AdaCLIP / method /adaclip.py
Caoyunkang's picture
first commit
a25563f verified
from typing import Union, List, Optional
import numpy as np
import torch
from pkg_resources import packaging
from torch import nn
from torch.nn import functional as F
from .clip_model import CLIP
from .simple_tokenizer import SimpleTokenizer as _Tokenizer
from sklearn.cluster import KMeans
class ProjectLayer(nn.Module):
def __init__(self, input_dim, output_dim, num_replicas, stack=False, is_array=True):
super(ProjectLayer, self).__init__()
self.head = nn.ModuleList([nn.Linear(input_dim, output_dim) for _ in range(num_replicas)])
self.num_replicas = num_replicas
self.stack = stack
self.is_array = is_array
def forward(self, tokens):
out_tokens = []
for i in range(self.num_replicas):
if self.is_array:
temp = self.head[i](tokens[i][:, 1:, :]) # for ViT, we exclude the class token and only extract patch tokens here.
else:
temp = self.head[i](tokens)
out_tokens.append(temp)
if self.stack:
out_tokens = torch.stack(out_tokens, dim=1)
return out_tokens
class PromptLayer(nn.Module):
def __init__(self, channel, length, depth, is_text, prompting_type, enabled=True):
super(PromptLayer, self).__init__()
self.channel = channel
self.length = length
self.depth = depth
self.is_text = is_text
self.enabled = enabled
self.prompting_type = prompting_type
if self.enabled: # only when enabled, the parameters should be constructed
if 'S' in prompting_type: # static prompts
# learnable
self.static_prompts = nn.ParameterList(
[nn.Parameter(torch.empty(self.length, self.channel))
for _ in range(self.depth)])
for single_para in self.static_prompts:
nn.init.normal_(single_para, std=0.02)
if 'D' in prompting_type: # dynamic prompts
self.dynamic_prompts = [0.] # place holder
def set_dynamic_prompts(self, dynamic_prompts):
self.dynamic_prompts = dynamic_prompts
def forward_text(self, resblock, indx, x, k_x=None, v_x=None, attn_mask: Optional[torch.Tensor] = None):
if self.enabled:
length = self.length
# only prompt the first J layers
if indx < self.depth:
if 'S' in self.prompting_type and 'D' in self.prompting_type: # both
static_prompts = self.static_prompts[indx].unsqueeze(0).expand(x.shape[1], -1, -1)
textual_context = self.dynamic_prompts + static_prompts
elif 'S' in self.prompting_type: # static
static_prompts = self.static_prompts[indx].unsqueeze(0).expand(x.shape[1], -1, -1)
textual_context = static_prompts
elif 'D' in self.prompting_type: # dynamic
textual_context = self.dynamic_prompts
else:
print('You should at least choose one type of prompts when the prompting branches are not none.')
raise NotImplementedError
if indx == 0: # for the first layer
x = x
else:
if indx < self.depth: # replace with learnalbe tokens
prefix = x[:1, :, :]
suffix = x[1 + length:, :, :]
textual_context = textual_context.permute(1, 0, 2).half()
x = torch.cat([prefix, textual_context, suffix], dim=0)
else: # keep the same
x = x
else:
x = x
x, attn_tmp = resblock(q_x=x, k_x=k_x, v_x= v_x, attn_mask=attn_mask)
return x, attn_tmp
def forward_visual(self, resblock, indx, x, k_x=None, v_x=None, attn_mask: Optional[torch.Tensor] = None):
if self.enabled:
length = self.length
# only prompt the first J layers
if indx < self.depth:
if 'S' in self.prompting_type and 'D' in self.prompting_type: # both
static_prompts = self.static_prompts[indx].unsqueeze(0).expand(x.shape[1], -1, -1)
visual_context = self.dynamic_prompts + static_prompts
elif 'S' in self.prompting_type: # static
static_prompts = self.static_prompts[indx].unsqueeze(0).expand(x.shape[1], -1, -1)
visual_context = static_prompts
elif 'D' in self.prompting_type: # dynamic
visual_context = self.dynamic_prompts
else:
print('You should at least choose one type of prompts when the prompting branches are not none.')
raise NotImplementedError
if indx == 0: # for the first layer
visual_context = visual_context.permute(1, 0, 2).half()
x = torch.cat([x, visual_context], dim=0)
else:
if indx < self.depth: # replace with learnalbe tokens
prefix = x[0:x.shape[0] - length, :, :]
visual_context = visual_context.permute(1, 0, 2).half()
x = torch.cat([prefix, visual_context], dim=0)
else: # keep the same
x = x
else:
x = x
x, attn_tmp = resblock(q_x=x, k_x=k_x, v_x= v_x, attn_mask=attn_mask)
if self.enabled:
tokens = x[:x.shape[0] - length, :, :]
else:
tokens = x
return x, tokens, attn_tmp
def forward(self, resblock, indx, x, k_x=None, v_x=None, attn_mask: Optional[torch.Tensor] = None):
if self.is_text:
return self.forward_text(resblock, indx, x, k_x, v_x, attn_mask)
else:
return self.forward_visual(resblock, indx, x, k_x, v_x, attn_mask)
class TextEmbebddingLayer(nn.Module):
def __init__(self, fixed):
super(TextEmbebddingLayer, self).__init__()
self.tokenizer = _Tokenizer()
self.ensemble_text_features = {}
self.prompt_normal = ['{}', 'flawless {}', 'perfect {}', 'unblemished {}', '{} without flaw',
'{} without defect',
'{} without damage']
self.prompt_abnormal = ['damaged {}', 'broken {}', '{} with flaw', '{} with defect', '{} with damage']
self.prompt_state = [self.prompt_normal, self.prompt_abnormal]
self.prompt_templates = ['a bad photo of a {}.',
'a low resolution photo of the {}.',
'a bad photo of the {}.',
'a cropped photo of the {}.',
]
self.fixed = fixed
def tokenize(self, texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> Union[
torch.IntTensor, torch.LongTensor]:
if isinstance(texts, str):
texts = [texts]
sot_token = self.tokenizer.encoder["<|startoftext|>"]
eot_token = self.tokenizer.encoder["<|endoftext|>"]
all_tokens = [[sot_token] + self.tokenizer.encode(text) + [eot_token] for text in texts]
if packaging.version.parse(torch.__version__) < packaging.version.parse("1.8.0"):
result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
else:
result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
for i, tokens in enumerate(all_tokens):
if len(tokens) > context_length:
if truncate:
tokens = tokens[:context_length]
tokens[-1] = eot_token
else:
raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")
result[i, :len(tokens)] = torch.tensor(tokens)
return result
## TODO: text layeer is not compitable with multiple batches...
def forward(self, model, texts, device):
text_feature_list = []
for indx, text in enumerate(texts):
if self.fixed:
if self.ensemble_text_features.get(text) is None:
text_features = self.encode_text(model, text, device)
self.ensemble_text_features[text] = text_features
else:
text_features = self.ensemble_text_features[text]
else:
text_features = self.encode_text(model, text, device)
self.ensemble_text_features[text] = text_features
text_feature_list.append(text_features)
text_features = torch.stack(text_feature_list, dim=0)
text_features = F.normalize(text_features, dim=1)
return text_features
def encode_text(self, model, text, device):
text_features = []
for i in range(len(self.prompt_state)):
text = text.replace('-', ' ')
prompted_state = [state.format(text) for state in self.prompt_state[i]]
prompted_sentence = []
for s in prompted_state:
for template in self.prompt_templates:
prompted_sentence.append(template.format(s))
prompted_sentence = self.tokenize(prompted_sentence, context_length=77).to(device)
class_embeddings = model.encode_text(prompted_sentence)
class_embeddings /= class_embeddings.norm(dim=-1, keepdim=True)
class_embedding = class_embeddings.mean(dim=0)
class_embedding /= class_embedding.norm()
text_features.append(class_embedding)
text_features = torch.stack(text_features, dim=1)
return text_features
class HybridSemanticFusion(nn.Module):
def __init__(self, k_clusters):
super(HybridSemanticFusion, self).__init__()
self.k_clusters = k_clusters
self.n_aggregate_patch_tokens = k_clusters * 5
self.cluster_performer = KMeans(n_clusters=self.k_clusters, n_init="auto")
# @torch.no_grad()
def forward(self, patch_tokens: list, anomaly_maps: list):
anomaly_map = torch.mean(torch.stack(anomaly_maps, dim=1), dim=1)
anomaly_map = torch.softmax(anomaly_map, dim=2)[:, :, 1] # B, L
# extract most abnormal feats
selected_abnormal_tokens = []
k = min(anomaly_map.shape[1], self.n_aggregate_patch_tokens)
top_k_indices = torch.topk(anomaly_map, k=k, dim=1).indices
for layer in range(len(patch_tokens)):
selected_tokens = patch_tokens[layer]. \
gather(dim=1, index=top_k_indices.unsqueeze(-1).
expand(-1, -1, patch_tokens[layer].shape[-1]))
selected_abnormal_tokens.append(selected_tokens)
# use kmeans to extract these centriods
# Stack the data_preprocess
stacked_data = torch.cat(selected_abnormal_tokens, dim=2)
batch_cluster_centers = []
# Perform K-Means clustering
for b in range(stacked_data.shape[0]):
cluster_labels = self.cluster_performer.fit_predict(stacked_data[b, :, :].detach().cpu().numpy())
# Initialize a list to store the cluster centers
cluster_centers = []
# Extract cluster centers for each cluster
for cluster_id in range(self.k_clusters):
collected_cluster_data = []
for abnormal_tokens in selected_abnormal_tokens:
cluster_data = abnormal_tokens[b, :, :][cluster_labels == cluster_id]
collected_cluster_data.append(cluster_data)
collected_cluster_data = torch.cat(collected_cluster_data, dim=0)
cluster_center = torch.mean(collected_cluster_data, dim=0, keepdim=True)
cluster_centers.append(cluster_center)
# Normalize the cluster centers
cluster_centers = torch.cat(cluster_centers, dim=0)
cluster_centers = torch.mean(cluster_centers, dim=0)
batch_cluster_centers.append(cluster_centers)
batch_cluster_centers = torch.stack(batch_cluster_centers, dim=0)
batch_cluster_centers = F.normalize(batch_cluster_centers, dim=1)
return batch_cluster_centers
# # preprocess
# # compute the anomaly map
# anomaly_map = torch.mean(torch.stack(anomaly_maps, dim=1), dim=1)
# anomaly_map = torch.softmax(anomaly_map, dim=2)[:, :, 1] # B, L
#
# # compute the average multi-hierarchy patch embeddings
# avg_patch_tokens = torch.mean(torch.stack(patch_tokens, dim=0), dim=0) # B, L, C
#
# # Initialize a list to store the centroids of clusters with the largest anomaly scores
# cluster_centroids = []
#
# # loop across the batch size
# for b in range(avg_patch_tokens.shape[0]):
# # step1: group features into clusters
# cluster_labels = self.cluster_performer.fit_predict(avg_patch_tokens[b, :, :].detach().cpu().numpy())
#
# # step2: compute the anomaly scores for individual clusters via the anomaly map
# # Convert cluster labels back to tensor
# cluster_labels = torch.tensor(cluster_labels).to(avg_patch_tokens.device)
# cluster_anomaly_scores = {}
# for label in torch.unique(cluster_labels):
# cluster_indices = torch.where(cluster_labels == label)[0]
# cluster_anomaly_scores[label.item()] = anomaly_map[b, cluster_indices].mean().item()
#
# # step3: select the cluster with the largest anomaly score and then compute its centroid by averaging the
# # corresponding avg_patch_tokens
# # Find the cluster with the largest anomaly score
# largest_anomaly_cluster = max(cluster_anomaly_scores, key=cluster_anomaly_scores.get)
#
# # Get the indices of the tokens belonging to the largest anomaly cluster
# largest_anomaly_cluster_indices = torch.where(cluster_labels == largest_anomaly_cluster)[0]
#
# # Compute the centroid of the largest anomaly cluster by averaging the corresponding avg_patch_tokens
# centroid = avg_patch_tokens[b, largest_anomaly_cluster_indices, :].mean(dim=0)
#
# # Append the centroid to the list of cluster centroids
# cluster_centroids.append(centroid)
#
# # Convert the list of centroids to a tensor
# cluster_centroids = torch.stack(cluster_centroids, dim=0)
# cluster_centroids = F.normalize(cluster_centroids, dim=1)
# return cluster_centroids
class AdaCLIP(nn.Module):
def __init__(self, freeze_clip: CLIP, text_channel: int, visual_channel: int,
prompting_length: int, prompting_depth: int, prompting_branch: str, prompting_type: str,
use_hsf: bool, k_clusters: int,
output_layers: list, device: str, image_size: int):
super(AdaCLIP, self).__init__()
self.freeze_clip = freeze_clip
self.visual = self.freeze_clip.visual
self.transformer = self.freeze_clip.transformer
self.token_embedding = self.freeze_clip.token_embedding
self.positional_embedding = self.freeze_clip.positional_embedding
self.ln_final = self.freeze_clip.ln_final
self.text_projection = self.freeze_clip.text_projection
self.attn_mask = self.freeze_clip.attn_mask
self.output_layers = output_layers
self.prompting_branch = prompting_branch
self.prompting_type = prompting_type
self.prompting_depth = prompting_depth
self.prompting_length = prompting_length
self.use_hsf = use_hsf
self.k_clusters = k_clusters
if 'L' in self.prompting_branch:
self.enable_text_prompt = True
else:
self.enable_text_prompt = False
if 'V' in self.prompting_branch:
self.enable_visual_prompt = True
else:
self.enable_visual_prompt = False
self.text_embedding_layer = TextEmbebddingLayer(fixed=(not self.enable_text_prompt))
self.text_prompter = PromptLayer(text_channel, prompting_length, prompting_depth, is_text=True,
prompting_type=prompting_type,
enabled=self.enable_text_prompt)
self.visual_prompter = PromptLayer(visual_channel, prompting_length, prompting_depth, is_text=False,
prompting_type=prompting_type,
enabled=self.enable_visual_prompt)
self.patch_token_layer = ProjectLayer(
visual_channel,
text_channel,
len(output_layers), stack=False, is_array=True
)
self.cls_token_layer = ProjectLayer(
text_channel,
text_channel,
1, stack=False, is_array=False
)
if 'D' in self.prompting_type: # dynamic prompts
self.dynamic_visual_prompt_generator = ProjectLayer(text_channel,
visual_channel,
prompting_length,
stack=True,
is_array=False)
self.dynamic_text_prompt_generator = ProjectLayer(text_channel,
text_channel,
prompting_length,
stack=True,
is_array=False)
if self.use_hsf:
self.HSF = HybridSemanticFusion(k_clusters)
self.image_size = image_size
self.device = device
def generate_and_set_dynamic_promtps(self, image):
with torch.no_grad():
# extract image features
image_features, _ = self.visual.forward(image, self.output_layers)
dynamic_visual_prompts = self.dynamic_visual_prompt_generator(image_features)
dynamic_text_prompts = self.dynamic_text_prompt_generator(image_features)
self.visual_prompter.set_dynamic_prompts(dynamic_visual_prompts)
self.text_prompter.set_dynamic_prompts(dynamic_text_prompts)
def encode_image(self, image):
x = image
# to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
if self.visual.input_patchnorm:
# einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
x = x.reshape(x.shape[0], x.shape[1],
self.visual.grid_size[0],
self.visual.patch_size[0],
self.visual.grid_size[1],
self.visual.patch_size[1])
x = x.permute(0, 2, 4, 1, 3, 5)
x = x.reshape(x.shape[0], self.visual.grid_size[0] * self.visual.grid_size[1], -1)
x = self.visual.patchnorm_pre_ln(x)
x = self.visual.conv1(x)
else:
x = self.visual.conv1(x) # shape = [*, width, grid, grid]
x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]
x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]
# class embeddings and positional embeddings
x = torch.cat(
[self.visual.class_embedding.to(x.dtype) +
torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
x], dim=1) # shape = [*, grid ** 2 + 1, width]
x = x + self.visual.positional_embedding.to(x.dtype)
# a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
x = self.visual.patch_dropout(x)
x = self.visual.ln_pre(x)
patch_embedding = x
x = x.permute(1, 0, 2) # NLD -> LND
patch_tokens = []
for indx, r in enumerate(self.visual.transformer.resblocks):
x, tokens, attn_tmp = self.visual_prompter(r, indx, x, k_x=None, v_x=None, attn_mask=None)
if (indx + 1) in self.output_layers:
patch_tokens.append(tokens)
x = x.permute(1, 0, 2) # LND -> NLD
patch_tokens = [patch_tokens[t].permute(1, 0, 2) for t in range(len(patch_tokens))] # LND -> NLD
if self.visual.attn_pool is not None:
x = self.visual.attn_pool(x)
x = self.visual.ln_post(x)
pooled, tokens = self.visual._global_pool(x)
else:
pooled, tokens = self.visual._global_pool(x)
pooled = self.visual.ln_post(pooled)
if self.visual.proj is not None:
pooled = pooled @ self.visual.proj
return pooled, patch_tokens, patch_embedding
def proj_visual_tokens(self, image_features, patch_tokens):
# for patch tokens
proj_patch_tokens = self.patch_token_layer(patch_tokens)
for layer in range(len(proj_patch_tokens)):
proj_patch_tokens[layer] /= proj_patch_tokens[layer].norm(dim=-1, keepdim=True)
# for cls tokens
proj_cls_tokens = self.cls_token_layer(image_features)[0]
proj_cls_tokens /= proj_cls_tokens.norm(dim=-1, keepdim=True)
return proj_cls_tokens, proj_patch_tokens
def encode_text(self, text):
cast_dtype = self.transformer.get_cast_dtype()
x = self.token_embedding(text).to(cast_dtype) # [batch_size, n_ctx, d_model]
x = x + self.positional_embedding.to(cast_dtype)
x = x.permute(1, 0, 2) # NLD -> LND
for indx, r in enumerate(self.transformer.resblocks):
# add prompt here
x, attn_tmp = self.text_prompter(r, indx, x, k_x=None, v_x=None, attn_mask=self.attn_mask)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_final(x) # [batch_size, n_ctx, transformer.width]
# take features from the eot embedding (eot_token is the highest number in each sequence)
x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
return x
def visual_text_similarity(self, image_feature, patch_token, text_feature, aggregation):
anomaly_maps = []
for layer in range(len(patch_token)):
anomaly_map = (100.0 * patch_token[layer] @ text_feature)
anomaly_maps.append(anomaly_map)
if self.use_hsf:
alpha = 0.2
clustered_feature = self.HSF.forward(patch_token, anomaly_maps)
# aggregate the class token and the clustered features for more comprehensive information
cur_image_feature = alpha * clustered_feature + (1 - alpha) * image_feature
cur_image_feature = F.normalize(cur_image_feature, dim=1)
else:
cur_image_feature = image_feature
anomaly_score = (100.0 * cur_image_feature.unsqueeze(1) @ text_feature)
anomaly_score = anomaly_score.squeeze(1)
anomaly_score = torch.softmax(anomaly_score, dim=1)
# NOTE: this bilinear interpolation is not unreproducible and may occasionally lead to unstable ZSAD performance.
for i in range(len(anomaly_maps)):
B, L, C = anomaly_maps[i].shape
H = int(np.sqrt(L))
anomaly_maps[i] = anomaly_maps[i].permute(0, 2, 1).view(B, 2, H, H)
anomaly_maps[i] = F.interpolate(anomaly_maps[i], size=self.image_size, mode='bilinear', align_corners=True)
if aggregation: # in the test stage, we firstly aggregate logits from all hierarchies and then do the softmax normalization
anomaly_map = torch.mean(torch.stack(anomaly_maps, dim=1), dim=1)
anomaly_map = torch.softmax(anomaly_map, dim=1)
anomaly_map = (anomaly_map[:, 1:, :, :] + 1 - anomaly_map[:, 0:1, :, :]) / 2.0
anomaly_score = anomaly_score[:, 1]
return anomaly_map, anomaly_score
else: # otherwise, we do the softmax normalization for individual hierarchies
for i in range(len(anomaly_maps)):
anomaly_maps[i] = torch.softmax(anomaly_maps[i], dim=1)
return anomaly_maps, anomaly_score
def extract_feat(self, image, cls_name):
if 'D' in self.prompting_type:
self.generate_and_set_dynamic_promtps(image) # generate and set dynamic prompts for corresponding prompters
if self.enable_visual_prompt:
image_features, patch_tokens, _ = self.encode_image(image)
else:
with torch.no_grad():
image_features, patch_tokens, _ = self.encode_image(image)
if self.enable_text_prompt:
text_features = self.text_embedding_layer(self, cls_name, self.device)
else:
with torch.no_grad():
text_features = self.text_embedding_layer(self, cls_name, self.device)
proj_cls_tokens, proj_patch_tokens = self.proj_visual_tokens(image_features, patch_tokens)
return proj_cls_tokens, proj_patch_tokens, text_features
@torch.cuda.amp.autocast()
def forward(self, image, cls_name, aggregation=True):
# extract features for images and texts
image_features, patch_tokens, text_features = self.extract_feat(image, cls_name)
anomaly_map, anomaly_score = self.visual_text_similarity(image_features, patch_tokens, text_features, aggregation)
if aggregation:
anomaly_map = anomaly_map # tensor
anomaly_score = anomaly_score
anomaly_map = anomaly_map.squeeze(1)
return anomaly_map, anomaly_score
else:
anomaly_maps = anomaly_map # list
anomaly_score = anomaly_score
return anomaly_maps, anomaly_score