() - self.proj = nn.Sequential(nn.Conv2d(in_chans, embed_dim, 3, s, 1, bias=True, groups=embed_dim), ) - self.s = s - - def forward(self, x, H, W): - B, N, C = x.shape - feat_token = x - cnn_feat = feat_token.transpose(1, 2).view(B, C, H, W) - if self.s == 1: - x = self.proj(cnn_feat) + cnn_feat - else: - x = self.proj(cnn_feat) - x = x.flatten(2).transpose(1, 2) - return x - - def no_weight_decay(self): - return ['proj.%d.weight' % i for i in range(4)] - - -class CPVTV2(PyramidVisionTransformer): - """ - Use useful results from CPVT. PEG and GAP. - Therefore, cls token is no longer required. - PEG is used to encode the absolute position on the fly, which greatly affects the performance when input resolution - changes during the training (such as segmentation, detection) - """ - def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512], - num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., - attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, - depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], block_cls=Block): - super(CPVTV2, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, mlp_ratios, - qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, norm_layer, depths, - sr_ratios, block_cls) - del self.pos_embeds - del self.cls_token - self.pos_block = nn.ModuleList( - [PosCNN(embed_dim, embed_dim) for embed_dim in embed_dims] - ) - self.apply(self._init_weights) - def _init_weights(self, m): - import math if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: @@ -454,98 +337,28 @@ class CPVTV2(PyramidVisionTransformer): m.weight.data.fill_(1.0) m.bias.data.zero_() - def no_weight_decay(self): - return set(['cls_token'] + ['pos_block.' + n for n, p in self.pos_block.named_parameters()]) - def forward_features(self, x): B = x.shape[0] - - for i in range(len(self.depths)): - x, (H, W) = self.patch_embeds[i](x) - x = self.pos_drops[i](x) - for j, blk in enumerate(self.blocks[i]): - x = blk(x, H, W) + for i, (embed, drop, blocks, pos_blk) in enumerate( + zip(self.patch_embeds, self.pos_drops, self.blocks, self.pos_block)): + x, size = embed(x) + x = drop(x) + for j, blk in enumerate(blocks): + x = blk(x, size) if j == 0: - x = self.pos_block[i](x, H, W) # PEG here + x = pos_blk(x, size) # PEG here if i < len(self.depths) - 1: - x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() - + x = x.reshape(B, *size, -1).permute(0, 3, 1, 2).contiguous() x = self.norm(x) - return x.mean(dim=1) # GAP here + def forward(self, x): + x = self.forward_features(x) + x = self.head(x) + return x -class Twins_PCPVT(CPVTV2): - def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256], - num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., - attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, - depths=[4, 4, 4], sr_ratios=[4, 2, 1], block_cls=SBlock): - super(Twins_PCPVT, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, - mlp_ratios, qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, - norm_layer, depths, sr_ratios, block_cls) - - -class Twins_SVT(Twins_PCPVT): - """ - alias Twins-SVT - """ - def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256], - num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., - attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, - depths=[4, 4, 4], sr_ratios=[4, 2, 1], block_cls=GroupBlock, wss=[7, 7, 7]): - super(Twins_SVT, self).__init__(img_size, patch_size, in_chans, num_classes, embed_dims, num_heads, - mlp_ratios, qkv_bias, qk_scale, drop_rate, attn_drop_rate, drop_path_rate, - norm_layer, depths, sr_ratios, block_cls) - del self.blocks - self.wss = wss - # transformer encoder - dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule - cur = 0 - self.blocks = nn.ModuleList() - for k in range(len(depths)): - _block = nn.ModuleList([block_cls( - dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], qkv_bias=qkv_bias, - qk_scale=qk_scale, - drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, - sr_ratio=sr_ratios[k], ws=1 if i % 2 == 1 else wss[k]) for i in range(depths[k])]) - self.blocks.append(_block) - cur += depths[k] - self.apply(self._init_weights) - - -def _conv_filter(state_dict, patch_size=16): - """ convert patch embedding weight from manual patchify + linear proj to conv""" - out_dict = {} - for k, v in state_dict.items(): - if 'patch_embed.proj.weight' in k: - v = v.reshape((v.shape[0], 3, patch_size, patch_size)) - out_dict[k] = v - - return out_dict - -def _create_twins_svt(variant, pretrained=False, default_cfg=None, **kwargs): - if default_cfg is None: - default_cfg = deepcopy(default_cfgs[variant]) - overlay_external_default_cfg(default_cfg, kwargs) - default_num_classes = default_cfg['num_classes'] - default_img_size = default_cfg['input_size'][-2:] - - num_classes = kwargs.pop('num_classes', default_num_classes) - img_size = kwargs.pop('img_size', default_img_size) - if kwargs.get('features_only', None): - raise RuntimeError('features_only not implemented for Vision Transformer models.') - - model = build_model_with_cfg( - Twins_SVT, variant, pretrained, - default_cfg=default_cfg, - img_size=img_size, - num_classes=num_classes, - pretrained_filter_fn=checkpoint_filter_fn, - **kwargs) - - return model -def _create_twins_pcpvt(variant, pretrained=False, default_cfg=None, **kwargs): +def _create_twins(variant, pretrained=False, default_cfg=None, **kwargs): if default_cfg is None: default_cfg = deepcopy(default_cfgs[variant]) overlay_external_default_cfg(default_cfg, kwargs) @@ -558,11 +371,10 @@ def _create_twins_pcpvt(variant, pretrained=False, default_cfg=None, **kwargs): raise RuntimeError('features_only not implemented for Vision Transformer models.') model = build_model_with_cfg( - CPVTV2, variant, pretrained, + Twins, variant, pretrained, default_cfg=default_cfg, img_size=img_size, num_classes=num_classes, - pretrained_filter_fn=checkpoint_filter_fn, **kwargs) return model @@ -571,55 +383,46 @@ def _create_twins_pcpvt(variant, pretrained=False, default_cfg=None, **kwargs): @register_model def twins_pcpvt_small(pretrained=False, **kwargs): model_kwargs = dict( - patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], - **kwargs) - return _create_twins_pcpvt('twins_pcpvt_small', pretrained=pretrained, **model_kwargs) + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], + depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_pcpvt_small', pretrained=pretrained, **model_kwargs) @register_model def twins_pcpvt_base(pretrained=False, **kwargs): model_kwargs = dict( - patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1], - **kwargs) - return _create_twins_pcpvt('twins_pcpvt_base', pretrained=pretrained, **model_kwargs) + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], + depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_pcpvt_base', pretrained=pretrained, **model_kwargs) @register_model def twins_pcpvt_large(pretrained=False, **kwargs): model_kwargs = dict( - patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], - **kwargs) - return _create_twins_pcpvt('twins_pcpvt_large', pretrained=pretrained, **model_kwargs) + patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], + depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_pcpvt_large', pretrained=pretrained, **model_kwargs) @register_model def twins_svt_small(pretrained=False, **kwargs): model_kwargs = dict( - patch_size=4, embed_dims=[64, 128, 256, 512], num_heads=[2, 4, 8, 16], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 10, 4], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], - **kwargs) - return _create_twins_svt('twins_svt_small', pretrained=pretrained, **model_kwargs) + patch_size=4, embed_dims=[64, 128, 256, 512], num_heads=[2, 4, 8, 16], mlp_ratios=[4, 4, 4, 4], + depths=[2, 2, 10, 4], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_svt_small', pretrained=pretrained, **model_kwargs) @register_model def twins_svt_base(pretrained=False, **kwargs): model_kwargs = dict( - patch_size=4, embed_dims=[96, 192, 384, 768], num_heads=[3, 6, 12, 24], mlp_ratios=[4, 4, 4, 4], qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], - **kwargs) - - return _create_twins_svt('twins_svt_base', pretrained=pretrained, **model_kwargs) + patch_size=4, embed_dims=[96, 192, 384, 768], num_heads=[3, 6, 12, 24], mlp_ratios=[4, 4, 4, 4], + depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_svt_base', pretrained=pretrained, **model_kwargs) @register_model def twins_svt_large(pretrained=False, **kwargs): model_kwargs = dict( patch_size=4, embed_dims=[128, 256, 512, 1024], num_heads=[4, 8, 16, 32], mlp_ratios=[4, 4, 4, 4], - qkv_bias=True, - norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], - **kwargs) - - return _create_twins_svt('twins_svt_large', pretrained=pretrained, **model_kwargs) + depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs) + return _create_twins('twins_svt_large', pretrained=pretrained, **model_kwargs) From a569635045b83bfd7f86881694b2515fed575592 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 21 May 2021 16:23:14 -0700 Subject: [PATCH 21/25] Update twin weights to a copy in GitHub releases for faster dl. Tweak model class comment. --- timm/models/twins.py | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) diff --git a/timm/models/twins.py b/timm/models/twins.py index ce51c497..a534d174 100644 --- a/timm/models/twins.py +++ b/timm/models/twins.py @@ -40,22 +40,22 @@ def _cfg(url='', **kwargs): default_cfgs = { 'twins_pcpvt_small': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_small.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_small-e70e7e7a.pth', ), 'twins_pcpvt_base': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_base.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_base-e5ecb09b.pth', ), 'twins_pcpvt_large': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/pcpvt_large.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_large-d273f802.pth', ), 'twins_svt_small': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_small.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_small-42e5f78c.pth', ), 'twins_svt_base': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_base.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_base-c2265010.pth', ), 'twins_svt_large': _cfg( - url='https://s3plus.meituan.net/v1/mss_9240d97c6bf34ab1b78859c3c2a2a3e4/automl-model-zoo/models/twins/alt_gvt_large.pth', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_large-90f6aaa9.pth', ), } @@ -266,7 +266,10 @@ class PatchEmbed(nn.Module): class Twins(nn.Module): - # Adapted from PVT https://github.com/whai362/PVT.git + """ Twins Vision Transfomer (Revisiting Spatial Attention) + + Adapted from PVT (PyramidVisionTransformer) class at https://github.com/whai362/PVT.git + """ def __init__( self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=(64, 128, 256, 512), num_heads=(1, 2, 4, 8), mlp_ratios=(4, 4, 4, 4), drop_rate=0., attn_drop_rate=0., drop_path_rate=0., From b7de82e835682c2f90b6a5fc9fd325d1457193b6 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 21 May 2021 17:04:23 -0700 Subject: [PATCH 22/25] ConViT cleanup, fix torchscript, bit of reformatting, reuse existing layers. --- timm/models/convit.py | 290 ++++++++++++++---------------------------- 1 file changed, 98 insertions(+), 192 deletions(-) diff --git a/timm/models/convit.py b/timm/models/convit.py index 31c05df3..f6ae3ec1 100644 --- a/timm/models/convit.py +++ b/timm/models/convit.py @@ -1,6 +1,24 @@ -"""These modules are adapted from those of timm, see -https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py +""" ConViT Model + +@article{d2021convit, + title={ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases}, + author={d'Ascoli, St{\'e}phane and Touvron, Hugo and Leavitt, Matthew and Morcos, Ari and Biroli, Giulio and Sagun, Levent}, + journal={arXiv preprint arXiv:2103.10697}, + year={2021} +} + +Paper link: https://arxiv.org/abs/2103.10697 +Original code: https://github.com/facebookresearch/convit, original copyright below """ +# Copyright (c) 2015-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the CC-by-NC license found in the +# LICENSE file in the root directory of this source tree. +# +'''These modules are adapted from those of timm, see +https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py +''' import torch import torch.nn as nn @@ -9,8 +27,9 @@ import torch.nn.functional as F from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .helpers import build_model_with_cfg -from timm.models.layers import DropPath, to_2tuple, trunc_normal_ -from timm.models.registry import register_model +from .layers import DropPath, to_2tuple, trunc_normal_, PatchEmbed, Mlp +from .registry import register_model +from .vision_transformer_hybrid import HybridEmbed import torch import torch.nn as nn @@ -29,7 +48,7 @@ def _cfg(url='', **kwargs): default_cfgs = { # ConViT 'convit_tiny': _cfg( - url="https://dl.fbaipublicfiles.com/convit/convit_tiny.pth"), + url="https://dl.fbaipublicfiles.com/convit/convit_tiny.pth"), 'convit_small': _cfg( url="https://dl.fbaipublicfiles.com/convit/convit_small.pth"), 'convit_base': _cfg( @@ -37,71 +56,31 @@ default_cfgs = { } -class Mlp(nn.Module): - def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) - self.drop = nn.Dropout(drop) - self.apply(self._init_weights) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - def forward(self, x): - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - x = self.fc2(x) - x = self.drop(x) - return x - - class GPSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., - locality_strength=1., use_local_init=True): + locality_strength=1.): super().__init__() self.num_heads = num_heads self.dim = dim head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 + self.locality_strength = locality_strength + + self.qk = nn.Linear(dim, dim * 2, bias=qkv_bias) + self.v = nn.Linear(dim, dim, bias=qkv_bias) - self.qk = nn.Linear(dim, dim * 2, bias=qkv_bias) - self.v = nn.Linear(dim, dim, bias=qkv_bias) - self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.pos_proj = nn.Linear(3, num_heads) self.proj_drop = nn.Dropout(proj_drop) self.locality_strength = locality_strength self.gating_param = nn.Parameter(torch.ones(self.num_heads)) - self.apply(self._init_weights) - if use_local_init: - self.local_init(locality_strength=locality_strength) + self.rel_indices: torch.Tensor = torch.zeros(1, 1, 1, 3) # silly torchscript hack, won't work with None - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - def forward(self, x): B, N, C = x.shape - if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: - self.get_rel_indices(N) - + if self.rel_indices is None or self.rel_indices.shape[1] != N: + self.rel_indices = self.get_rel_indices(N) attn = self.get_attention(x) v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) x = (attn @ v).transpose(1, 2).reshape(B, N, C) @@ -110,61 +89,58 @@ class GPSA(nn.Module): return x def get_attention(self, x): - B, N, C = x.shape + B, N, C = x.shape qk = self.qk(x).reshape(B, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k = qk[0], qk[1] - pos_score = self.rel_indices.expand(B, -1, -1,-1) - pos_score = self.pos_proj(pos_score).permute(0,3,1,2) + pos_score = self.rel_indices.expand(B, -1, -1, -1) + pos_score = self.pos_proj(pos_score).permute(0, 3, 1, 2) patch_score = (q @ k.transpose(-2, -1)) * self.scale patch_score = patch_score.softmax(dim=-1) pos_score = pos_score.softmax(dim=-1) - gating = self.gating_param.view(1,-1,1,1) - attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score + gating = self.gating_param.view(1, -1, 1, 1) + attn = (1. - torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score attn /= attn.sum(dim=-1).unsqueeze(-1) attn = self.attn_drop(attn) return attn - def get_attention_map(self, x, return_map = False): - - attn_map = self.get_attention(x).mean(0) # average over batch - distances = self.rel_indices.squeeze()[:,:,-1]**.5 - dist = torch.einsum('nm,hnm->h', (distances, attn_map)) - dist /= distances.size(0) + def get_attention_map(self, x, return_map=False): + attn_map = self.get_attention(x).mean(0) # average over batch + distances = self.rel_indices.squeeze()[:, :, -1] ** .5 + dist = torch.einsum('nm,hnm->h', (distances, attn_map)) / distances.size(0) if return_map: return dist, attn_map else: return dist - - def local_init(self, locality_strength=1.): - + + def local_init(self): self.v.weight.data.copy_(torch.eye(self.dim)) - locality_distance = 1 #max(1,1/locality_strength**.5) - - kernel_size = int(self.num_heads**.5) - center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 + locality_distance = 1 # max(1,1/locality_strength**.5) + + kernel_size = int(self.num_heads ** .5) + center = (kernel_size - 1) / 2 if kernel_size % 2 == 0 else kernel_size // 2 for h1 in range(kernel_size): for h2 in range(kernel_size): - position = h1+kernel_size*h2 - self.pos_proj.weight.data[position,2] = -1 - self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance - self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance - self.pos_proj.weight.data *= locality_strength - - def get_rel_indices(self, num_patches): - img_size = int(num_patches**.5) - rel_indices = torch.zeros(1, num_patches, num_patches, 3) - ind = torch.arange(img_size).view(1,-1) - torch.arange(img_size).view(-1, 1) - indx = ind.repeat(img_size,img_size) - indy = ind.repeat_interleave(img_size,dim=0).repeat_interleave(img_size,dim=1) - indd = indx**2 + indy**2 - rel_indices[:,:,:,2] = indd.unsqueeze(0) - rel_indices[:,:,:,1] = indy.unsqueeze(0) - rel_indices[:,:,:,0] = indx.unsqueeze(0) + position = h1 + kernel_size * h2 + self.pos_proj.weight.data[position, 2] = -1 + self.pos_proj.weight.data[position, 1] = 2 * (h1 - center) * locality_distance + self.pos_proj.weight.data[position, 0] = 2 * (h2 - center) * locality_distance + self.pos_proj.weight.data *= self.locality_strength + + def get_rel_indices(self, num_patches: int) -> torch.Tensor: + img_size = int(num_patches ** .5) + rel_indices = torch.zeros(1, num_patches, num_patches, 3) + ind = torch.arange(img_size).view(1, -1) - torch.arange(img_size).view(-1, 1) + indx = ind.repeat(img_size, img_size) + indy = ind.repeat_interleave(img_size, dim=0).repeat_interleave(img_size, dim=1) + indd = indx ** 2 + indy ** 2 + rel_indices[:, :, :, 2] = indd.unsqueeze(0) + rel_indices[:, :, :, 1] = indy.unsqueeze(0) + rel_indices[:, :, :, 0] = indx.unsqueeze(0) device = self.qk.weight.device - self.rel_indices = rel_indices.to(device) + return rel_indices.to(device) + - class MHSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() @@ -176,41 +152,28 @@ class MHSA(nn.Module): self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) - self.apply(self._init_weights) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - def get_attention_map(self, x, return_map = False): + def get_attention_map(self, x, return_map=False): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn_map = (q @ k.transpose(-2, -1)) * self.scale attn_map = attn_map.softmax(dim=-1).mean(0) - img_size = int(N**.5) - ind = torch.arange(img_size).view(1,-1) - torch.arange(img_size).view(-1, 1) - indx = ind.repeat(img_size,img_size) - indy = ind.repeat_interleave(img_size,dim=0).repeat_interleave(img_size,dim=1) - indd = indx**2 + indy**2 - distances = indd**.5 + img_size = int(N ** .5) + ind = torch.arange(img_size).view(1, -1) - torch.arange(img_size).view(-1, 1) + indx = ind.repeat(img_size, img_size) + indy = ind.repeat_interleave(img_size, dim=0).repeat_interleave(img_size, dim=1) + indd = indx ** 2 + indy ** 2 + distances = indd ** .5 distances = distances.to('cuda') - dist = torch.einsum('nm,hnm->h', (distances, attn_map)) - dist /= N - + dist = torch.einsum('nm,hnm->h', (distances, attn_map)) / N if return_map: return dist, attn_map else: return dist - def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) @@ -228,15 +191,19 @@ class MHSA(nn.Module): class Block(nn.Module): - def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): super().__init__() self.norm1 = norm_layer(dim) self.use_gpsa = use_gpsa if self.use_gpsa: - self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) + self.attn = GPSA( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, + proj_drop=drop, **kwargs) else: - self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) + self.attn = MHSA( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, + proj_drop=drop, **kwargs) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) @@ -246,75 +213,12 @@ class Block(nn.Module): x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x - - -class PatchEmbed(nn.Module): - """ Image to Patch Embedding, from timm - """ - def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) - self.img_size = img_size - self.patch_size = patch_size - self.num_patches = num_patches - - self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) - self.apply(self._init_weights) - - def forward(self, x): - B, C, H, W = x.shape - assert H == self.img_size[0] and W == self.img_size[1], \ - f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." - x = self.proj(x).flatten(2).transpose(1, 2) - return x - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - -class HybridEmbed(nn.Module): - """ CNN Feature Map Embedding, from timm - """ - def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768): - super().__init__() - assert isinstance(backbone, nn.Module) - img_size = to_2tuple(img_size) - self.img_size = img_size - self.backbone = backbone - if feature_size is None: - with torch.no_grad(): - training = backbone.training - if training: - backbone.eval() - o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1] - feature_size = o.shape[-2:] - feature_dim = o.shape[1] - backbone.train(training) - else: - feature_size = to_2tuple(feature_size) - feature_dim = self.backbone.feature_info.channels()[-1] - self.num_patches = feature_size[0] * feature_size[1] - self.proj = nn.Linear(feature_dim, embed_dim) - self.apply(self._init_weights) - - def forward(self, x): - x = self.backbone(x)[-1] - x = x.flatten(2).transpose(1, 2) - x = self.proj(x) - return x class ConViT(nn.Module): """ Vision Transformer with support for patch or hybrid CNN input stage """ + def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=nn.LayerNorm, global_pool=None, @@ -335,7 +239,7 @@ class ConViT(nn.Module): img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) num_patches = self.patch_embed.num_patches self.num_patches = num_patches - + self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.pos_drop = nn.Dropout(p=drop_rate) @@ -350,7 +254,7 @@ class ConViT(nn.Module): drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=True, locality_strength=locality_strength) - if i 0 else nn.Identity() trunc_normal_(self.cls_token, std=.02) - self.head.apply(self._init_weights) + self.apply(self._init_weights) + for n, m in self.named_modules(): + if hasattr(m, 'local_init'): + m.local_init() def _init_weights(self, m): if isinstance(m, nn.Linear): @@ -395,8 +302,8 @@ class ConViT(nn.Module): x = x + self.pos_embed x = self.pos_drop(x) - for u,blk in enumerate(self.blocks): - if u == self.local_up_to_layer : + for u, blk in enumerate(self.blocks): + if u == self.local_up_to_layer: x = torch.cat((cls_tokens, x), dim=1) x = blk(x) @@ -415,30 +322,29 @@ def _create_convit(variant, pretrained=False, **kwargs): default_cfg=default_cfgs[variant], **kwargs) - + @register_model def convit_tiny(pretrained=False, **kwargs): model_args = dict( - local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, num_heads=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) - model = _create_convit( - variant='convit_tiny', pretrained=pretrained, **model_args) + model = _create_convit(variant='convit_tiny', pretrained=pretrained, **model_args) return model + @register_model def convit_small(pretrained=False, **kwargs): model_args = dict( - local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, num_heads=9, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) - model = _create_convit( - variant='convit_small', pretrained=pretrained, **model_args) + model = _create_convit(variant='convit_small', pretrained=pretrained, **model_args) return model + @register_model def convit_base(pretrained=False, **kwargs): model_args = dict( - local_up_to_layer=10, locality_strength=1.0, embed_dim=48, + local_up_to_layer=10, locality_strength=1.0, embed_dim=48, num_heads=16, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) - model = _create_convit( - variant='convit_base', pretrained=pretrained, **model_args) + model = _create_convit(variant='convit_base', pretrained=pretrained, **model_args) return model From 30b9880d06a7f65edbd6a65aba4b6fca4c735060 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 21 May 2021 17:20:33 -0700 Subject: [PATCH 23/25] Minor adjustment, mutable default arg, extra check of valid len... --- timm/models/vision_transformer.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py index 1acdd808..bef6dfb0 100644 --- a/timm/models/vision_transformer.py +++ b/timm/models/vision_transformer.py @@ -352,7 +352,7 @@ def _init_vit_weights(m, n: str = '', head_bias: float = 0., jax_impl: bool = Fa nn.init.ones_(m.weight) -def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=[]): +def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()): # Rescale the grid of position embeddings when loading from state_dict. Adapted from # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224 _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape) @@ -363,8 +363,9 @@ def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=[]): else: posemb_tok, posemb_grid = posemb[:, :0], posemb[0] gs_old = int(math.sqrt(len(posemb_grid))) - if not len(gs_new): # backwards compatibility - gs_new = [int(math.sqrt(ntok_new))]*2 + if not len(gs_new): # backwards compatibility + gs_new = [int(math.sqrt(ntok_new))] * 2 + assert len(gs_new) >= 2 _logger.info('Position embedding grid-size from %s to %s', [gs_old, gs_old], gs_new) posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) posemb_grid = F.interpolate(posemb_grid, size=gs_new, mode='bilinear') From c2ba229d995c33aaaf20e00a5686b4dc857044be Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 21 May 2021 17:47:49 -0700 Subject: [PATCH 24/25] Prep for effcientnetv2_rw_m model weights that started training before official release.. --- timm/models/efficientnet.py | 16 ++++++++++++++-- timm/models/efficientnet_builder.py | 8 ++++++-- 2 files changed, 20 insertions(+), 4 deletions(-) diff --git a/timm/models/efficientnet.py b/timm/models/efficientnet.py index 0c0464b5..37c1c745 100644 --- a/timm/models/efficientnet.py +++ b/timm/models/efficientnet.py @@ -162,6 +162,9 @@ default_cfgs = { 'efficientnetv2_rw_s': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnet_v2s_ra2_288-a6477665.pth', input_size=(3, 288, 288), test_input_size=(3, 384, 384), pool_size=(9, 9), crop_pct=1.0), + 'efficientnetv2_rw_m': _cfg( + url='', + input_size=(3, 320, 320), test_input_size=(3, 416, 416), pool_size=(10, 10), crop_pct=1.0), 'efficientnetv2_s': _cfg( url='', @@ -173,7 +176,6 @@ default_cfgs = { url='', input_size=(3, 384, 384), test_input_size=(3, 480, 480), pool_size=(12, 12), crop_pct=1.0), - 'tf_efficientnet_b0': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b0_aa-827b6e33.pth', input_size=(3, 224, 224)), @@ -1461,7 +1463,7 @@ def efficientnet_b3_pruned(pretrained=False, **kwargs): @register_model def efficientnetv2_rw_s(pretrained=False, **kwargs): - """ EfficientNet-V2 Small. + """ EfficientNet-V2 Small RW variant. NOTE: This is my initial (pre official code release) w/ some differences. See efficientnetv2_s and tf_efficientnetv2_s for versions that match the official w/ PyTorch vs TF padding """ @@ -1469,6 +1471,16 @@ def efficientnetv2_rw_s(pretrained=False, **kwargs): return model +@register_model +def efficientnetv2_rw_m(pretrained=False, **kwargs): + """ EfficientNet-V2 Medium RW variant. + """ + model = _gen_efficientnetv2_s( + 'efficientnetv2_rw_m', channel_multiplier=1.2, depth_multiplier=(1.2,) * 4 + (1.6,) * 2, rw=True, + pretrained=pretrained, **kwargs) + return model + + @register_model def efficientnetv2_s(pretrained=False, **kwargs): """ EfficientNet-V2 Small. """ diff --git a/timm/models/efficientnet_builder.py b/timm/models/efficientnet_builder.py index 9d5853c7..30739454 100644 --- a/timm/models/efficientnet_builder.py +++ b/timm/models/efficientnet_builder.py @@ -237,7 +237,11 @@ def _scale_stage_depth(stack_args, repeats, depth_multiplier=1.0, depth_trunc='c def decode_arch_def(arch_def, depth_multiplier=1.0, depth_trunc='ceil', experts_multiplier=1, fix_first_last=False): arch_args = [] - for stack_idx, block_strings in enumerate(arch_def): + if isinstance(depth_multiplier, tuple): + assert len(depth_multiplier) == len(arch_def) + else: + depth_multiplier = (depth_multiplier,) * len(arch_def) + for stack_idx, (block_strings, multiplier) in enumerate(zip(arch_def, depth_multiplier)): assert isinstance(block_strings, list) stack_args = [] repeats = [] @@ -251,7 +255,7 @@ def decode_arch_def(arch_def, depth_multiplier=1.0, depth_trunc='ceil', experts_ if fix_first_last and (stack_idx == 0 or stack_idx == len(arch_def) - 1): arch_args.append(_scale_stage_depth(stack_args, repeats, 1.0, depth_trunc)) else: - arch_args.append(_scale_stage_depth(stack_args, repeats, depth_multiplier, depth_trunc)) + arch_args.append(_scale_stage_depth(stack_args, repeats, multiplier, depth_trunc)) return arch_args From 23c18a33e4168dc7cb11439c1f9acd38dc8e9824 Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Fri, 21 May 2021 21:16:25 -0700 Subject: [PATCH 25/25] Add efficientnetv2_rw_m weights trained in PyTorch. 84.8 top-1 @ 416 test. 53M params. --- timm/models/efficientnet.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/timm/models/efficientnet.py b/timm/models/efficientnet.py index 37c1c745..8aa61ec5 100644 --- a/timm/models/efficientnet.py +++ b/timm/models/efficientnet.py @@ -163,7 +163,7 @@ default_cfgs = { url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnet_v2s_ra2_288-a6477665.pth', input_size=(3, 288, 288), test_input_size=(3, 384, 384), pool_size=(9, 9), crop_pct=1.0), 'efficientnetv2_rw_m': _cfg( - url='', + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnetv2_rw_m_agc-3d90cb1e.pth', input_size=(3, 320, 320), test_input_size=(3, 416, 416), pool_size=(10, 10), crop_pct=1.0), 'efficientnetv2_s': _cfg(