networks/timm_deit.py

# Copyright (c) 2015-present, Facebook, Inc.
# All rights reserved.
import math
import torch
import torch.nn as nn
from functools import partial

from networks.timm_vit import VisionTransformer, _cfg
from timm.models.registry import register_model
from timm.models.layers import trunc_normal_


__all__ = [
    'deit_tiny_patch16_224', 'deit_small_patch16_224', 'deit_base_patch16_224',
    'deit_tiny_distilled_patch16_224', 'deit_small_distilled_patch16_224',
    'deit_base_distilled_patch16_224', 'deit_base_patch16_384',
    'deit_base_distilled_patch16_384',
]


class DistilledVisionTransformer(VisionTransformer):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
        num_patches = self.patch_embed.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 2, self.embed_dim))
        self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity()

        trunc_normal_(self.dist_token, std=.02)
        trunc_normal_(self.pos_embed, std=.02)
        self.head_dist.apply(self._init_weights)

    def forward_features(self, x):
        # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
        # with slight modifications to add the dist_token
        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
        dist_token = self.dist_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, dist_token, x), dim=1)

        x = x + self.pos_embed
        x = self.pos_drop(x)

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

        x = self.norm(x)
        return x[:, 0], x[:, 1]

    def forward(self, x):
        x, x_dist = self.forward_features(x)
        x = self.head(x)
        x_dist = self.head_dist(x_dist)
        if self.training:
            return x, x_dist
        else:
            # during inference, return the average of both classifier predictions
            return (x + x_dist) / 2

    def interpolate_pos_encoding(self, x, pos_embed):
        """Interpolate the learnable positional encoding to match the number of patches.

        x: B x (1 + 1 + N patches) x dim_embedding
        pos_embed: B x (1 + 1 + N patches) x dim_embedding

        return interpolated positional embedding
        """

        npatch = x.shape[1] - 2  # (H // patch_size * W // patch_size)
        N = pos_embed.shape[1] - 2  # 784 (= 28 x 28)

        if npatch == N:
            return pos_embed

        class_emb, distil_token, pos_embed = pos_embed[:, 0], pos_embed[:, 1], pos_embed[:, 2:]  # a learnable CLS token, learnable position embeddings

        dim = x.shape[-1]  # dimension of embeddings
        pos_embed = nn.functional.interpolate(
            pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),  # B x dim x 28 x 28
            scale_factor=math.sqrt(npatch / N) + 1e-5,  # noel: this can be a float, but the output shape will be integer.
            recompute_scale_factor=True,
            mode='bicubic'
        )
        # print("pos_embed", pos_embed.shape, npatch, N, math.sqrt(npatch/N), math.sqrt(npatch/N) * int(math.sqrt(N)))
        # exit(12)
        pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        pos_embed = torch.cat((class_emb.unsqueeze(0), distil_token.unsqueeze(0), pos_embed), dim=1)
        return pos_embed

    def get_tokens(
            self,
            x,
            layers: list,
            patch_tokens: bool = False,
            norm: bool = True,
            input_tokens: bool = False,
            post_pe: bool = False
    ):
        """Return intermediate tokens."""
        list_tokens: list = []

        B = x.shape[0]
        x = self.patch_embed(x)

        cls_tokens = self.cls_token.expand(B, -1, -1)
        dist_token = self.dist_token.expand(B, -1, -1)

        x = torch.cat((cls_tokens, dist_token, x), dim=1)

        if input_tokens:
            list_tokens.append(x)

        pos_embed = self.interpolate_pos_encoding(x, self.pos_embed)
        x = x + pos_embed

        if post_pe:
            list_tokens.append(x)

        x = self.pos_drop(x)

        for i, blk in enumerate(self.blocks):
            x = blk(x)  # B x # patches x dim
            if layers is None or i in layers:
                list_tokens.append(self.norm(x) if norm else x)

        tokens = torch.stack(list_tokens, dim=1)  # B x n_layers x (1 + # patches) x dim

        if not patch_tokens:
            return tokens[:, :, 0, :]  # index [CLS] tokens only, B x n_layers x dim

        else:
            return torch.cat((tokens[:, :, 0, :].unsqueeze(dim=2), tokens[:, :, 2:, :]), dim=2)  # exclude distil token.


@register_model
def deit_tiny_patch16_224(pretrained=False, **kwargs):
    model = VisionTransformer(
        patch_size=16, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_small_patch16_224(pretrained=False, **kwargs):
    model = VisionTransformer(
        patch_size=16, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_small_patch16_224-cd65a155.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_base_patch16_224(pretrained=False, **kwargs):
    model = VisionTransformer(
        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 = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_tiny_distilled_patch16_224(pretrained=False, **kwargs):
    model = DistilledVisionTransformer(
        patch_size=16, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_small_distilled_patch16_224(pretrained=False, **kwargs):
    model = DistilledVisionTransformer(
        patch_size=16, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_base_distilled_patch16_224(pretrained=False, **kwargs):
    model = DistilledVisionTransformer(
        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 = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_224-df68dfff.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_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, qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
    model.default_cfg = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_base_patch16_384-8de9b5d1.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def deit_base_distilled_patch16_384(pretrained=False, **kwargs):
    model = DistilledVisionTransformer(
        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 = _cfg()
    if pretrained:
        checkpoint = torch.hub.load_state_dict_from_url(
            url="https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_384-d0272ac0.pth",
            map_location="cpu", check_hash=True
        )
        model.load_state_dict(checkpoint["model"])
    return model