diff --git a/tests/test_models.py b/tests/test_models.py index 50838c8a..c0d0e901 100644 --- a/tests/test_models.py +++ b/tests/test_models.py @@ -17,7 +17,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'): # 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_*', 'twins_*', - 'convit_*', 'levit*', 'visformer*', 'deit*', 'jx_nest_*', 'nest_*', 'xcit_*'] + 'convit_*', 'levit*', 'visformer*', 'deit*', 'jx_nest_*', 'nest_*', 'xcit_*', 'crossvit_*', 'beit_*'] NUM_NON_STD = len(NON_STD_FILTERS) # exclude models that cause specific test failures @@ -188,23 +188,22 @@ def test_model_default_cfgs_non_std(model_name, batch_size): input_tensor = torch.randn((batch_size, *input_size)) - # test forward_features (always unpooled) outputs = model.forward_features(input_tensor) - if isinstance(outputs, tuple): + if isinstance(outputs, (tuple, list)): outputs = outputs[0] assert outputs.shape[1] == model.num_features # test forward after deleting the classifier, output should be poooled, size(-1) == model.num_features model.reset_classifier(0) outputs = model.forward(input_tensor) - if isinstance(outputs, tuple): + if isinstance(outputs, (tuple, list)): outputs = outputs[0] assert len(outputs.shape) == 2 assert outputs.shape[1] == model.num_features model = create_model(model_name, pretrained=False, num_classes=0).eval() outputs = model.forward(input_tensor) - if isinstance(outputs, tuple): + if isinstance(outputs, (tuple, list)): outputs = outputs[0] assert len(outputs.shape) == 2 assert outputs.shape[1] == model.num_features diff --git a/tests/test_optim.py b/tests/test_optim.py index 41e6d5e9..737674e5 100644 --- a/tests/test_optim.py +++ b/tests/test_optim.py @@ -319,10 +319,10 @@ def test_sgd(optimizer): # lambda opt: ReduceLROnPlateau(opt)] # ) _test_basic_cases( - lambda weight, bias: create_optimizer_v2([weight, bias], optimizer, lr=1e-3, momentum=1) + lambda weight, bias: create_optimizer_v2([weight, bias], optimizer, lr=3e-3, momentum=1) ) _test_basic_cases( - lambda weight, bias: create_optimizer_v2([weight, bias], optimizer, lr=1e-3, momentum=1, weight_decay=.1) + lambda weight, bias: create_optimizer_v2([weight, bias], optimizer, lr=3e-3, momentum=1, weight_decay=.1) ) _test_rosenbrock( lambda params: create_optimizer_v2(params, optimizer, lr=1e-3) diff --git a/timm/models/__init__.py b/timm/models/__init__.py index 56c812d4..56a753b1 100644 --- a/timm/models/__init__.py +++ b/timm/models/__init__.py @@ -1,8 +1,10 @@ +from .beit import * from .byoanet import * from .byobnet import * from .cait import * from .coat import * from .convit import * +from .crossvit import * from .cspnet import * from .densenet import * from .dla import * @@ -36,6 +38,7 @@ from .sknet import * from .swin_transformer import * from .tnt import * from .tresnet import * +from .twins import * from .vgg import * from .visformer import * from .vision_transformer import * @@ -44,7 +47,6 @@ from .vovnet import * from .xception import * from .xception_aligned import * from .xcit 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/beit.py b/timm/models/beit.py new file mode 100644 index 00000000..e8d1dd2c --- /dev/null +++ b/timm/models/beit.py @@ -0,0 +1,420 @@ +""" BEIT: BERT Pre-Training of Image Transformers (https://arxiv.org/abs/2106.08254) + +Model from official source: https://github.com/microsoft/unilm/tree/master/beit + +At this point only the 1k fine-tuned classification weights and model configs have been added, +see original source above for pre-training models and procedure. + +Modifications by / Copyright 2021 Ross Wightman, original copyrights below +""" +# -------------------------------------------------------- +# BEIT: BERT Pre-Training of Image Transformers (https://arxiv.org/abs/2106.08254) +# Github source: https://github.com/microsoft/unilm/tree/master/beit +# Copyright (c) 2021 Microsoft +# Licensed under The MIT License [see LICENSE for details] +# By Hangbo Bao +# Based on timm and DeiT code bases +# https://github.com/rwightman/pytorch-image-models/tree/master/timm +# https://github.com/facebookresearch/deit/ +# https://github.com/facebookresearch/dino +# --------------------------------------------------------' +import math +from functools import partial +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .helpers import build_model_with_cfg +from .layers import PatchEmbed, Mlp, DropPath, trunc_normal_ +from .registry import register_model +from .vision_transformer import checkpoint_filter_fn + + +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': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5), + 'first_conv': 'patch_embed.proj', 'classifier': 'head', + **kwargs + } + + +default_cfgs = { + 'beit_base_patch16_224': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_base_patch16_224_pt22k_ft22kto1k.pth'), + 'beit_base_patch16_384': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_base_patch16_384_pt22k_ft22kto1k.pth', + input_size=(3, 384, 384), crop_pct=1.0, + ), + 'beit_base_patch16_224_in22k': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_base_patch16_224_pt22k_ft22k.pth', + num_classes=21841, + ), + 'beit_large_patch16_224': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_large_patch16_224_pt22k_ft22kto1k.pth'), + 'beit_large_patch16_384': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_large_patch16_384_pt22k_ft22kto1k.pth', + input_size=(3, 384, 384), crop_pct=1.0, + ), + 'beit_large_patch16_512': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_large_patch16_512_pt22k_ft22kto1k.pth', + input_size=(3, 512, 512), crop_pct=1.0, + ), + 'beit_large_patch16_224_in22k': _cfg( + url='https://unilm.blob.core.windows.net/beit/beit_large_patch16_224_pt22k_ft22k.pth', + num_classes=21841, + ), +} + + +class Attention(nn.Module): + def __init__( + self, dim, num_heads=8, qkv_bias=False, attn_drop=0., + proj_drop=0., window_size=None, attn_head_dim=None): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + if attn_head_dim is not None: + head_dim = attn_head_dim + all_head_dim = head_dim * self.num_heads + self.scale = head_dim ** -0.5 + + self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False) + if qkv_bias: + self.q_bias = nn.Parameter(torch.zeros(all_head_dim)) + self.v_bias = nn.Parameter(torch.zeros(all_head_dim)) + else: + self.q_bias = None + self.v_bias = None + + if window_size: + self.window_size = window_size + self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3 + self.relative_position_bias_table = nn.Parameter( + torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH + # cls to token & token 2 cls & cls to cls + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(window_size[0]) + coords_w = torch.arange(window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * window_size[1] - 1 + relative_position_index = \ + torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype) + relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + relative_position_index[0, 0:] = self.num_relative_distance - 3 + relative_position_index[0:, 0] = self.num_relative_distance - 2 + relative_position_index[0, 0] = self.num_relative_distance - 1 + + self.register_buffer("relative_position_index", relative_position_index) + else: + self.window_size = None + self.relative_position_bias_table = None + self.relative_position_index = None + + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(all_head_dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x, rel_pos_bias: Optional[torch.Tensor] = None): + B, N, C = x.shape + qkv_bias = None + if self.q_bias is not None: + if torch.jit.is_scripting(): + # FIXME requires_grad breaks w/ torchscript + qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias), self.v_bias)) + else: + qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias)) + qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias) + qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + if self.relative_position_bias_table is not None: + relative_position_bias = \ + self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1] + 1, + self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if rel_pos_bias is not None: + attn = attn + rel_pos_bias + + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B, N, -1) + 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, drop=0., attn_drop=0., + drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm, + window_size=None, attn_head_dim=None): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop, + window_size=window_size, attn_head_dim=attn_head_dim) + # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here + 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) + + if init_values: + self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)), requires_grad=True) + self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)), requires_grad=True) + else: + self.gamma_1, self.gamma_2 = None, None + + def forward(self, x, rel_pos_bias: Optional[torch.Tensor] = None): + if self.gamma_1 is None: + x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + else: + x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias)) + x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x))) + return x + + +class RelativePositionBias(nn.Module): + + def __init__(self, window_size, num_heads): + super().__init__() + self.window_size = window_size + self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3 + self.relative_position_bias_table = nn.Parameter( + torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH + # cls to token & token 2 cls & cls to cls + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(window_size[0]) + coords_w = torch.arange(window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * window_size[1] - 1 + relative_position_index = \ + torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype) + relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + relative_position_index[0, 0:] = self.num_relative_distance - 3 + relative_position_index[0:, 0] = self.num_relative_distance - 2 + relative_position_index[0, 0] = self.num_relative_distance - 1 + + self.register_buffer("relative_position_index", relative_position_index) + + # trunc_normal_(self.relative_position_bias_table, std=.02) + + def forward(self): + relative_position_bias = \ + self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1] + 1, + self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH + return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + + +class Beit(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=True, drop_rate=0., attn_drop_rate=0., + drop_path_rate=0., norm_layer=partial(nn.LayerNorm, eps=1e-6), init_values=None, + use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False, + use_mean_pooling=True, init_scale=0.001): + super().__init__() + self.num_classes = num_classes + self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models + + 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.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + if use_abs_pos_emb: + self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) + else: + self.pos_embed = None + self.pos_drop = nn.Dropout(p=drop_rate) + + if use_shared_rel_pos_bias: + self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.grid_size, num_heads=num_heads) + else: + self.rel_pos_bias = None + + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule + self.use_rel_pos_bias = use_rel_pos_bias + self.blocks = nn.ModuleList([ + Block( + dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, + drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, + init_values=init_values, window_size=self.patch_embed.grid_size if use_rel_pos_bias else None) + for i in range(depth)]) + self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim) + self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None + self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity() + + self.apply(self._init_weights) + if self.pos_embed is not None: + trunc_normal_(self.pos_embed, std=.02) + trunc_normal_(self.cls_token, std=.02) + # trunc_normal_(self.mask_token, std=.02) + self.fix_init_weight() + if isinstance(self.head, nn.Linear): + trunc_normal_(self.head.weight, std=.02) + self.head.weight.data.mul_(init_scale) + self.head.bias.data.mul_(init_scale) + + def fix_init_weight(self): + def rescale(param, layer_id): + param.div_(math.sqrt(2.0 * layer_id)) + + for layer_id, layer in enumerate(self.blocks): + rescale(layer.attn.proj.weight.data, layer_id + 1) + rescale(layer.mlp.fc2.weight.data, layer_id + 1) + + 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_num_layers(self): + return len(self.blocks) + + @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): + x = self.patch_embed(x) + batch_size, seq_len, _ = x.size() + + cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks + x = torch.cat((cls_tokens, x), dim=1) + if self.pos_embed is not None: + x = x + self.pos_embed + x = self.pos_drop(x) + + rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None + for blk in self.blocks: + x = blk(x, rel_pos_bias=rel_pos_bias) + + x = self.norm(x) + if self.fc_norm is not None: + t = x[:, 1:, :] + return self.fc_norm(t.mean(1)) + else: + return x[:, 0] + + def forward(self, x): + x = self.forward_features(x) + x = self.head(x) + return x + + +def _create_beit(variant, pretrained=False, default_cfg=None, **kwargs): + default_cfg = default_cfg or default_cfgs[variant] + if kwargs.get('features_only', None): + raise RuntimeError('features_only not implemented for Beit models.') + + model = build_model_with_cfg( + Beit, variant, pretrained, + default_cfg=default_cfg, + # FIXME an updated filter fn needed to interpolate rel pos emb if fine tuning to diff model sizes + pretrained_filter_fn=checkpoint_filter_fn, + **kwargs) + return model + + +@register_model +def beit_base_patch16_224(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=0.1, **kwargs) + model = _create_beit('beit_base_patch16_224', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_base_patch16_384(pretrained=False, **kwargs): + model_kwargs = dict( + img_size=384, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=0.1, **kwargs) + model = _create_beit('beit_base_patch16_384', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_base_patch16_224_in22k(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=0.1, **kwargs) + model = _create_beit('beit_base_patch16_224_in22k', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_large_patch16_224(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=1e-5, **kwargs) + model = _create_beit('beit_large_patch16_224', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_large_patch16_384(pretrained=False, **kwargs): + model_kwargs = dict( + img_size=384, patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=1e-5, **kwargs) + model = _create_beit('beit_large_patch16_384', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_large_patch16_512(pretrained=False, **kwargs): + model_kwargs = dict( + img_size=512, patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=1e-5, **kwargs) + model = _create_beit('beit_large_patch16_512', pretrained=pretrained, **model_kwargs) + return model + + +@register_model +def beit_large_patch16_224_in22k(pretrained=False, **kwargs): + model_kwargs = dict( + patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + use_abs_pos_emb=False, use_rel_pos_bias=True, init_values=1e-5, **kwargs) + model = _create_beit('beit_large_patch16_224_in22k', pretrained=pretrained, **model_kwargs) + return model diff --git a/timm/models/crossvit.py b/timm/models/crossvit.py new file mode 100644 index 00000000..6e0160f9 --- /dev/null +++ b/timm/models/crossvit.py @@ -0,0 +1,497 @@ +""" CrossViT Model + +@inproceedings{ + chen2021crossvit, + title={{CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification}}, + author={Chun-Fu (Richard) Chen and Quanfu Fan and Rameswar Panda}, + booktitle={International Conference on Computer Vision (ICCV)}, + year={2021} +} + +Paper link: https://arxiv.org/abs/2103.14899 +Original code: https://github.com/IBM/CrossViT/blob/main/models/crossvit.py + +NOTE: model names have been renamed from originals to represent actual input res all *_224 -> *_240 and *_384 -> *_408 +""" + +# Copyright IBM All Rights Reserved. +# SPDX-License-Identifier: Apache-2.0 + + +""" +Modifed from Timm. https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py + +""" + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.hub +from functools import partial +from typing import List + +from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD +from .helpers import build_model_with_cfg +from .layers import DropPath, to_2tuple, trunc_normal_ +from .registry import register_model +from .vision_transformer import Mlp, Block + + +def _cfg(url='', **kwargs): + return { + 'url': url, + 'num_classes': 1000, 'input_size': (3, 240, 240), 'pool_size': None, 'crop_pct': 0.875, + 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, 'fixed_input_size': True, + 'first_conv': ('patch_embed.0.proj', 'patch_embed.1.proj'), + 'classifier': ('head.0', 'head.1'), + **kwargs + } + + +default_cfgs = { + 'crossvit_15_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_224.pth'), + 'crossvit_15_dagger_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_dagger_224.pth', + first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), + ), + 'crossvit_15_dagger_408': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_dagger_384.pth', + input_size=(3, 408, 408), first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), crop_pct=1.0, + ), + 'crossvit_18_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_224.pth'), + 'crossvit_18_dagger_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_dagger_224.pth', + first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), + ), + 'crossvit_18_dagger_408': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_dagger_384.pth', + input_size=(3, 408, 408), first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), crop_pct=1.0, + ), + 'crossvit_9_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_9_224.pth'), + 'crossvit_9_dagger_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_9_dagger_224.pth', + first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), + ), + 'crossvit_base_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_base_224.pth'), + 'crossvit_small_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_small_224.pth'), + 'crossvit_tiny_240': _cfg( + url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_tiny_224.pth'), +} + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + """ + + def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, multi_conv=False): + 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 + if multi_conv: + if patch_size[0] == 12: + self.proj = nn.Sequential( + nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3), + nn.ReLU(inplace=True), + nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=3, padding=0), + nn.ReLU(inplace=True), + nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=1, padding=1), + ) + elif patch_size[0] == 16: + self.proj = nn.Sequential( + nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3), + nn.ReLU(inplace=True), + nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=2, padding=1), + nn.ReLU(inplace=True), + nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=2, padding=1), + ) + else: + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + + def forward(self, x): + B, C, H, W = x.shape + # FIXME look at relaxing size constraints + 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 + + +class CrossAttention(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 + # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights + self.scale = qk_scale or head_dim ** -0.5 + + self.wq = nn.Linear(dim, dim, bias=qkv_bias) + self.wk = nn.Linear(dim, dim, bias=qkv_bias) + self.wv = nn.Linear(dim, dim, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x): + B, N, C = x.shape + # B1C -> B1H(C/H) -> BH1(C/H) + q = self.wq(x[:, 0:1, ...]).reshape(B, 1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + # BNC -> BNH(C/H) -> BHN(C/H) + k = self.wk(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + # BNC -> BNH(C/H) -> BHN(C/H) + v = self.wv(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + + attn = (q @ k.transpose(-2, -1)) * self.scale # BH1(C/H) @ BH(C/H)N -> BH1N + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B, 1, C) # (BH1N @ BHN(C/H)) -> BH1(C/H) -> B1H(C/H) -> B1C + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class CrossAttentionBlock(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): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = CrossAttention( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + + def forward(self, x): + x = x[:, 0:1, ...] + self.drop_path(self.attn(self.norm1(x))) + + return x + + +class MultiScaleBlock(nn.Module): + + def __init__(self, dim, patches, depth, num_heads, mlp_ratio, qkv_bias=False, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + + num_branches = len(dim) + self.num_branches = num_branches + # different branch could have different embedding size, the first one is the base + self.blocks = nn.ModuleList() + for d in range(num_branches): + tmp = [] + for i in range(depth[d]): + tmp.append(Block( + dim=dim[d], num_heads=num_heads[d], mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias, + drop=drop, attn_drop=attn_drop, drop_path=drop_path[i], norm_layer=norm_layer)) + if len(tmp) != 0: + self.blocks.append(nn.Sequential(*tmp)) + + if len(self.blocks) == 0: + self.blocks = None + + self.projs = nn.ModuleList() + for d in range(num_branches): + if dim[d] == dim[(d + 1) % num_branches] and False: + tmp = [nn.Identity()] + else: + tmp = [norm_layer(dim[d]), act_layer(), nn.Linear(dim[d], dim[(d + 1) % num_branches])] + self.projs.append(nn.Sequential(*tmp)) + + self.fusion = nn.ModuleList() + for d in range(num_branches): + d_ = (d + 1) % num_branches + nh = num_heads[d_] + if depth[-1] == 0: # backward capability: + self.fusion.append( + CrossAttentionBlock( + dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias, + drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer)) + else: + tmp = [] + for _ in range(depth[-1]): + tmp.append(CrossAttentionBlock( + dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias, + drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer)) + self.fusion.append(nn.Sequential(*tmp)) + + self.revert_projs = nn.ModuleList() + for d in range(num_branches): + if dim[(d + 1) % num_branches] == dim[d] and False: + tmp = [nn.Identity()] + else: + tmp = [norm_layer(dim[(d + 1) % num_branches]), act_layer(), + nn.Linear(dim[(d + 1) % num_branches], dim[d])] + self.revert_projs.append(nn.Sequential(*tmp)) + + def forward(self, x: List[torch.Tensor]) -> List[torch.Tensor]: + + outs_b = [] + for i, block in enumerate(self.blocks): + outs_b.append(block(x[i])) + + # only take the cls token out + proj_cls_token = torch.jit.annotate(List[torch.Tensor], []) + for i, proj in enumerate(self.projs): + proj_cls_token.append(proj(outs_b[i][:, 0:1, ...])) + + # cross attention + outs = [] + for i, (fusion, revert_proj) in enumerate(zip(self.fusion, self.revert_projs)): + tmp = torch.cat((proj_cls_token[i], outs_b[(i + 1) % self.num_branches][:, 1:, ...]), dim=1) + tmp = fusion(tmp) + reverted_proj_cls_token = revert_proj(tmp[:, 0:1, ...]) + tmp = torch.cat((reverted_proj_cls_token, outs_b[i][:, 1:, ...]), dim=1) + outs.append(tmp) + return outs + + +def _compute_num_patches(img_size, patches): + return [i[0] // p * i[1] // p for i, p in zip(img_size, patches)] + + +class CrossViT(nn.Module): + """ Vision Transformer with support for patch or hybrid CNN input stage + """ + + def __init__( + self, img_size=224, img_scale=(1.0, 1.0), patch_size=(8, 16), in_chans=3, num_classes=1000, + embed_dim=(192, 384), depth=((1, 3, 1), (1, 3, 1), (1, 3, 1)), num_heads=(6, 12), mlp_ratio=(2., 2., 4.), + qkv_bias=True, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., + norm_layer=partial(nn.LayerNorm, eps=1e-6), multi_conv=False, crop_scale=False, + ): + super().__init__() + + self.num_classes = num_classes + self.img_size = to_2tuple(img_size) + img_scale = to_2tuple(img_scale) + self.img_size_scaled = [tuple([int(sj * si) for sj in self.img_size]) for si in img_scale] + self.crop_scale = crop_scale # crop instead of interpolate for scale + num_patches = _compute_num_patches(self.img_size_scaled, patch_size) + self.num_branches = len(patch_size) + self.embed_dim = embed_dim + self.num_features = embed_dim[0] # to pass the tests + self.patch_embed = nn.ModuleList() + + # hard-coded for torch jit script + for i in range(self.num_branches): + setattr(self, f'pos_embed_{i}', nn.Parameter(torch.zeros(1, 1 + num_patches[i], embed_dim[i]))) + setattr(self, f'cls_token_{i}', nn.Parameter(torch.zeros(1, 1, embed_dim[i]))) + + for im_s, p, d in zip(self.img_size_scaled, patch_size, embed_dim): + self.patch_embed.append( + PatchEmbed(img_size=im_s, patch_size=p, in_chans=in_chans, embed_dim=d, multi_conv=multi_conv)) + + self.pos_drop = nn.Dropout(p=drop_rate) + + total_depth = sum([sum(x[-2:]) for x in depth]) + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, total_depth)] # stochastic depth decay rule + dpr_ptr = 0 + self.blocks = nn.ModuleList() + for idx, block_cfg in enumerate(depth): + curr_depth = max(block_cfg[:-1]) + block_cfg[-1] + dpr_ = dpr[dpr_ptr:dpr_ptr + curr_depth] + blk = MultiScaleBlock( + embed_dim, num_patches, block_cfg, num_heads=num_heads, mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr_, norm_layer=norm_layer) + dpr_ptr += curr_depth + self.blocks.append(blk) + + self.norm = nn.ModuleList([norm_layer(embed_dim[i]) for i in range(self.num_branches)]) + self.head = nn.ModuleList([ + nn.Linear(embed_dim[i], num_classes) if num_classes > 0 else nn.Identity() + for i in range(self.num_branches)]) + + for i in range(self.num_branches): + trunc_normal_(getattr(self, f'pos_embed_{i}'), std=.02) + trunc_normal_(getattr(self, f'cls_token_{i}'), std=.02) + + 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) + + @torch.jit.ignore + def no_weight_decay(self): + out = set() + for i in range(self.num_branches): + out.add(f'cls_token_{i}') + pe = getattr(self, f'pos_embed_{i}', None) + if pe is not None and pe.requires_grad: + out.add(f'pos_embed_{i}') + return out + + def get_classifier(self): + return self.head + + def reset_classifier(self, num_classes, global_pool=''): + self.num_classes = num_classes + self.head = nn.ModuleList( + [nn.Linear(self.embed_dim[i], num_classes) if num_classes > 0 else nn.Identity() for i in + range(self.num_branches)]) + + def forward_features(self, x): + B, C, H, W = x.shape + xs = [] + for i, patch_embed in enumerate(self.patch_embed): + x_ = x + ss = self.img_size_scaled[i] + if H != ss[0] or W != ss[1]: + if self.crop_scale and ss[0] <= H and ss[1] <= W: + cu, cl = int(round((H - ss[0]) / 2.)), int(round((W - ss[1]) / 2.)) + x_ = x_[:, :, cu:cu + ss[0], cl:cl + ss[1]] + else: + x_ = torch.nn.functional.interpolate(x_, size=ss, mode='bicubic', align_corners=False) + x_ = patch_embed(x_) + cls_tokens = self.cls_token_0 if i == 0 else self.cls_token_1 # hard-coded for torch jit script + cls_tokens = cls_tokens.expand(B, -1, -1) + x_ = torch.cat((cls_tokens, x_), dim=1) + pos_embed = self.pos_embed_0 if i == 0 else self.pos_embed_1 # hard-coded for torch jit script + x_ = x_ + pos_embed + x_ = self.pos_drop(x_) + xs.append(x_) + + for i, blk in enumerate(self.blocks): + xs = blk(xs) + + # NOTE: was before branch token section, move to here to assure all branch token are before layer norm + xs = [norm(xs[i]) for i, norm in enumerate(self.norm)] + return [xo[:, 0] for xo in xs] + + def forward(self, x): + xs = self.forward_features(x) + ce_logits = [head(xs[i]) for i, head in enumerate(self.head)] + if not isinstance(self.head[0], nn.Identity): + ce_logits = torch.mean(torch.stack(ce_logits, dim=0), dim=0) + return ce_logits + + +def _create_crossvit(variant, pretrained=False, **kwargs): + if kwargs.get('features_only', None): + raise RuntimeError('features_only not implemented for Vision Transformer models.') + + def pretrained_filter_fn(state_dict): + new_state_dict = {} + for key in state_dict.keys(): + if 'pos_embed' in key or 'cls_token' in key: + new_key = key.replace(".", "_") + else: + new_key = key + new_state_dict[new_key] = state_dict[key] + return new_state_dict + + return build_model_with_cfg( + CrossViT, variant, pretrained, + default_cfg=default_cfgs[variant], + pretrained_filter_fn=pretrained_filter_fn, + **kwargs) + + +@register_model +def crossvit_tiny_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[96, 192], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]], + num_heads=[3, 3], mlp_ratio=[4, 4, 1], **kwargs) + model = _create_crossvit(variant='crossvit_tiny_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_small_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]], + num_heads=[6, 6], mlp_ratio=[4, 4, 1], **kwargs) + model = _create_crossvit(variant='crossvit_small_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_base_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[384, 768], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]], + num_heads=[12, 12], mlp_ratio=[4, 4, 1], **kwargs) + model = _create_crossvit(variant='crossvit_base_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_9_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[128, 256], depth=[[1, 3, 0], [1, 3, 0], [1, 3, 0]], + num_heads=[4, 4], mlp_ratio=[3, 3, 1], **kwargs) + model = _create_crossvit(variant='crossvit_9_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_15_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]], + num_heads=[6, 6], mlp_ratio=[3, 3, 1], **kwargs) + model = _create_crossvit(variant='crossvit_15_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_18_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224 / 240), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]], + num_heads=[7, 7], mlp_ratio=[3, 3, 1], **kwargs) + model = _create_crossvit(variant='crossvit_18_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_9_dagger_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224 / 240), patch_size=[12, 16], embed_dim=[128, 256], depth=[[1, 3, 0], [1, 3, 0], [1, 3, 0]], + num_heads=[4, 4], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs) + model = _create_crossvit(variant='crossvit_9_dagger_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_15_dagger_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]], + num_heads=[6, 6], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs) + model = _create_crossvit(variant='crossvit_15_dagger_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_15_dagger_408(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 384/408), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]], + num_heads=[6, 6], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs) + model = _create_crossvit(variant='crossvit_15_dagger_408', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_18_dagger_240(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]], + num_heads=[7, 7], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs) + model = _create_crossvit(variant='crossvit_18_dagger_240', pretrained=pretrained, **model_args) + return model + + +@register_model +def crossvit_18_dagger_408(pretrained=False, **kwargs): + model_args = dict( + img_scale=(1.0, 384/408), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]], + num_heads=[7, 7], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs) + model = _create_crossvit(variant='crossvit_18_dagger_408', pretrained=pretrained, **model_args) + return model diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py index de8248fe..ca8f52de 100644 --- a/timm/models/vision_transformer.py +++ b/timm/models/vision_transformer.py @@ -683,7 +683,7 @@ def vit_large_patch16_384(pretrained=False, **kwargs): def vit_base_patch16_sam_224(pretrained=False, **kwargs): """ ViT-Base (ViT-B/16) w/ SAM pretrained weights. Paper: https://arxiv.org/abs/2106.01548 """ - # NOTE original SAM weights releaes worked with representation_size=768 + # NOTE original SAM weights release worked with representation_size=768 model_kwargs = dict(patch_size=16, embed_dim=768, depth=12, num_heads=12, representation_size=0, **kwargs) model = _create_vision_transformer('vit_base_patch16_sam_224', pretrained=pretrained, **model_kwargs) return model @@ -693,7 +693,7 @@ def vit_base_patch16_sam_224(pretrained=False, **kwargs): def vit_base_patch32_sam_224(pretrained=False, **kwargs): """ ViT-Base (ViT-B/32) w/ SAM pretrained weights. Paper: https://arxiv.org/abs/2106.01548 """ - # NOTE original SAM weights releaes worked with representation_size=768 + # NOTE original SAM weights release worked with representation_size=768 model_kwargs = dict(patch_size=32, embed_dim=768, depth=12, num_heads=12, representation_size=0, **kwargs) model = _create_vision_transformer('vit_base_patch32_sam_224', pretrained=pretrained, **model_kwargs) return model