From 736f209e7d7ebe9a6ac9acf9967a7aba0a86aa4e Mon Sep 17 00:00:00 2001 From: Ross Wightman Date: Mon, 26 Oct 2020 18:42:11 -0700 Subject: [PATCH] Update vision transformers to be compatible with official code. Port official ViT weights from jax impl. --- README.md | 8 ++ timm/models/vision_transformer.py | 184 +++++++++++++++++------------- 2 files changed, 112 insertions(+), 80 deletions(-) diff --git a/README.md b/README.md index a89c58ec..5bc8322c 100644 --- a/README.md +++ b/README.md @@ -2,6 +2,14 @@ ## What's New +### Oct 26, 2020 +* Update Vision Transformer models to be compatible with official code release at https://github.com/google-research/vision_transformer +* Add Vision Transformer weights (ImageNet-21k pretrain) for 384x384 base and large models converted from official jax impl + * ViT-B/16 - 84.2 + * ViT-B/32 - 81.7 + * ViT-L/16 - 85.2 + * ViT-L/32 - 81.5 + ### Oct 21, 2020 * Weights added for Vision Transformer (ViT) models. 77.86 top-1 for 'small' and 79.35 for 'base'. Thanks to [Christof](https://www.kaggle.com/christofhenkel) for training the base model w/ lots of GPUs. diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py index 1e57c095..6cefda28 100644 --- a/timm/models/vision_transformer.py +++ b/timm/models/vision_transformer.py @@ -1,23 +1,18 @@ """ Vision Transformer (ViT) in PyTorch -This is a WIP attempt to implement Vision Transformers as described in -'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale' - -https://openreview.net/pdf?id=YicbFdNTTy +A PyTorch implement of Vision Transformers as described in +'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale' - https://arxiv.org/abs/2010.11929 -The paper is currently under review and there is no official reference impl. The -code here is likely to change in the future and I will not make an effort to maintain -backwards weight compatibility when it does. +The official jax code is released and available at https://github.com/google-research/vision_transformer Status/TODO: -* Trained (supervised on ImageNet-1k) my custom 'small' patch model to ~75 top-1 after 4 days, 2x GPU, -no dropout or stochastic depth active -* Need more time for supervised training results with dropout and drop connect active, hparam tuning -* Need more GPUs for SSL or unsupervised pretraining on OpenImages w/ ImageNet fine-tune -* There are likely mistakes. If you notice any, I'd love to improve this. This is my first time -fiddling with transformers/multi-head attn. -* Hopefully end up with worthwhile pretrained model at some point... +* Models updated to be compatible with official impl. Args added to support backward compat for old PyTorch weights. +* Weights ported from official jax impl for 384x384 base and small models, 16x16 and 32x32 patches. +* Trained (supervised on ImageNet-1k) my custom 'small' patch model to 77.9, 'base' to 79.4 top-1 with this code. +* Hopefully find time and GPUs for SSL or unsupervised pretraining on OpenImages w/ ImageNet fine-tune in future. Acknowledgments: +* The paper authors for releasing code and weights, thanks! * I fixed my class token impl based on Phil Wang's https://github.com/lucidrains/vit-pytorch ... check it out for some einops/einsum fun * Simple transformer style inspired by Andrej Karpathy's https://github.com/karpathy/minGPT @@ -27,6 +22,7 @@ Hacked together by / Copyright 2020 Ross Wightman """ import torch import torch.nn as nn +from functools import partial from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .helpers import load_pretrained @@ -52,13 +48,21 @@ default_cfgs = { url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth', ), 'vit_base_patch16_224': _cfg( - url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_base_p16_224-4e355ebd.pth' + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_base_p16_224-4e355ebd.pth', ), - 'vit_base_patch16_384': _cfg(input_size=(3, 384, 384)), - 'vit_base_patch32_384': _cfg(input_size=(3, 384, 384)), + 'vit_base_patch16_384': _cfg( + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_384-83fb41ba.pth', + input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), + 'vit_base_patch32_384': _cfg( + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p32_384-830016f5.pth', + input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_large_patch16_224': _cfg(), - 'vit_large_patch16_384': _cfg(input_size=(3, 384, 384)), - 'vit_large_patch32_384': _cfg(input_size=(3, 384, 384)), + 'vit_large_patch16_384': _cfg( + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_384-b3be5167.pth', + input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), + 'vit_large_patch32_384': _cfg( + url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p32_384-9b920ba8.pth', + input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0), 'vit_huge_patch16_224': _cfg(), 'vit_huge_patch32_384': _cfg(input_size=(3, 384, 384)), # hybrid models @@ -77,38 +81,35 @@ class Mlp(nn.Module): self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) - self.dropout = nn.Dropout(drop) # seems more common to have Transformer MLP drouput here? + self.drop = nn.Dropout(drop) def forward(self, x): x = self.fc1(x) x = self.act(x) + x = self.drop(x) x = self.fc2(x) - x = self.dropout(x) + x = self.drop(x) return x class Attention(nn.Module): - def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0.): + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() - self.scale = 1. / dim ** 0.5 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.qkv = nn.Linear(dim, dim * 3, bias=False) + 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) - def forward(self, x, attn_mask=None): + def forward(self, x): B, N, C = x.shape - qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) - q, k, v = qkv[:, :, 0].transpose(1, 2), qkv[:, :, 1].transpose(1, 2), qkv[:, :, 2].transpose(1, 2) - - # TODO benchmark vs above - #qkv = qkv.reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) - #q, k, v = qkv + q, k, v = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) attn = (q @ k.transpose(-2, -1)) * self.scale - # FIXME support masking attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) @@ -120,52 +121,44 @@ class Attention(nn.Module): class Block(nn.Module): - def __init__(self, dim, num_heads, mlp_ratio=4., act_layer=nn.GELU, drop=0., drop_path=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): super().__init__() - self.norm1 = nn.LayerNorm(dim) - self.attn = Attention(dim, num_heads=num_heads, attn_drop=drop, proj_drop=drop) + 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) # 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 = nn.LayerNorm(dim) + 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, attn_mask=None): - x = x + self.drop_path(self.attn(self.norm1(x), attn_mask=attn_mask)) + 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 - Unfold image into fixed size patches, flatten into seq, project to embedding dim. """ - def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, flatten_channels_last=False): + 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) - assert img_size[0] % patch_size[0] == 0, 'image height must be divisible by the patch height' - assert img_size[1] % patch_size[1] == 0, 'image width must be divisible by the patch width' num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) - patch_dim = in_chans * patch_size[0] * patch_size[1] self.img_size = img_size self.patch_size = patch_size - self.flatten_channels_last = flatten_channels_last self.num_patches = num_patches - self.proj = nn.Linear(patch_dim, embed_dim) + 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 - Ph, Pw = self.patch_size + # 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]})." - if self.flatten_channels_last: - # flatten patches with channels last like the paper (likely using TF) - x = x.unfold(2, Ph, Ph).unfold(3, Pw, Pw).permute(0, 2, 3, 4, 5, 1).reshape(B, -1, Ph * Pw * C) - else: - x = x.permute(0, 2, 3, 1).unfold(1, Ph, Ph).unfold(2, Pw, Pw).reshape(B, -1, C * Ph * Pw) - x = self.proj(x) + x = self.proj(x).flatten(2).transpose(1, 2) return x @@ -208,37 +201,37 @@ class VisionTransformer(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., mlp_head=False, drop_rate=0., drop_path_rate=0., - flatten_channels_last=False, hybrid_backbone=None): + 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): super().__init__() 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, - flatten_channels_last=flatten_channels_last) + img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) num_patches = self.patch_embed.num_patches - self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) + 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 self.blocks = nn.ModuleList([ - Block(dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, drop=drop_rate, drop_path=dpr[i]) + 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) for i in range(depth)]) + self.norm = norm_layer(embed_dim) - self.norm = nn.LayerNorm(embed_dim) - if mlp_head: - # paper diagram suggests 'MLP head', but results in 4M extra parameters vs paper - self.head = Mlp(embed_dim, int(embed_dim * mlp_ratio), num_classes) - else: - # with a single Linear layer as head, the param count within rounding of paper - self.head = nn.Linear(embed_dim, num_classes) + # NOTE as per official impl, we could have a pre-logits representation dense layer + tanh here + #self.repr = nn.Linear(embed_dim, representation_size) + #self.repr_act = nn.Tanh() - # FIXME not quite sure what the proper weight init is supposed to be, - # normal / trunc normal w/ std == .02 similar to other Bert like transformers - trunc_normal_(self.pos_embed, std=.02) # embeddings same as weights? + # Classifier head + self.head = nn.Linear(embed_dim, num_classes) + + trunc_normal_(self.pos_embed, std=.02) trunc_normal_(self.cls_token, std=.02) self.apply(self._init_weights) @@ -255,55 +248,80 @@ class VisionTransformer(nn.Module): def no_weight_decay(self): return {'pos_embed', 'cls_token'} - def forward(self, x, attn_mask=None): + def forward(self, x): B = x.shape[0] x = self.patch_embed(x) cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks x = torch.cat((cls_tokens, x), dim=1) - x += self.pos_embed + x = x + self.pos_embed + x = self.pos_drop(x) for blk in self.blocks: - x = blk(x, attn_mask=attn_mask) + x = blk(x) - x = self.norm(x[:, 0]) - x = self.head(x) + x = self.norm(x) + x = self.head(x[:, 0]) return x +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 + + @register_model def vit_small_patch16_224(pretrained=False, **kwargs): + if pretrained: + # NOTE my scale was wrong for original weights, leaving this here until I have better ones for this model + kwargs.setdefault('qk_scale', 768 ** -0.5) model = VisionTransformer(patch_size=16, embed_dim=768, depth=8, num_heads=8, mlp_ratio=3., **kwargs) model.default_cfg = default_cfgs['vit_small_patch16_224'] if pretrained: load_pretrained( - model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3), filter_fn=_conv_filter) return model @register_model def vit_base_patch16_224(pretrained=False, **kwargs): + if pretrained: + # NOTE my scale was wrong for original weights, leaving this here until I have better ones for this model + kwargs.setdefault('qk_scale', 768 ** -0.5) model = VisionTransformer(patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, **kwargs) model.default_cfg = default_cfgs['vit_base_patch16_224'] if pretrained: load_pretrained( - model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3), filter_fn=_conv_filter) return model @register_model def vit_base_patch16_384(pretrained=False, **kwargs): model = VisionTransformer( - img_size=384, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, **kwargs) + img_size=384, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) model.default_cfg = default_cfgs['vit_base_patch16_384'] + if pretrained: + load_pretrained( + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) return model @register_model def vit_base_patch32_384(pretrained=False, **kwargs): model = VisionTransformer( - img_size=384, patch_size=32, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, **kwargs) + img_size=384, patch_size=32, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) model.default_cfg = default_cfgs['vit_base_patch32_384'] + if pretrained: + load_pretrained( + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) return model @@ -317,16 +335,24 @@ def vit_large_patch16_224(pretrained=False, **kwargs): @register_model def vit_large_patch16_384(pretrained=False, **kwargs): model = VisionTransformer( - img_size=384, patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, **kwargs) + img_size=384, patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) model.default_cfg = default_cfgs['vit_large_patch16_384'] + if pretrained: + load_pretrained( + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) return model @register_model def vit_large_patch32_384(pretrained=False, **kwargs): model = VisionTransformer( - img_size=384, patch_size=32, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, **kwargs) + img_size=384, patch_size=32, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs) model.default_cfg = default_cfgs['vit_large_patch32_384'] + if pretrained: + load_pretrained( + model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3)) return model @@ -383,5 +409,3 @@ def vit_base_resnet50d_224(pretrained=False, **kwargs): img_size=224, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, hybrid_backbone=backbone, **kwargs) model.default_cfg = default_cfgs['vit_base_resnet50d_224'] return model - -