pytorch-image-models/timm/models/vision_transformer.py

""" 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

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.

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...

Acknowledgments:
* 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
* Bert reference code checks against Huggingface Transformers and Tensorflow Bert

Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
import torch.nn as nn

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from .helpers import load_pretrained
from .layers import DropPath, to_2tuple, trunc_normal_
from .resnet import resnet26d, resnet50d
from .registry import register_model


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': 1.0, 'interpolation': 'bicubic',
        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
        'first_conv': '', 'classifier': 'head',
        **kwargs
    }


default_cfgs = {
    # patch models
    'vit_small_patch16_224': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth',
    ),
    'vit_base_patch16_224': _cfg(),
    'vit_base_patch16_384': _cfg(input_size=(3, 384, 384)),
    'vit_base_patch32_384': _cfg(input_size=(3, 384, 384)),
    '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_huge_patch16_224': _cfg(),
    'vit_huge_patch32_384': _cfg(input_size=(3, 384, 384)),
    # hybrid models
    'vit_small_resnet26d_224': _cfg(),
    'vit_small_resnet50d_s3_224': _cfg(),
    'vit_base_resnet26d_224': _cfg(),
    'vit_base_resnet50d_224': _cfg(),
}


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.dropout = nn.Dropout(drop)  # seems more common to have Transformer MLP drouput here?

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.fc2(x)
        x = self.dropout(x)
        return x


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.scale = 1. / dim ** 0.5
        self.num_heads = num_heads

        self.qkv = nn.Linear(dim, dim * 3, bias=False)
        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):
        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

        attn = (q @ k.transpose(-2, -1)) * self.scale
        # FIXME support masking
        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., act_layer=nn.GELU, drop=0., drop_path=0.):
        super().__init__()
        self.norm1 = nn.LayerNorm(dim)
        self.attn = Attention(dim, num_heads=num_heads, attn_drop=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)
        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))
        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):
        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)

    def forward(self, x):
        B, C, H, W = x.shape
        Ph, Pw = self.patch_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]})."
        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)
        return x


class HybridEmbed(nn.Module):
    """ CNN Feature Map Embedding
    Extract feature map from CNN, flatten, project to embedding dim.
    """
    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():
                # FIXME this is hacky, but most reliable way of determining the exact dim of the output feature
                # map for all networks, the feature metadata has reliable channel and stride info, but using
                # stride to calc feature dim requires info about padding of each stage that isn't captured.
                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)

    def forward(self, x):
        x = self.backbone(x)[-1]
        x = x.flatten(2).transpose(1, 2)
        x = self.proj(x)
        return x


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):
        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)
        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))

        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])
            for i in range(depth)])

        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)

        # 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?
        trunc_normal_(self.cls_token, 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)

    @property
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def forward(self, x, attn_mask=None):
        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

        for blk in self.blocks:
            x = blk(x, attn_mask=attn_mask)

        x = self.norm(x[:, 0])
        x = self.head(x)
        return x


@register_model
def vit_small_patch16_224(pretrained=False, **kwargs):
    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))
    return model


@register_model
def vit_base_patch16_224(pretrained=False, **kwargs):
    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']
    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)
    model.default_cfg = default_cfgs['vit_base_patch16_384']
    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)
    model.default_cfg = default_cfgs['vit_base_patch32_384']
    return model


@register_model
def vit_large_patch16_224(pretrained=False, **kwargs):
    model = VisionTransformer(patch_size=16, embed_dim=1024, depth=24, num_heads=16, mlp_ratio=4, **kwargs)
    model.default_cfg = default_cfgs['vit_large_patch16_224']
    return model


@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)
    model.default_cfg = default_cfgs['vit_large_patch16_384']
    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)
    model.default_cfg = default_cfgs['vit_large_patch32_384']
    return model


@register_model
def vit_huge_patch16_224(pretrained=False, **kwargs):
    model = VisionTransformer(patch_size=16, embed_dim=1280, depth=32, num_heads=16, mlp_ratio=4, **kwargs)
    model.default_cfg = default_cfgs['vit_huge_patch16_224']
    return model


@register_model
def vit_huge_patch32_384(pretrained=False, **kwargs):
    model = VisionTransformer(
        img_size=384, patch_size=32, embed_dim=1280, depth=32, num_heads=16, mlp_ratio=4, **kwargs)
    model.default_cfg = default_cfgs['vit_huge_patch32_384']
    return model


@register_model
def vit_small_resnet26d_224(pretrained=False, **kwargs):
    pretrained_backbone = kwargs.get('pretrained_backbone', True)  # default to True for now, for testing
    backbone = resnet26d(pretrained=pretrained_backbone, features_only=True, out_indices=[4])
    model = VisionTransformer(
        img_size=224, embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, hybrid_backbone=backbone, **kwargs)
    model.default_cfg = default_cfgs['vit_small_resnet26d_224']
    return model


@register_model
def vit_small_resnet50d_s3_224(pretrained=False, **kwargs):
    pretrained_backbone = kwargs.get('pretrained_backbone', True)  # default to True for now, for testing
    backbone = resnet50d(pretrained=pretrained_backbone, features_only=True, out_indices=[3])
    model = VisionTransformer(
        img_size=224, embed_dim=768, depth=8, num_heads=8, mlp_ratio=3, hybrid_backbone=backbone, **kwargs)
    model.default_cfg = default_cfgs['vit_small_resnet50d_s3_224']
    return model


@register_model
def vit_base_resnet26d_224(pretrained=False, **kwargs):
    pretrained_backbone = kwargs.get('pretrained_backbone', True)  # default to True for now, for testing
    backbone = resnet26d(pretrained=pretrained_backbone, features_only=True, out_indices=[4])
    model = VisionTransformer(
        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_resnet26d_224']
    return model


@register_model
def vit_base_resnet50d_224(pretrained=False, **kwargs):
    pretrained_backbone = kwargs.get('pretrained_backbone', True)  # default to True for now, for testing
    backbone = resnet50d(pretrained=pretrained_backbone, features_only=True, out_indices=[4])
    model = VisionTransformer(
        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