Source code for models.l2p_utils.vit_prompt

""" 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

`How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers`
    - https://arxiv.org/abs/2106.10270

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

Hacked together by / Copyright 2020, Ross Wightman
# ------------------------------------------
# Modification:
# Added code for l2p implementation
# -- Jaeho Lee, dlwogh9344@khu.ac.kr
# ------------------------------------------
"""
import math
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint

from backbone.vit import create_vision_transformer, VisionTransformer as MammothVP
from models.l2p_utils.prompt import Prompt


[docs] class VisionTransformer(MammothVP): def __init__( self, prompt_length=None, embedding_key='cls', prompt_init='uniform', prompt_pool=False, prompt_key=False, pool_size=None, top_k=None, batchwise_prompt=False, prompt_key_init='uniform', head_type='token', use_prompt_mask=False, prompt_shuffle=False, args=None, **kwargs): super().__init__(args=args, **kwargs) self.num_prefix_tokens = 1 if self.class_token else 0 embed_len = self.pos_embed.shape[1] if prompt_length is not None and pool_size is not None and prompt_pool: embed_len += prompt_length * top_k self.pos_embed = nn.Parameter(torch.randn(1, embed_len, self.embed_dim) * .02) self.prompt_pool = prompt_pool self.head_type = head_type self.use_prompt_mask = use_prompt_mask self.prompt_shuffle = prompt_shuffle if prompt_length is not None and pool_size is not None and prompt_pool: self.prompt = Prompt(length=prompt_length, embed_dim=self.embed_dim, embedding_key=embedding_key, prompt_init=prompt_init, prompt_pool=prompt_pool, prompt_key=prompt_key, pool_size=pool_size, top_k=top_k, batchwise_prompt=batchwise_prompt, prompt_key_init=prompt_key_init, prompt_shuffle=self.prompt_shuffle) # Classifier Head self.head = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity() if self.weight_init != 'skip': self.init_weights(self.weight_init)
[docs] def forward_features(self, x, task_id=-1, cls_features=None, train=False): x = self.patch_embed(x) if hasattr(self, 'prompt'): if self.use_prompt_mask and train: start = task_id * self.prompt.top_k end = (task_id + 1) * self.prompt.top_k single_prompt_mask = torch.arange(start, end).to(x.device) prompt_mask = single_prompt_mask.unsqueeze(0).expand(x.shape[0], -1) if end > self.prompt.pool_size: prompt_mask = None else: prompt_mask = None res = self.prompt(x, prompt_mask=prompt_mask, cls_features=cls_features) self.total_prompt_len = res['total_prompt_len'] x = res['prompted_embedding'] else: res = dict() if self.cls_token is not None: x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1) x = self.pos_drop(x + self.pos_embed) x = self.blocks(x) x = self.norm(x) res['x'] = x return res
[docs] def forward_head(self, res, pre_logits: bool = False): x = res['x'] if self.class_token and self.head_type == 'token': x = x[:, 0] elif self.head_type == 'gap' and self.global_pool == 'avg': x = x.mean(dim=1) elif self.head_type == 'prompt' and self.prompt_pool: x = x[:, 1:(1 + self.total_prompt_len)] if self.class_token else x[:, 0:self.total_prompt_len] x = x.mean(dim=1) elif self.head_type == 'token+prompt' and self.prompt_pool and self.class_token: x = x[:, 0:self.total_prompt_len + 1] x = x.mean(dim=1) else: raise ValueError(f'Invalid classifier={self.classifier}') res['pre_logits'] = x x = self.fc_norm(x) res['logits'] = self.head(x) return res
[docs] def forward(self, x, task_id=-1, cls_features=None, train=False, returnt='out'): assert returnt in ('out', 'features', 'both') feats = self.forward_features(x, task_id=task_id, cls_features=cls_features, train=train) if returnt == 'features': return feats res = self.forward_head(feats) if returnt == 'both': return res, feats return res
[docs] def resize_pos_embed(posemb, posemb_new, num_prefix_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 # modify logging.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape) ntok_new = posemb_new.shape[1] if num_prefix_tokens: posemb_prefix, posemb_grid = posemb[:, :num_prefix_tokens], posemb[0, num_prefix_tokens:] # ntok_new -= num_prefix_tokens else: posemb_prefix, posemb_grid = posemb[:, :0], posemb[0] gs_old = int(math.sqrt(len(posemb_grid))) if ntok_new > gs_old ** 2: ntok_new -= gs_old ** 2 # expand cls's pos embedding for prompt tokens posemb_prefix = posemb_prefix.expand(-1, ntok_new, -1) if not len(gs_new): # backwards compatibility gs_new = [int(math.sqrt(ntok_new))] * 2 assert len(gs_new) >= 2 logging.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_prefix, posemb_grid], dim=1) return posemb
[docs] def checkpoint_filter_fn(state_dict, model, adapt_layer_scale=False): """ convert patch embedding weight from manual patchify + linear proj to conv""" import re 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[1] != model.pos_embed.shape[1]: # To resize pos embedding when using model at different size from pretrained weights v = resize_pos_embed( v, model.pos_embed, 0 if getattr(model, 'no_embed_class') else getattr(model, 'num_prefix_tokens', 1), model.patch_embed.grid_size ) elif adapt_layer_scale and 'gamma_' in k: # remap layer-scale gamma into sub-module (deit3 models) k = re.sub(r'gamma_([0-9])', r'ls\1.gamma', k) elif 'pre_logits' in k: # NOTE representation layer removed as not used in latest 21k/1k pretrained weights continue out_dict[k] = v return out_dict
[docs] def vit_base_patch16_224_l2p(pretrained=False, **kwargs): """ ViT-Base (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929). ImageNet-1k weights fine-tuned from in21k @ 224x224, 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_224', base_class=VisionTransformer, filter_fn=checkpoint_filter_fn, pretrained=pretrained, **model_kwargs) return model