vit_l2.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
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
DeiT model defs and weights from https://github.com/facebookresearch/deit,
paper `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877
Hacked together by / Copyright 2020 Ross Wightman
"""
import math
import logging
from functools import partial
from collections import OrderedDict
from copy import Error, deepcopy

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

from utils import batch_index_select

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

# coding=UTF-8
import os

_logger = logging.getLogger(__name__)

file = 'score_placeholder.json'


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': .9, 'interpolation': 'bicubic',
        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
        'first_conv': 'patch_embed.proj', 'classifier': 'head',
        **kwargs
    }


default_cfgs = {
    # patch models (my experiments)
    'vit_small_patch16_224': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth',
    ),

    # patch models (weights ported from official Google JAX impl)
    'vit_base_patch16_224': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth',
        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5),
    ),
    'vit_base_patch32_224': _cfg(
        url='',  # no official model weights for this combo, only for in21k
        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    '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(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_224-4ee7a4dc.pth',
        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    'vit_large_patch32_224': _cfg(
        url='',  # no official model weights for this combo, only for in21k
        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    '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),

    # patch models, imagenet21k (weights ported from official Google JAX impl)
    'vit_base_patch16_224_in21k': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch16_224_in21k-e5005f0a.pth',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    'vit_base_patch32_224_in21k': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch32_224_in21k-8db57226.pth',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    'vit_large_patch16_224_in21k': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch16_224_in21k-606da67d.pth',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    'vit_large_patch32_224_in21k': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch32_224_in21k-9046d2e7.pth',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    'vit_huge_patch14_224_in21k': _cfg(
        hf_hub='timm/vit_huge_patch14_224_in21k',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),

    # hybrid models (weights ported from official Google JAX impl)
    'vit_base_resnet50_224_in21k': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_224_in21k-6f7c7740.pth',
        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=0.9, first_conv='patch_embed.backbone.stem.conv'),
    'vit_base_resnet50_384': _cfg(
        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_384-9fd3c705.pth',
        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0, first_conv='patch_embed.backbone.stem.conv'),

    # hybrid models (my experiments)
    'vit_small_resnet26d_224': _cfg(),
    'vit_small_resnet50d_s3_224': _cfg(),
    'vit_base_resnet26d_224': _cfg(),
    'vit_base_resnet50d_224': _cfg(),

    # deit models (FB weights)
    'vit_deit_tiny_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth'),
    'vit_deit_small_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_small_patch16_224-cd65a155.pth'),
    'vit_deit_base_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth',),
    'vit_deit_base_patch16_384': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_384-8de9b5d1.pth',
        input_size=(3, 384, 384), crop_pct=1.0),
    'vit_deit_tiny_distilled_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth'),
    'vit_deit_small_distilled_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth'),
    'vit_deit_base_distilled_patch16_224': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_224-df68dfff.pth', ),
    'vit_deit_base_distilled_patch16_384': _cfg(
        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_384-d0272ac0.pth',
        input_size=(3, 384, 384), crop_pct=1.0),
}


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
        # 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=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def softmax_with_policy(self, attn, policy, eps=1e-6):
        B, N, _ = policy.size()
        B, H, N, N = attn.size()
        attn_policy = policy.reshape(B, 1, 1, N)  # * policy.reshape(B, 1, N, 1)
        eye = torch.eye(N, dtype=attn_policy.dtype, device=attn_policy.device).view(1, 1, N, N)
        attn_policy = attn_policy + (1.0 - attn_policy) * eye
        max_att = torch.max(attn, dim=-1, keepdim=True)[0]
        attn = attn - max_att

        attn = attn.to(torch.float32).exp_() * attn_policy.to(torch.float32)
        attn = (attn + eps/N) / (attn.sum(dim=-1, keepdim=True) + eps)
        return attn.type_as(max_att)

    def forward(self, x, policy):
        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)
        #with torch.cuda.amp.autocast(enabled=False):
        q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)

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

        if policy is None:
                attn = attn.softmax(dim=-1)
        elif not self.training:
                attn = self.softmax_with_policy(attn, policy, 0)
        else:
                attn = self.softmax_with_policy(attn, policy, 1e-6)

        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., 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, policy=None, i=None):

        if i is None:  # first 3 layers, no need to save rep. token for placeholder
            x = x + self.drop_path(self.attn(self.norm1(x), policy=policy))
            x = x + self.drop_path(self.mlp(self.norm2(x)))
            return x
        else:
            x_1 = self.drop_path(self.attn(self.norm1(x), policy=policy))
            rep_1 = x_1[:,-1:]   # output rep token of MSA [96, 1, 384]
            x = x + x_1   #skip connection

            x_2 = self.drop_path(self.mlp(self.norm2(x)))
            rep_2 = x_2[:, -1:]   # output rep token of FFN [96, 1, 384]
            x = x + x_2
            return x, rep_1, rep_2


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]))
                if isinstance(o, (list, tuple)):
                    o = o[-1]  # last feature if backbone outputs list/tuple of features
                feature_size = o.shape[-2:]
                feature_dim = o.shape[1]
                backbone.train(training)
        else:
            feature_size = to_2tuple(feature_size)
            if hasattr(self.backbone, 'feature_info'):
                feature_dim = self.backbone.feature_info.channels()[-1]
            else:
                feature_dim = self.backbone.num_features
        self.num_patches = feature_size[0] * feature_size[1]
        self.proj = nn.Conv2d(feature_dim, embed_dim, 1)

    def forward(self, x):
        x = self.backbone(x)
        if isinstance(x, (list, tuple)):
            x = x[-1]  # last feature if backbone outputs list/tuple of features
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x


class PredictorLG(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, embed_dim=384):
        super().__init__()
        self.in_conv = nn.Sequential(
            nn.LayerNorm(embed_dim),
            nn.Linear(embed_dim, embed_dim),
            nn.GELU()
        )

        self.out_conv = nn.Sequential(
            nn.Linear(embed_dim, embed_dim // 2),
            nn.GELU(),
            nn.Linear(embed_dim // 2, embed_dim // 4),
            nn.GELU(),
            nn.Linear(embed_dim // 4, 2),
            nn.LogSoftmax(dim=-1)
        )

    def forward(self, x, policy):
        x = self.in_conv(x)
        B, N, C = x.size()
        local_x = x[:,:, :C//2]
        global_x = (x[:,:, C//2:] * policy).sum(dim=1, keepdim=True) / torch.sum(policy, dim=1, keepdim=True)
        x = torch.cat([local_x, global_x.expand(B, N, C//2)], dim=-1) # 后一半 global 大家都一样。自己的和全局的做merge。
        return self.out_conv(x)

class MultiheadPredictorLG(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, num_heads=6, embed_dim=384):
        super().__init__()

        self.num_heads=num_heads
        self.embed_dim = embed_dim
        onehead_in_conv = nn.Sequential(
            nn.LayerNorm(embed_dim // num_heads),
            nn.Linear(embed_dim // num_heads, embed_dim // num_heads),
            nn.GELU()
        )

        onehead_out_conv = nn.Sequential(
            nn.Linear(embed_dim // num_heads, embed_dim // num_heads  // 2),
            nn.GELU(),
            nn.Linear(embed_dim // num_heads // 2, embed_dim // num_heads // 4),
            nn.GELU(),
            nn.Linear(embed_dim // num_heads // 4, 2),
            #nn.LogSoftmax(dim=-1)
        )

        in_conv_list = [onehead_in_conv for _ in range(num_heads)]
        out_conv_list = [onehead_out_conv for _ in range(num_heads)]

        self.in_conv = nn.ModuleList(in_conv_list)
        self.out_conv = nn.ModuleList(out_conv_list)

    def forward(self, x, policy):

        multihead_score = 0
        multihead_softmax_score = 0
        for i in range(self.num_heads):
            x_single = x[:,:,self.embed_dim//self.num_heads*i:self.embed_dim//self.num_heads*(i+1)]   #([96, 196, 64])
            x_single = self.in_conv[i](x_single)
            B, N, C = x_single.size()       #([96, 196, 64])
            local_x = x_single[:,:, :C//2]  #([96, 196, 32])
            global_x = (x_single[:,:, C//2:] * policy).sum(dim=1, keepdim=True) / torch.sum(policy, dim=1, keepdim=True)  #([96, 1, 32])
            x_single = torch.cat([local_x, global_x.expand(B, N, C//2)], dim=-1)  #([96, 196, 64])
            x_single = self.out_conv[i](x_single) #([96, 196, 2])

            # for placeholder
            m = nn.Softmax(dim=-1)
            score_softmax = m(x_single)
            multihead_softmax_score += score_softmax

            # for gumble
            n = nn.LogSoftmax(dim=-1)
            score_single = n(x_single)
            multihead_score += score_single

        # for gumble
        multihead_score = multihead_score / self.num_heads  # ([96, 196, 2])

        # for placeholder
        multihead_softmax_score = multihead_softmax_score / self.num_heads  # get softmax keep/drop probability

        return multihead_score, multihead_softmax_score  #, represent_token, placeholder_weights, placeholder_score3


class VisionTransformerDiffPruning(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, embed_dim=768, depth=12,
                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None,
                 pruning_loc=None, token_ratio=None, distill=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
            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
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            hybrid_backbone (nn.Module): CNN backbone to use in-place of PatchEmbed module
            norm_layer: (nn.Module): normalization layer
        """
        super().__init__()

        print('## diff vit pruning method')
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        self.depth = depth

        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)
        num_patches = self.patch_embed.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        #self.pre_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) # 咱们多生成一个 pre_token
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
        #self.pos_embed_re = nn.Parameter(torch.zeros(1, 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, 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)

        # Representation layer
        if representation_size:
            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
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

        predictor_list = [MultiheadPredictorLG(num_heads,embed_dim) for _ in range(len(pruning_loc))]

        self.score_predictor = nn.ModuleList(predictor_list)

        self.distill = distill

        self.pruning_loc = pruning_loc
        self.token_ratio = token_ratio

        trunc_normal_(self.pos_embed, std=.02)
        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)

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

    def get_classifier(self):
        return self.head

    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()

    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 = x + self.pos_embed

        x = self.pos_drop(x)

        p_count = 0
        out_pred_prob = []
        init_n = 14 * 14
        sparse = []
        score_dict = {}
        policy = torch.ones(B, init_n + 1, 1, dtype=x.dtype, device=x.device)

        prev_decision = torch.ones(B, init_n, 1, dtype=x.dtype, device=x.device)
        for i, blk in enumerate(self.blocks):
            if i in self.pruning_loc:

                spatial_x = x[:, 1:]
                if i != self.pruning_loc[0]:
                    rep_decision = torch.ones(B, 1, 1, dtype=x.dtype, device=x.device)
                    prev_decision = torch.cat([prev_decision, rep_decision], dim=1)
                pred_score, softmax_score = self.score_predictor[p_count](spatial_x, prev_decision)
                pred_score = pred_score.reshape(B, -1, 2)
                softmax_score = softmax_score.reshape(B, -1, 2)
                #-------------------- 确定 informative token 和 placeholder 的 mask
                if i == self.pruning_loc[0]:
                    hard_keep_decision = F.gumbel_softmax(pred_score, hard=True)[:, :, 0:1] *  prev_decision
                    hard_drop_decision = (1 - hard_keep_decision) - (1 - prev_decision) #  current drop decision
                else:
                    hard_keep_decision_all = F.gumbel_softmax(pred_score, hard=True)[:, :, 0:1] *  prev_decision
                    hard_keep_decision = torch.cat([hard_keep_decision_all[:,:-1], rep_decision], dim=1)
                    hard_drop_decision = (1 - hard_keep_decision) - (1 - prev_decision)
                ############### end

                ###get representative token  (regularization)
                softmax_score = softmax_score[:, :, 0:1]  # softmax score of all tokens to keep
                placeholder_score = softmax_score * hard_drop_decision  #keep score of only placeholder tokens
                x2 = spatial_x * placeholder_score  # placehoder score [96, 196, 384]
                x2_sum = torch.sum(x2, dim=1)  # sum by the N dimension, output (B,N,C)-->(B,C) [96, 384]
                x2_sum = torch.unsqueeze(x2_sum, dim=1)  # resize to (B,1,C)  [96, 1, 384]
                #--------------------
                placeholder_score_sum = torch.sum(placeholder_score, dim=1)  # sum of token score, [96, 196, 1]-->[96, 1]
                placeholder_score_sum = torch.unsqueeze(placeholder_score_sum, dim=1)  # resize to [96, 1, 1]
                #--------------------
                represent_token = x2_sum / placeholder_score_sum  # regularization --> [96, 1, 384] representitave token

                rep_mean = represent_token.mean()
                if torch.isnan(rep_mean):
                    print('has nan')
                    represent_token = torch.nan_to_num(represent_token, nan = 1e-6)

                if i == self.pruning_loc[0]:
                    x = torch.cat((x,represent_token), dim=1)
                else:
                    represent_token = x[:, -1:, :] + represent_token
                    x = x[:,:-1]
                    x = torch.cat((x,represent_token), dim=1)
                    hard_keep_decision = hard_keep_decision[:,:-1]

                if self.training:
                    out_pred_prob.append(hard_keep_decision.reshape(B, init_n))
                    cls_policy = torch.ones(B, 1, 1, dtype=hard_keep_decision.dtype, device=hard_keep_decision.device)
                    rep_policy = torch.ones(B, 1, 1, dtype=hard_keep_decision.dtype, device=hard_keep_decision.device)
                    policy = torch.cat([cls_policy, hard_keep_decision, rep_policy], dim=1)
                    x = blk(x, policy=policy)   #when i=None, means no output rep. token. Such as first 3 layers.
                    prev_decision = hard_keep_decision
                else:
                    cls_policy = torch.ones(B, 1,1, dtype=hard_keep_decision.dtype, device=hard_keep_decision.device)
                    rep_policy = torch.ones(B, 1, 1, dtype=hard_keep_decision.dtype, device=hard_keep_decision.device)
                    policy = torch.cat([cls_policy, hard_keep_decision, rep_policy], dim=1)
                    x = blk(x, policy=policy)   #when i=None, means no output rep. token. Such as first 3 layers.
                    prev_decision = hard_keep_decision
                    zeros, unzeros = test_irregular_sparsity(p_count, policy)
                    sparse.append([zeros, unzeros])
                    score = pred_score[:, :, 0:1].cpu().numpy().tolist()
                    score_dict[p_count] = score[0]
                p_count += 1

            ### first 3 layers. No rep token, placeholder, etc.
            else:
                x = blk(x, policy)
            ############### end

        x = self.norm(x)
        features = x[:, 1:-1]
        x = x[:, 0]
        x = self.pre_logits(x)
        x = self.head(x)

        if self.training:
            if self.distill:
                return x, features, prev_decision.detach(), out_pred_prob
            else:
                return x, out_pred_prob
        else:
            with open(file, 'a') as f: # ins
                json.dump(score_dict, f)
                f.write('\n')
            sparse = torch.FloatTensor(sparse).cuda()
            return x, sparse.detach()

class VisionTransformerTeacher(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, embed_dim=768, depth=12,
                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None):
        """
        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
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            hybrid_backbone (nn.Module): CNN backbone to use in-place of PatchEmbed module
            norm_layer: (nn.Module): normalization layer
        """
        super().__init__()

        print('## diff vit pruning method')
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)

        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)
        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 + 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, 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)

        # Representation layer
        if representation_size:
            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
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

        trunc_normal_(self.pos_embed, std=.02)
        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)

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

    def get_classifier(self):
        return self.head

    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()

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

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

        feature = self.norm(x)
        cls = feature[:, 0]
        tokens = feature[:, 1:]
        cls = self.pre_logits(cls)
        cls = self.head(cls)
        return cls, tokens

def resize_pos_embed(posemb, posemb_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 True:
        posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]
        ntok_new -= 1
    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)
        out_dict[k] = v
    return out_dict

def test_irregular_sparsity(name,matrix):

    # continue
    zeros = np.sum(matrix.cpu().detach().numpy() == 0)

    non_zeros = np.sum(matrix.cpu().detach().numpy() != 0)

    # print(name, non_zeros)
    #print(" {}, all weights: {}, irregular zeros: {}, irregular sparsity is: {:.4f}".format( name, zeros+non_zeros, zeros, zeros / (zeros + non_zeros)))
    # print(non_zeros+zeros)
    # total_nonzeros += 128000

    return zeros,non_zeros