vision_transformer_diverse.py

""" Vision Transformer (ViT) in PyTorch
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 official jax code is released and available at https://github.com/google-research/vision_transformer
DeiT model defs and weights from https://github.com/facebookresearch/deit,
paper `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877
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
* Bert reference code checks against Huggingface Transformers and Tensorflow Bert
Hacked together by / Copyright 2020 Ross Wightman
"""
'''
gumbel softmax on the top layer
'''



import math
import logging
from functools import partial
from collections import OrderedDict
from copy import deepcopy

import torch
import numpy as np 
import torch.nn as nn
import torch.nn.functional as F

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.models.layers import DropPath, to_2tuple, trunc_normal_


_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
    }


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)

    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 Attention(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)

    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]   # make torchscript happy (cannot use tensor as tuple)

        attn_origin = (q @ k.transpose(-2, -1)) * self.scale
        attn_origin = torch.softmax(attn_origin, dim=-1) #(B, Heads, N, N)
        attn = self.attn_drop(attn_origin)

        x = attn @ v
        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x, attn_origin


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):
        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)
        # 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)

    # 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

    def forward(self, x):

        shortcut = x 
        x, affinity_attn = self.attn(self.norm1(x))
        x = shortcut + self.drop_path(x)

        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x, affinity_attn


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)
        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)

    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 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
    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`
        - https://arxiv.org/abs/2010.11929
    Includes distillation token & head support for `DeiT: Data-efficient Image Transformers`
        - https://arxiv.org/abs/2012.12877
    """

    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, qk_scale=None, representation_size=None, distilled=False,
                drop_rate=0., attn_drop_rate=0., drop_path_rate=0., embed_layer=PatchEmbed, norm_layer=None,
                act_layer=None, weight_init=''):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_chans (int): number of input channels
            num_classes (int): number of classes for classification head
            embed_dim (int): embedding dimension
            depth (int): depth of transformer
            num_heads (int): number of attention heads
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            qkv_bias (bool): enable bias for qkv if True
            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
            representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
            distilled (bool): model includes a distillation token and head as in DeiT models
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            embed_layer (nn.Module): patch embedding layer
            norm_layer: (nn.Module): normalization layer
            weight_init: (str): weight init scheme
        """
        super().__init__()
        
        if distilled:
            assert False 
    
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
        self.num_tokens = 2 if distilled else 1
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        act_layer = act_layer or nn.GELU

        self.patch_embed = embed_layer(
            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.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
        self.pos_drop = nn.Dropout(p=drop_rate)
        self.features = []
        self.token_grad = None

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.blocks = nn.Sequential(*[
            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, act_layer=act_layer)
            for i in range(depth)])
        self.norm = norm_layer(embed_dim)

        # Representation layer
        if representation_size and not distilled:
            self.num_features = representation_size
            self.pre_logits = nn.Sequential(OrderedDict([
                ('fc', nn.Linear(embed_dim, representation_size)),
                ('act', nn.Tanh())
            ]))
        else:
            self.pre_logits = nn.Identity()

        # Classifier head(s)
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
        self.head_dist = None
        if distilled:
            self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()

        # Weight init
        assert weight_init in ('jax', 'jax_nlhb', 'nlhb', '')
        head_bias = -math.log(self.num_classes) if 'nlhb' in weight_init else 0.
        trunc_normal_(self.pos_embed, std=.02)
        if self.dist_token is not None:
            trunc_normal_(self.dist_token, std=.02)
        if weight_init.startswith('jax'):
            # leave cls token as zeros to match jax impl
            for n, m in self.named_modules():
                _init_vit_weights(m, n, head_bias=head_bias, jax_impl=True)
        else:
            trunc_normal_(self.cls_token, std=.02)
            self.apply(_init_vit_weights)

    def _init_weights(self, m):
        # this fn left here for compat with downstream users
        _init_vit_weights(m)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token', 'dist_token'}

    def get_classifier(self):
        if self.dist_token is None:
            return self.head
        else:
            return self.head, self.head_dist

    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()
        if self.num_tokens == 2:
            self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()

    def forward(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
        x = torch.cat((cls_tokens, x), dim=1)
        x1 = self.pos_drop(x + self.pos_embed)


        xtemp = x1 
        for ii, blk in enumerate(self.blocks):
            xtemp, affinity_attn = blk(xtemp)
            if ii == 0:
                attn1 = affinity_attn

        xl = xtemp 
        xo = self.norm(xl)
        xo = self.head(xo)

        if self.training:
            return xo, x1, xl, attn1
        else:
            return xo[:,0]


def _init_vit_weights(m, n: str = '', head_bias: float = 0., jax_impl: bool = False):
    """ ViT weight initialization
    * When called without n, head_bias, jax_impl args it will behave exactly the same
    as my original init for compatibility with prev hparam / downstream use cases (ie DeiT).
    * When called w/ valid n (module name) and jax_impl=True, will (hopefully) match JAX impl
    """
    if isinstance(m, nn.Linear):
        if n.startswith('head'):
            nn.init.zeros_(m.weight)
            nn.init.constant_(m.bias, head_bias)
        elif n.startswith('pre_logits'):
            lecun_normal_(m.weight)
            nn.init.zeros_(m.bias)
        else:
            if jax_impl:
                nn.init.xavier_uniform_(m.weight)
                if m.bias is not None:
                    if 'mlp' in n:
                        nn.init.normal_(m.bias, std=1e-6)
                    else:
                        nn.init.zeros_(m.bias)
            else:
                trunc_normal_(m.weight, std=.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
    elif jax_impl and isinstance(m, nn.Conv2d):
        # NOTE conv was left to pytorch default in my original init
        lecun_normal_(m.weight)
        if m.bias is not None:
            nn.init.zeros_(m.bias)
    elif isinstance(m, nn.LayerNorm):
        nn.init.zeros_(m.bias)
        nn.init.ones_(m.weight)


def resize_pos_embed(posemb, posemb_new, num_tokens=1):
    # 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)
    ntok_new = posemb_new.shape[1]
    if num_tokens:
        posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]
        ntok_new -= num_tokens
    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)
    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 = torch.cat([posemb_tok, posemb_grid], dim=1)
    return posemb


def checkpoint_filter_fn(state_dict, model):
    """ convert patch embedding weight from manual patchify + linear proj to conv"""
    out_dict = {}
    if 'model' in state_dict:
        # For deit models
        state_dict = state_dict['model']
    for k, v in state_dict.items():
        if 'patch_embed.proj.weight' in k and len(v.shape) < 4:
            # For old models that I trained prior to conv based patchification
            O, I, H, W = model.patch_embed.proj.weight.shape
            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))
        out_dict[k] = v
    return out_dict


def _create_vision_transformer(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)
    repr_size = kwargs.pop('representation_size', None)
    if repr_size is not None and num_classes != default_num_classes:
        # Remove representation layer if fine-tuning. This may not always be the desired action,
        # but I feel better than doing nothing by default for fine-tuning. Perhaps a better interface?
        _logger.warning("Removing representation layer for fine-tuning.")
        repr_size = None

    if kwargs.get('features_only', None):
        raise RuntimeError('features_only not implemented for Vision Transformer models.')

    model = build_model_with_cfg(
        VisionTransformer, variant, pretrained,
        default_cfg=default_cfg,
        img_size=img_size,
        num_classes=num_classes,
        representation_size=repr_size,
        pretrained_filter_fn=checkpoint_filter_fn,
        **kwargs)

    return model