vits.py

# =============================================================================================
# Based on timm/models/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/blob/main/timm/models/vision_transformer.py
# =============================================================================================

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

import torch
import torch.nn as nn
import torch.nn.functional as F

from torch.utils.checkpoint import checkpoint
from timm.data import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
from timm.models.helpers import build_model_with_cfg, named_apply, adapt_input_conv
from timm.models.layers import Mlp, DropPath, trunc_normal_, lecun_normal_, to_2tuple
from timm.models.registry import register_model
_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_INCEPTION_MEAN, 'std': IMAGENET_INCEPTION_STD,
        'first_conv': 'patch_embed.proj', 'classifier': 'head',
        **kwargs
    }


default_cfgs = {
    'vit_tiny_patch16_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_tiny_patch32_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_small_patch32_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_small_patch16_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_base_patch32_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'B_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_base_patch16_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz',
        input_size=(3, 384, 384), 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), crop_pct=1.0),
    'vit_large_patch16_384': _cfg(
        url='https://storage.googleapis.com/vit_models/augreg/'
            'L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_384.npz',
        input_size=(3, 384, 384), crop_pct=1.0),
}


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = 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.unbind(0)   # make torchscript happy (cannot use tensor as tuple)

        attn = (q @ k.transpose(-2, -1)) * self.scale
        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, attn


class LayerScale(nn.Module):
    def __init__(self, dim, init_values=1e-5, inplace=False):
        super().__init__()
        self.inplace = inplace
        self.gamma = nn.Parameter(init_values * torch.ones(dim))

    def forward(self, x):
        return x.mul_(self.gamma) if self.inplace else x * self.gamma


class Block(nn.Module):

    def __init__(
            self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0., init_values=None,
            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, attn_drop=attn_drop, proj_drop=drop)
        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path1 = 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)
        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x, return_attention=False):
        y, attn = self.attn(self.norm1(x))
        if return_attention:
            return attn
        x = x + self.drop_path1(self.ls1(y))
        x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
        return x


class ParallelBlock(nn.Module):
    def __init__(
            self, dim, num_heads, num_parallel=2, mlp_ratio=4., qkv_bias=False, init_values=None,
            drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.num_parallel = num_parallel
        self.attns = nn.ModuleList()
        self.ffns = nn.ModuleList()
        for _ in range(num_parallel):
            self.attns.append(nn.Sequential(OrderedDict([
                ('norm', norm_layer(dim)),
                ('attn', Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)),
                ('ls', LayerScale(dim, init_values=init_values) if init_values else nn.Identity()),
                ('drop_path', DropPath(drop_path) if drop_path > 0. else nn.Identity())
            ])))
            self.ffns.append(nn.Sequential(OrderedDict([
                ('norm', norm_layer(dim)),
                ('mlp', Mlp(dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)),
                ('ls', LayerScale(dim, init_values=init_values) if init_values else nn.Identity()),
                ('drop_path', DropPath(drop_path) if drop_path > 0. else nn.Identity())
            ])))

    def _forward_jit(self, x):
        x = x + torch.stack([attn(x) for attn in self.attns]).sum(dim=0)
        x = x + torch.stack([ffn(x) for ffn in self.ffns]).sum(dim=0)
        return x

    @torch.jit.ignore
    def _forward(self, x):
        x = x + sum(attn(x) for attn in self.attns)
        x = x + sum(ffn(x) for ffn in self.ffns)
        return x

    def forward(self, x):
        if torch.jit.is_scripting() or torch.jit.is_tracing():
            return self._forward_jit(x)
        else:
            return self._forward(x)


class PatchEmbed(nn.Module):
    """ 2D Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
        self.num_patches = self.grid_size[0] * self.grid_size[1]
        self.flatten = flatten

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        B, C, H, W = x.shape
        x = self.proj(x)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)
        return x


class PatchSampler(object):
    def __init__(self, mask_size, patch_size, mask_ratio=0.25):
        self.mask_size = mask_size
        self.patch_size = patch_size
        self.mask_ratio = mask_ratio

    def __call__(self, pmap):
        B, C, H, W = pmap.shape
        num_sample = int((1 - self.mask_ratio) * H * W)

        feat_idx = pmap.flatten(1).argsort(descending=True)[:,:num_sample]
        feat_idx += 1 # class embedding concat before the image embedding
        cls_idx = torch.zeros((B, 1), dtype=torch.int64, device=pmap.device)
        active_idx = torch.cat([cls_idx, feat_idx], dim=1)
        return active_idx


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
    """

    def __init__(
            self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, global_pool='token',
            embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, representation_size=None,
            drop_rate=0., attn_drop_rate=0., drop_path_rate=0., weight_init='', init_values=None,
            embed_layer=PatchEmbed, norm_layer=None, act_layer=None, block_fn=Block, mask_ratio=0.25, feat_concat=False):
        """
        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
            global_pool (str): type of global pooling for final sequence (default: 'token')
            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
            init_values: (float): layer-scale init values
            class_token (bool): use class token
            fc_norm (Optional[bool]): pre-fc norm after pool, set if global_pool == 'avg' if None (default: None)
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            weight_init (str): weight init scheme
            embed_layer (nn.Module): patch embedding layer
            norm_layer: (nn.Module): normalization layer
            act_layer: (nn.Module): MLP activation layer
        """
        super().__init__()
        assert global_pool in ('', 'avg', 'token')
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        act_layer = act_layer or nn.GELU

        # store grid_size
        self.grid_sizes = {}
        self.patch_size = patch_size

        self.num_classes = num_classes
        self.global_pool = global_pool
        self.feat_concat = feat_concat
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
        self.num_tokens = 1
        self.grad_checkpointing = False

        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.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, 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.Sequential(*[
            block_fn(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, init_values=init_values,
                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)])
        use_fc_norm = self.global_pool == 'avg'
        self.norm = norm_layer(embed_dim) if not use_fc_norm else nn.Identity()

        # Representation layer. Used for original ViT models w/ in21k pretraining.
        self.representation_size = representation_size
        self.pre_logits = nn.Identity()
        if representation_size:
            self._reset_representation(representation_size)

        # Classifier Head
        final_chs = self.representation_size if self.representation_size else self.embed_dim
        final_chs = final_chs * 2 if self.feat_concat else final_chs
        self.fc_norm = norm_layer(final_chs) if use_fc_norm else nn.Identity()
        self.head = nn.Linear(final_chs, num_classes) if num_classes > 0 else nn.Identity()

        self.mask_ratio = mask_ratio
        self.patch_sampler = PatchSampler(mask_size=img_size, patch_size=patch_size, mask_ratio=mask_ratio)

        if weight_init != 'skip':
            self.init_weights(weight_init)

    def init_weights(self, mode=''):
        assert mode in ('jax', 'jax_nlhb', 'moco', '')
        head_bias = -math.log(self.num_classes) if 'nlhb' in mode else 0.
        trunc_normal_(self.pos_embed, std=.02)
        nn.init.normal_(self.cls_token, std=1e-6)
        named_apply(get_init_weights_vit(mode, head_bias), self)

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

    @torch.jit.ignore()
    def load_pretrained(self, checkpoint_path, prefix=''):
        _load_weights(self, checkpoint_path, prefix)

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

    @torch.jit.ignore
    def group_matcher(self, coarse=False):
        return dict(
            stem=r'^cls_token|pos_embed|patch_embed',  # stem and embed
            blocks=[(r'^blocks\.(\d+)', None), (r'^norm', (99999,))]
        )

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.grad_checkpointing = enable

    @torch.jit.ignore
    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes: int, global_pool=None, representation_size=None):
        self.num_classes = num_classes
        if global_pool is not None:
            assert global_pool in ('', 'avg', 'token')
            self.global_pool = global_pool
        if representation_size is not None:
            self._reset_representation(representation_size)
        final_chs = self.representation_size if self.representation_size else self.embed_dim
        self.head = nn.Linear(final_chs, num_classes) if num_classes > 0 else nn.Identity()

    def forward_features(self, x, pmap=None):
        x = self.patch_embed(x)
        if self.cls_token is not None:
            x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
        x = x + self.pos_embed

        if pmap is not None and self.mask_ratio < 1:
            active_idx = self.patch_sampler(pmap)
            active_idx = active_idx.unsqueeze(-1).repeat(1, 1, self.embed_dim)
            x = torch.gather(x, dim=1, index=active_idx)

        x = self.pos_drop(x)
        if self.grad_checkpointing and not torch.jit.is_scripting():
            x = checkpoint_seq(self.blocks, x)
        else:
            x = self.blocks(x)
        x = self.norm(x)
        return x

    def forward_head(self, x, pre_logits: bool = False):
        if self.feat_concat:
            feats = x[:, 1:].mean(dim=1)
            x = torch.cat((x[:, 0], feats), dim=1)
        elif self.global_pool:
            x = x[:, 1:].mean(dim=1) if self.global_pool == 'avg' else x[:, 0]
        x = self.fc_norm(x)
        x = self.pre_logits(x)
        return x if pre_logits else self.head(x)

    def forward(self, x, pmap=None):
        f = self.forward_features(x, pmap)
        x = self.forward_head(f)
        return x, f

    def get_last_selfattention(self, x):
        x = self.patch_embed(x)
        if self.cls_token is not None:
            x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
        x = x + self.pos_embed
        x = self.pos_drop(x)

        for i, blk in enumerate(self.blocks):
            if i < len(self.blocks) - 1:
                x = blk(x)
            else:
                # return attention of the last block
                return blk(x, return_attention=True)


def init_weights_vit_timm(module: nn.Module, name: str = ''):
    """ ViT weight initialization, original timm impl (for reproducibility) """
    if isinstance(module, nn.Linear):
        trunc_normal_(module.weight, std=.02)
        if module.bias is not None:
            nn.init.zeros_(module.bias)


def init_weights_vit_jax(module: nn.Module, name: str = '', head_bias: float = 0.):
    """ ViT weight initialization, matching JAX (Flax) impl """
    if isinstance(module, nn.Linear):
        if name.startswith('head'):
            nn.init.zeros_(module.weight)
            nn.init.constant_(module.bias, head_bias)
        elif name.startswith('pre_logits'):
            lecun_normal_(module.weight)
            nn.init.zeros_(module.bias)
        else:
            nn.init.xavier_uniform_(module.weight)
            if module.bias is not None:
                nn.init.normal_(module.bias, std=1e-6) if 'mlp' in name else nn.init.zeros_(module.bias)
    elif isinstance(module, nn.Conv2d):
        lecun_normal_(module.weight)
        if module.bias is not None:
            nn.init.zeros_(module.bias)


def init_weights_vit_moco(module: nn.Module, name: str = ''):
    """ ViT weight initialization, matching moco-v3 impl minus fixed PatchEmbed """
    if isinstance(module, nn.Linear):
        if 'qkv' in name:
            # treat the weights of Q, K, V separately
            val = math.sqrt(6. / float(module.weight.shape[0] // 3 + module.weight.shape[1]))
            nn.init.uniform_(module.weight, -val, val)
        else:
            nn.init.xavier_uniform_(module.weight)
        if module.bias is not None:
            nn.init.zeros_(module.bias)


def get_init_weights_vit(mode='jax', head_bias: float = 0.):
    if 'jax' in mode:
        return partial(init_weights_vit_jax, head_bias=head_bias)
    elif 'moco' in mode:
        return init_weights_vit_moco
    else:
        return init_weights_vit_timm


@torch.no_grad()
def _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = ''):
    """ Load weights from .npz checkpoints for official Google Brain Flax implementation
    """
    import numpy as np

    def _n2p(w, t=True):
        if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:
            w = w.flatten()
        if t:
            if w.ndim == 4:
                w = w.transpose([3, 2, 0, 1])
            elif w.ndim == 3:
                w = w.transpose([2, 0, 1])
            elif w.ndim == 2:
                w = w.transpose([1, 0])
        return torch.from_numpy(w)

    w = np.load(checkpoint_path)
    if not prefix and 'opt/target/embedding/kernel' in w:
        prefix = 'opt/target/'

    if hasattr(model.patch_embed, 'backbone'):
        # hybrid
        backbone = model.patch_embed.backbone
        stem_only = not hasattr(backbone, 'stem')
        stem = backbone if stem_only else backbone.stem
        stem.conv.weight.copy_(adapt_input_conv(stem.conv.weight.shape[1], _n2p(w[f'{prefix}conv_root/kernel'])))
        stem.norm.weight.copy_(_n2p(w[f'{prefix}gn_root/scale']))
        stem.norm.bias.copy_(_n2p(w[f'{prefix}gn_root/bias']))
        if not stem_only:
            for i, stage in enumerate(backbone.stages):
                for j, block in enumerate(stage.blocks):
                    bp = f'{prefix}block{i + 1}/unit{j + 1}/'
                    for r in range(3):
                        getattr(block, f'conv{r + 1}').weight.copy_(_n2p(w[f'{bp}conv{r + 1}/kernel']))
                        getattr(block, f'norm{r + 1}').weight.copy_(_n2p(w[f'{bp}gn{r + 1}/scale']))
                        getattr(block, f'norm{r + 1}').bias.copy_(_n2p(w[f'{bp}gn{r + 1}/bias']))
                    if block.downsample is not None:
                        block.downsample.conv.weight.copy_(_n2p(w[f'{bp}conv_proj/kernel']))
                        block.downsample.norm.weight.copy_(_n2p(w[f'{bp}gn_proj/scale']))
                        block.downsample.norm.bias.copy_(_n2p(w[f'{bp}gn_proj/bias']))
        embed_conv_w = _n2p(w[f'{prefix}embedding/kernel'])
    else:
        embed_conv_w = adapt_input_conv(
            model.patch_embed.proj.weight.shape[1], _n2p(w[f'{prefix}embedding/kernel']))
    model.patch_embed.proj.weight.copy_(embed_conv_w)
    model.patch_embed.proj.bias.copy_(_n2p(w[f'{prefix}embedding/bias']))
    model.cls_token.copy_(_n2p(w[f'{prefix}cls'], t=False))
    pos_embed_w = _n2p(w[f'{prefix}Transformer/posembed_input/pos_embedding'], t=False)
    if pos_embed_w.shape != model.pos_embed.shape:
        pos_embed_w = resize_pos_embed(  # resize pos embedding when different size from pretrained weights
            pos_embed_w, model.pos_embed, getattr(model, 'num_tokens', 1), model.patch_embed.grid_size)
    model.pos_embed.copy_(pos_embed_w)
    model.norm.weight.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/scale']))
    model.norm.bias.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/bias']))
    if isinstance(model.head, nn.Linear) and model.head.bias.shape[0] == w[f'{prefix}head/bias'].shape[-1]:
        model.head.weight.copy_(_n2p(w[f'{prefix}head/kernel']))
        model.head.bias.copy_(_n2p(w[f'{prefix}head/bias']))
    if isinstance(getattr(model.pre_logits, 'fc', None), nn.Linear) and f'{prefix}pre_logits/bias' in w:
        model.pre_logits.fc.weight.copy_(_n2p(w[f'{prefix}pre_logits/kernel']))
        model.pre_logits.fc.bias.copy_(_n2p(w[f'{prefix}pre_logits/bias']))
    for i, block in enumerate(model.blocks.children()):
        block_prefix = f'{prefix}Transformer/encoderblock_{i}/'
        mha_prefix = block_prefix + 'MultiHeadDotProductAttention_1/'
        block.norm1.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/scale']))
        block.norm1.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/bias']))
        block.attn.qkv.weight.copy_(torch.cat([
            _n2p(w[f'{mha_prefix}{n}/kernel'], t=False).flatten(1).T for n in ('query', 'key', 'value')]))
        block.attn.qkv.bias.copy_(torch.cat([
            _n2p(w[f'{mha_prefix}{n}/bias'], t=False).reshape(-1) for n in ('query', 'key', 'value')]))
        block.attn.proj.weight.copy_(_n2p(w[f'{mha_prefix}out/kernel']).flatten(1))
        block.attn.proj.bias.copy_(_n2p(w[f'{mha_prefix}out/bias']))
        for r in range(2):
            getattr(block.mlp, f'fc{r + 1}').weight.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/kernel']))
            getattr(block.mlp, f'fc{r + 1}').bias.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/bias']))
        block.norm2.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/scale']))
        block.norm2.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/bias']))


def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()):
    # 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)))
    if not len(gs_new):  # backwards compatibility
        gs_new = [int(math.sqrt(ntok_new))] * 2
    assert len(gs_new) >= 2
    _logger.info('Position embedding grid-size from %s to %s', [gs_old, 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, mode='bicubic', align_corners=False)
    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -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), model.patch_embed.grid_size)
        out_dict[k] = v
    return out_dict


def _create_vision_transformer(variant, pretrained=False, **kwargs):
    if kwargs.get('features_only', None):
        raise RuntimeError('features_only not implemented for Vision Transformer models.')

    # NOTE this extra code to support handling of repr size for in21k pretrained models
    pretrained_cfg = resolve_pretrained_cfg(variant, kwargs=kwargs)
    default_num_classes = pretrained_cfg['num_classes']
    num_classes = kwargs.get('num_classes', default_num_classes)
    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

    model = build_model_with_cfg(
        VisionTransformer, variant, pretrained,
        default_cfg=pretrained_cfg,
        representation_size=repr_size,
        pretrained_filter_fn=checkpoint_filter_fn,
        pretrained_custom_load='npz' in pretrained_cfg['url'],
        **kwargs)
    return model


@register_model
def vit_tiny_patch16_384(pretrained=False, **kwargs):
    """ ViT-Tiny (Vit-Ti/16) @ 384x384.
    """
    model_kwargs = dict(patch_size=16, embed_dim=192, depth=12, num_heads=3, **kwargs)
    model = _create_vision_transformer('vit_tiny_patch16_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_tiny_patch32_384(pretrained=False, **kwargs):
    """ ViT-Tiny (Vit-Ti/16) @ 384x384.
    """
    model_kwargs = dict(patch_size=32, embed_dim=192, depth=12, num_heads=3, **kwargs)
    model = _create_vision_transformer('vit_tiny_patch16_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_small_patch32_384(pretrained=False, **kwargs):
    """ ViT-Small (ViT-S/32) at 384x384.
    """
    model_kwargs = dict(patch_size=32, embed_dim=384, depth=12, num_heads=6, **kwargs)
    model = _create_vision_transformer('vit_small_patch32_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_small_patch16_384(pretrained=False, **kwargs):
    """ ViT-Small (ViT-S/16)
    NOTE I've replaced my previous 'small' model definition and weights with the small variant from the DeiT paper
    """
    model_kwargs = dict(patch_size=16, embed_dim=384, depth=12, num_heads=6, **kwargs)
    model = _create_vision_transformer('vit_small_patch16_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_base_patch32_384(pretrained=False, **kwargs):
    """ ViT-Base model (ViT-B/32) from original paper (https://arxiv.org/abs/2010.11929).
    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
    """
    model_kwargs = dict(patch_size=32, embed_dim=768, depth=12, num_heads=12, **kwargs)
    model = _create_vision_transformer('vit_base_patch32_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_base_patch16_384(pretrained=False, **kwargs):
    """ ViT-Base model (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929).
    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
    """
    model_kwargs = dict(patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs)
    model = _create_vision_transformer('vit_base_patch16_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_large_patch32_384(pretrained=False, **kwargs):
    """ ViT-Large model (ViT-L/32) from original paper (https://arxiv.org/abs/2010.11929).
    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
    """
    model_kwargs = dict(patch_size=32, embed_dim=1024, depth=24, num_heads=16, **kwargs)
    model = _create_vision_transformer('vit_large_patch32_384', pretrained=pretrained, **model_kwargs)
    return model


@register_model
def vit_large_patch16_384(pretrained=False, **kwargs):
    """ ViT-Large model (ViT-L/16) from original paper (https://arxiv.org/abs/2010.11929).
    ImageNet-1k weights fine-tuned from in21k @ 384x384, source https://github.com/google-research/vision_transformer.
    """
    model_kwargs = dict(patch_size=16, embed_dim=1024, depth=24, num_heads=16, **kwargs)
    model = _create_vision_transformer('vit_large_patch16_384', pretrained=pretrained, **model_kwargs)
    return model


def resolve_pretrained_cfg(variant: str, pretrained_cfg=None, kwargs=None):
    if pretrained_cfg and isinstance(pretrained_cfg, dict):
        # highest priority, pretrained_cfg available and passed explicitly
        return deepcopy(pretrained_cfg)
    if kwargs and 'pretrained_cfg' in kwargs:
        # next highest, pretrained_cfg in a kwargs dict, pop and return
        pretrained_cfg = kwargs.pop('pretrained_cfg', {})
        if pretrained_cfg:
            return deepcopy(pretrained_cfg)
    # lookup pretrained cfg in model registry by variant
    pretrained_cfg = get_pretrained_cfg(variant)
    assert pretrained_cfg
    return pretrained_cfg


def get_pretrained_cfg(model_name):
    if model_name in default_cfgs:
        return deepcopy(default_cfgs[model_name])
    return {}


def checkpoint_seq(
        functions,
        x,
        every=1,
        flatten=False,
        skip_last=False,
        preserve_rng_state=True
):
    r"""A helper function for checkpointing sequential models.
    Sequential models execute a list of modules/functions in order
    (sequentially). Therefore, we can divide such a sequence into segments
    and checkpoint each segment. All segments except run in :func:`torch.no_grad`
    manner, i.e., not storing the intermediate activations. The inputs of each
    checkpointed segment will be saved for re-running the segment in the backward pass.
    See :func:`~torch.utils.checkpoint.checkpoint` on how checkpointing works.
    .. warning::
        Checkpointing currently only supports :func:`torch.autograd.backward`
        and only if its `inputs` argument is not passed. :func:`torch.autograd.grad`
        is not supported.
    .. warning:
        At least one of the inputs needs to have :code:`requires_grad=True` if
        grads are needed for model inputs, otherwise the checkpointed part of the
        model won't have gradients.
    Args:
        functions: A :class:`torch.nn.Sequential` or the list of modules or functions to run sequentially.
        x: A Tensor that is input to :attr:`functions`
        every: checkpoint every-n functions (default: 1)
        flatten (bool): flatten nn.Sequential of nn.Sequentials
        skip_last (bool): skip checkpointing the last function in the sequence if True
        preserve_rng_state (bool, optional, default=True):  Omit stashing and restoring
            the RNG state during each checkpoint.
    Returns:
        Output of running :attr:`functions` sequentially on :attr:`*inputs`
    Example:
        >>> model = nn.Sequential(...)
        >>> input_var = checkpoint_seq(model, input_var, every=2)
    """
    def run_function(start, end, functions):
        def forward(_x):
            for j in range(start, end + 1):
                _x = functions[j](_x)
            return _x
        return forward

    if isinstance(functions, torch.nn.Sequential):
        functions = functions.children()
    if flatten:
        functions = chain.from_iterable(functions)
    if not isinstance(functions, (tuple, list)):
        functions = tuple(functions)

    num_checkpointed = len(functions)
    if skip_last:
        num_checkpointed -= 1
    end = -1
    for start in range(0, num_checkpointed, every):
        end = min(start + every - 1, num_checkpointed - 1)
        x = checkpoint(run_function(start, end, functions), x, preserve_rng_state=preserve_rng_state)
    if skip_last:
        return run_function(end + 1, len(functions) - 1, functions)(x)
    return x


archs = {
    'ViT-T-p16': vit_tiny_patch16_384,
    'ViT-T-p32': vit_tiny_patch32_384,
    'ViT-S-p16': vit_small_patch16_384,
    'ViT-S-p32': vit_small_patch32_384,
    'ViT-B-p16': vit_base_patch16_384,
    'ViT-B-p32': vit_base_patch32_384,
    'ViT-L-p16': vit_large_patch16_384,
    'ViT-L-p32': vit_large_patch32_384,
}