Delete CLIP_Explainability/app.py
Browse files- CLIP_Explainability/app.py +0 -801
CLIP_Explainability/app.py
DELETED
@@ -1,801 +0,0 @@
|
|
1 |
-
# from base64 import b64encode
|
2 |
-
from io import BytesIO
|
3 |
-
from math import ceil
|
4 |
-
|
5 |
-
import clip
|
6 |
-
from multilingual_clip import legacy_multilingual_clip, pt_multilingual_clip
|
7 |
-
import numpy as np
|
8 |
-
import pandas as pd
|
9 |
-
from PIL import Image
|
10 |
-
import requests
|
11 |
-
import streamlit as st
|
12 |
-
import torch
|
13 |
-
from torchvision.transforms import ToPILImage
|
14 |
-
from transformers import AutoTokenizer, AutoModel, BertTokenizer
|
15 |
-
|
16 |
-
from CLIP_Explainability.clip_ import load, tokenize
|
17 |
-
from CLIP_Explainability.rn_cam import (
|
18 |
-
# interpret_rn,
|
19 |
-
interpret_rn_overlapped,
|
20 |
-
rn_perword_relevance,
|
21 |
-
)
|
22 |
-
from CLIP_Explainability.vit_cam import (
|
23 |
-
# interpret_vit,
|
24 |
-
vit_perword_relevance,
|
25 |
-
interpret_vit_overlapped,
|
26 |
-
)
|
27 |
-
|
28 |
-
from pytorch_grad_cam.grad_cam import GradCAM
|
29 |
-
|
30 |
-
RUN_LITE = True # Load models for CAM viz for M-CLIP and J-CLIP only
|
31 |
-
|
32 |
-
MAX_IMG_WIDTH = 500
|
33 |
-
MAX_IMG_HEIGHT = 800
|
34 |
-
|
35 |
-
st.set_page_config(layout="wide")
|
36 |
-
|
37 |
-
|
38 |
-
# The `find_best_matches` function compares the text feature vector to the feature vectors of all images and finds the best matches. The function returns the IDs of the best matching images.
|
39 |
-
def find_best_matches(text_features, image_features, image_ids):
|
40 |
-
# Compute the similarity between the search query and each image using the Cosine similarity
|
41 |
-
similarities = (image_features @ text_features.T).squeeze(1)
|
42 |
-
|
43 |
-
# Sort the images by their similarity score
|
44 |
-
best_image_idx = (-similarities).argsort()
|
45 |
-
|
46 |
-
# Return the image IDs of the best matches
|
47 |
-
return [[image_ids[i], similarities[i].item()] for i in best_image_idx]
|
48 |
-
|
49 |
-
|
50 |
-
# The `encode_search_query` function takes a text description and encodes it into a feature vector using the CLIP model.
|
51 |
-
def encode_search_query(search_query, model_type):
|
52 |
-
with torch.no_grad():
|
53 |
-
# Encode and normalize the search query using the multilingual model
|
54 |
-
if model_type == "M-CLIP (multilingual ViT)":
|
55 |
-
text_encoded = st.session_state.ml_model.forward(
|
56 |
-
search_query, st.session_state.ml_tokenizer
|
57 |
-
)
|
58 |
-
text_encoded /= text_encoded.norm(dim=-1, keepdim=True)
|
59 |
-
elif model_type == "J-CLIP (日本語 ViT)":
|
60 |
-
t_text = st.session_state.ja_tokenizer(
|
61 |
-
search_query,
|
62 |
-
padding=True,
|
63 |
-
return_tensors="pt",
|
64 |
-
device=st.session_state.device,
|
65 |
-
)
|
66 |
-
text_encoded = st.session_state.ja_model.get_text_features(**t_text)
|
67 |
-
text_encoded /= text_encoded.norm(dim=-1, keepdim=True)
|
68 |
-
else: # model_type == legacy
|
69 |
-
text_encoded = st.session_state.rn_model(search_query)
|
70 |
-
text_encoded /= text_encoded.norm(dim=-1, keepdim=True)
|
71 |
-
|
72 |
-
# Retrieve the feature vector
|
73 |
-
return text_encoded.to(st.session_state.device)
|
74 |
-
|
75 |
-
|
76 |
-
def clip_search(search_query):
|
77 |
-
if st.session_state.search_field_value != search_query:
|
78 |
-
st.session_state.search_field_value = search_query
|
79 |
-
|
80 |
-
model_type = st.session_state.active_model
|
81 |
-
|
82 |
-
if len(search_query) >= 1:
|
83 |
-
text_features = encode_search_query(search_query, model_type)
|
84 |
-
|
85 |
-
# Compute the similarity between the descrption and each photo using the Cosine similarity
|
86 |
-
# similarities = list((text_features @ photo_features.T).squeeze(0))
|
87 |
-
|
88 |
-
# Sort the photos by their similarity score
|
89 |
-
if model_type == "M-CLIP (multilingual ViT)":
|
90 |
-
matches = find_best_matches(
|
91 |
-
text_features,
|
92 |
-
st.session_state.ml_image_features,
|
93 |
-
st.session_state.image_ids,
|
94 |
-
)
|
95 |
-
elif model_type == "J-CLIP (日本語 ViT)":
|
96 |
-
matches = find_best_matches(
|
97 |
-
text_features,
|
98 |
-
st.session_state.ja_image_features,
|
99 |
-
st.session_state.image_ids,
|
100 |
-
)
|
101 |
-
else: # model_type == legacy
|
102 |
-
matches = find_best_matches(
|
103 |
-
text_features,
|
104 |
-
st.session_state.rn_image_features,
|
105 |
-
st.session_state.image_ids,
|
106 |
-
)
|
107 |
-
|
108 |
-
st.session_state.search_image_ids = [match[0] for match in matches]
|
109 |
-
st.session_state.search_image_scores = {match[0]: match[1] for match in matches}
|
110 |
-
|
111 |
-
|
112 |
-
def string_search():
|
113 |
-
st.session_state.disable_uploader = (
|
114 |
-
RUN_LITE and st.session_state.active_model == "Legacy (multilingual ResNet)"
|
115 |
-
)
|
116 |
-
|
117 |
-
if "search_field_value" in st.session_state:
|
118 |
-
clip_search(st.session_state.search_field_value)
|
119 |
-
|
120 |
-
|
121 |
-
def load_image_features():
|
122 |
-
# Load the image feature vectors
|
123 |
-
if st.session_state.vision_mode == "tiled":
|
124 |
-
ml_image_features = np.load("./image_features/tiled_ml_features.npy")
|
125 |
-
ja_image_features = np.load("./image_features/tiled_ja_features.npy")
|
126 |
-
rn_image_features = np.load("./image_features/tiled_rn_features.npy")
|
127 |
-
elif st.session_state.vision_mode == "stretched":
|
128 |
-
ml_image_features = np.load("./image_features/resized_ml_features.npy")
|
129 |
-
ja_image_features = np.load("./image_features/resized_ja_features.npy")
|
130 |
-
rn_image_features = np.load("./image_features/resized_rn_features.npy")
|
131 |
-
else: # st.session_state.vision_mode == "cropped":
|
132 |
-
ml_image_features = np.load("./image_features/cropped_ml_features.npy")
|
133 |
-
ja_image_features = np.load("./image_features/cropped_ja_features.npy")
|
134 |
-
rn_image_features = np.load("./image_features/cropped_rn_features.npy")
|
135 |
-
|
136 |
-
# Convert features to Tensors: Float32 on CPU and Float16 on GPU
|
137 |
-
device = st.session_state.device
|
138 |
-
if device == "cpu":
|
139 |
-
ml_image_features = torch.from_numpy(ml_image_features).float().to(device)
|
140 |
-
ja_image_features = torch.from_numpy(ja_image_features).float().to(device)
|
141 |
-
rn_image_features = torch.from_numpy(rn_image_features).float().to(device)
|
142 |
-
else:
|
143 |
-
ml_image_features = torch.from_numpy(ml_image_features).to(device)
|
144 |
-
ja_image_features = torch.from_numpy(ja_image_features).to(device)
|
145 |
-
rn_image_features = torch.from_numpy(rn_image_features).to(device)
|
146 |
-
|
147 |
-
st.session_state.ml_image_features = ml_image_features / ml_image_features.norm(
|
148 |
-
dim=-1, keepdim=True
|
149 |
-
)
|
150 |
-
st.session_state.ja_image_features = ja_image_features / ja_image_features.norm(
|
151 |
-
dim=-1, keepdim=True
|
152 |
-
)
|
153 |
-
st.session_state.rn_image_features = rn_image_features / rn_image_features.norm(
|
154 |
-
dim=-1, keepdim=True
|
155 |
-
)
|
156 |
-
|
157 |
-
string_search()
|
158 |
-
|
159 |
-
|
160 |
-
def init():
|
161 |
-
st.session_state.current_page = 1
|
162 |
-
|
163 |
-
# device = "cuda" if torch.cuda.is_available() else "cpu"
|
164 |
-
device = "cpu"
|
165 |
-
|
166 |
-
st.session_state.device = device
|
167 |
-
|
168 |
-
# Load the open CLIP models
|
169 |
-
|
170 |
-
with st.spinner("Loading models and data, please wait..."):
|
171 |
-
ml_model_name = "M-CLIP/XLM-Roberta-Large-Vit-B-16Plus"
|
172 |
-
ml_model_path = "./models/vit_b_16_plus_240-laion400m_e32-699c4b84.pt"
|
173 |
-
|
174 |
-
st.session_state.ml_image_model, st.session_state.ml_image_preprocess = load(
|
175 |
-
ml_model_path, device=device, jit=False
|
176 |
-
)
|
177 |
-
|
178 |
-
st.session_state.ml_model = (
|
179 |
-
pt_multilingual_clip.MultilingualCLIP.from_pretrained(ml_model_name)
|
180 |
-
).to(device)
|
181 |
-
st.session_state.ml_tokenizer = AutoTokenizer.from_pretrained(ml_model_name)
|
182 |
-
|
183 |
-
ja_model_name = "hakuhodo-tech/japanese-clip-vit-h-14-bert-wider"
|
184 |
-
ja_model_path = "./models/ViT-H-14-laion2B-s32B-b79K.bin"
|
185 |
-
|
186 |
-
st.session_state.ja_image_model, st.session_state.ja_image_preprocess = load(
|
187 |
-
ja_model_path, device=device, jit=False
|
188 |
-
)
|
189 |
-
|
190 |
-
st.session_state.ja_model = AutoModel.from_pretrained(
|
191 |
-
ja_model_name, trust_remote_code=True
|
192 |
-
).to(device)
|
193 |
-
st.session_state.ja_tokenizer = AutoTokenizer.from_pretrained(
|
194 |
-
ja_model_name, trust_remote_code=True
|
195 |
-
)
|
196 |
-
|
197 |
-
if not RUN_LITE:
|
198 |
-
st.session_state.rn_image_model, st.session_state.rn_image_preprocess = (
|
199 |
-
clip.load("RN50x4", device=device)
|
200 |
-
)
|
201 |
-
|
202 |
-
st.session_state.rn_model = legacy_multilingual_clip.load_model(
|
203 |
-
"M-BERT-Base-69"
|
204 |
-
).to(device)
|
205 |
-
st.session_state.rn_tokenizer = BertTokenizer.from_pretrained(
|
206 |
-
"bert-base-multilingual-cased"
|
207 |
-
)
|
208 |
-
|
209 |
-
# Load the image IDs
|
210 |
-
st.session_state.images_info = pd.read_csv("./metadata.csv")
|
211 |
-
st.session_state.images_info.set_index("filename", inplace=True)
|
212 |
-
|
213 |
-
with open("./images_list.txt", "r", encoding="utf-8") as images_list:
|
214 |
-
st.session_state.image_ids = list(images_list.read().strip().split("\n"))
|
215 |
-
|
216 |
-
st.session_state.active_model = "M-CLIP (multilingual ViT)"
|
217 |
-
|
218 |
-
st.session_state.vision_mode = "tiled"
|
219 |
-
st.session_state.search_image_ids = []
|
220 |
-
st.session_state.search_image_scores = {}
|
221 |
-
st.session_state.text_table_df = None
|
222 |
-
st.session_state.disable_uploader = (
|
223 |
-
RUN_LITE and st.session_state.active_model == "Legacy (multilingual ResNet)"
|
224 |
-
)
|
225 |
-
|
226 |
-
with st.spinner("Loading models and data, please wait..."):
|
227 |
-
load_image_features()
|
228 |
-
|
229 |
-
|
230 |
-
if "images_info" not in st.session_state:
|
231 |
-
init()
|
232 |
-
|
233 |
-
|
234 |
-
def get_overlay_vis(image, img_dim, image_model):
|
235 |
-
orig_img_dims = image.size
|
236 |
-
|
237 |
-
##### If the features are based on tiled image slices
|
238 |
-
tile_behavior = None
|
239 |
-
|
240 |
-
if st.session_state.vision_mode == "tiled":
|
241 |
-
scaled_dims = [img_dim, img_dim]
|
242 |
-
|
243 |
-
if orig_img_dims[0] > orig_img_dims[1]:
|
244 |
-
scale_ratio = round(orig_img_dims[0] / orig_img_dims[1])
|
245 |
-
if scale_ratio > 1:
|
246 |
-
scaled_dims = [scale_ratio * img_dim, img_dim]
|
247 |
-
tile_behavior = "width"
|
248 |
-
elif orig_img_dims[0] < orig_img_dims[1]:
|
249 |
-
scale_ratio = round(orig_img_dims[1] / orig_img_dims[0])
|
250 |
-
if scale_ratio > 1:
|
251 |
-
scaled_dims = [img_dim, scale_ratio * img_dim]
|
252 |
-
tile_behavior = "height"
|
253 |
-
|
254 |
-
resized_image = image.resize(scaled_dims, Image.LANCZOS)
|
255 |
-
|
256 |
-
if tile_behavior == "width":
|
257 |
-
image_tiles = []
|
258 |
-
for x in range(0, scale_ratio):
|
259 |
-
box = (x * img_dim, 0, (x + 1) * img_dim, img_dim)
|
260 |
-
image_tiles.append(resized_image.crop(box))
|
261 |
-
|
262 |
-
elif tile_behavior == "height":
|
263 |
-
image_tiles = []
|
264 |
-
for y in range(0, scale_ratio):
|
265 |
-
box = (0, y * img_dim, img_dim, (y + 1) * img_dim)
|
266 |
-
image_tiles.append(resized_image.crop(box))
|
267 |
-
|
268 |
-
else:
|
269 |
-
image_tiles = [resized_image]
|
270 |
-
|
271 |
-
elif st.session_state.vision_mode == "stretched":
|
272 |
-
image_tiles = [image.resize((img_dim, img_dim), Image.LANCZOS)]
|
273 |
-
|
274 |
-
else: # vision_mode == "cropped"
|
275 |
-
if orig_img_dims[0] > orig_img_dims[1]:
|
276 |
-
scale_factor = orig_img_dims[0] / orig_img_dims[1]
|
277 |
-
resized_img_dims = (round(scale_factor * img_dim), img_dim)
|
278 |
-
resized_img = image.resize(resized_img_dims)
|
279 |
-
elif orig_img_dims[0] < orig_img_dims[1]:
|
280 |
-
scale_factor = orig_img_dims[1] / orig_img_dims[0]
|
281 |
-
resized_img_dims = (img_dim, round(scale_factor * img_dim))
|
282 |
-
else:
|
283 |
-
resized_img_dims = (img_dim, img_dim)
|
284 |
-
|
285 |
-
resized_img = image.resize(resized_img_dims)
|
286 |
-
|
287 |
-
left = round((resized_img_dims[0] - img_dim) / 2)
|
288 |
-
top = round((resized_img_dims[1] - img_dim) / 2)
|
289 |
-
x_right = round(resized_img_dims[0] - img_dim) - left
|
290 |
-
x_bottom = round(resized_img_dims[1] - img_dim) - top
|
291 |
-
right = resized_img_dims[0] - x_right
|
292 |
-
bottom = resized_img_dims[1] - x_bottom
|
293 |
-
|
294 |
-
# Crop the center of the image
|
295 |
-
image_tiles = [resized_img.crop((left, top, right, bottom))]
|
296 |
-
|
297 |
-
image_visualizations = []
|
298 |
-
image_features = []
|
299 |
-
image_similarities = []
|
300 |
-
|
301 |
-
if st.session_state.active_model == "M-CLIP (multilingual ViT)":
|
302 |
-
text_features = st.session_state.ml_model.forward(
|
303 |
-
st.session_state.search_field_value, st.session_state.ml_tokenizer
|
304 |
-
)
|
305 |
-
|
306 |
-
if st.session_state.device == "cpu":
|
307 |
-
text_features = text_features.float().to(st.session_state.device)
|
308 |
-
else:
|
309 |
-
text_features = text_features.to(st.session_state.device)
|
310 |
-
|
311 |
-
for altered_image in image_tiles:
|
312 |
-
p_image = (
|
313 |
-
st.session_state.ml_image_preprocess(altered_image)
|
314 |
-
.unsqueeze(0)
|
315 |
-
.to(st.session_state.device)
|
316 |
-
)
|
317 |
-
|
318 |
-
vis_t, img_feats, similarity = interpret_vit_overlapped(
|
319 |
-
p_image.type(image_model.dtype),
|
320 |
-
text_features.type(image_model.dtype),
|
321 |
-
image_model.visual,
|
322 |
-
st.session_state.device,
|
323 |
-
img_dim=img_dim,
|
324 |
-
)
|
325 |
-
|
326 |
-
image_visualizations.append(vis_t)
|
327 |
-
image_features.append(img_feats)
|
328 |
-
image_similarities.append(similarity.item())
|
329 |
-
|
330 |
-
elif st.session_state.active_model == "J-CLIP (日本語 ViT)":
|
331 |
-
t_text = st.session_state.ja_tokenizer(
|
332 |
-
st.session_state.search_field_value,
|
333 |
-
return_tensors="pt",
|
334 |
-
device=st.session_state.device,
|
335 |
-
)
|
336 |
-
|
337 |
-
text_features = st.session_state.ja_model.get_text_features(**t_text)
|
338 |
-
|
339 |
-
if st.session_state.device == "cpu":
|
340 |
-
text_features = text_features.float().to(st.session_state.device)
|
341 |
-
else:
|
342 |
-
text_features = text_features.to(st.session_state.device)
|
343 |
-
|
344 |
-
for altered_image in image_tiles:
|
345 |
-
p_image = (
|
346 |
-
st.session_state.ja_image_preprocess(altered_image)
|
347 |
-
.unsqueeze(0)
|
348 |
-
.to(st.session_state.device)
|
349 |
-
)
|
350 |
-
|
351 |
-
vis_t, img_feats, similarity = interpret_vit_overlapped(
|
352 |
-
p_image.type(image_model.dtype),
|
353 |
-
text_features.type(image_model.dtype),
|
354 |
-
image_model.visual,
|
355 |
-
st.session_state.device,
|
356 |
-
img_dim=img_dim,
|
357 |
-
)
|
358 |
-
|
359 |
-
image_visualizations.append(vis_t)
|
360 |
-
image_features.append(img_feats)
|
361 |
-
image_similarities.append(similarity.item())
|
362 |
-
|
363 |
-
else: # st.session_state.active_model == Legacy
|
364 |
-
text_features = st.session_state.rn_model(st.session_state.search_field_value)
|
365 |
-
|
366 |
-
if st.session_state.device == "cpu":
|
367 |
-
text_features = text_features.float().to(st.session_state.device)
|
368 |
-
else:
|
369 |
-
text_features = text_features.to(st.session_state.device)
|
370 |
-
|
371 |
-
for altered_image in image_tiles:
|
372 |
-
p_image = (
|
373 |
-
st.session_state.rn_image_preprocess(altered_image)
|
374 |
-
.unsqueeze(0)
|
375 |
-
.to(st.session_state.device)
|
376 |
-
)
|
377 |
-
|
378 |
-
vis_t = interpret_rn_overlapped(
|
379 |
-
p_image.type(image_model.dtype),
|
380 |
-
text_features.type(image_model.dtype),
|
381 |
-
image_model.visual,
|
382 |
-
GradCAM,
|
383 |
-
st.session_state.device,
|
384 |
-
img_dim=img_dim,
|
385 |
-
)
|
386 |
-
|
387 |
-
text_features_norm = text_features.norm(dim=-1, keepdim=True)
|
388 |
-
text_features_new = text_features / text_features_norm
|
389 |
-
|
390 |
-
image_feats = image_model.encode_image(p_image.type(image_model.dtype))
|
391 |
-
image_feats_norm = image_feats.norm(dim=-1, keepdim=True)
|
392 |
-
image_feats_new = image_feats / image_feats_norm
|
393 |
-
|
394 |
-
similarity = image_feats_new[0].dot(text_features_new[0])
|
395 |
-
|
396 |
-
image_visualizations.append(vis_t)
|
397 |
-
image_features.append(p_image)
|
398 |
-
image_similarities.append(similarity.item())
|
399 |
-
|
400 |
-
transform = ToPILImage()
|
401 |
-
|
402 |
-
vis_images = [transform(vis_t) for vis_t in image_visualizations]
|
403 |
-
|
404 |
-
if st.session_state.vision_mode == "cropped":
|
405 |
-
resized_img.paste(vis_images[0], (left, top))
|
406 |
-
vis_images = [resized_img]
|
407 |
-
|
408 |
-
if orig_img_dims[0] > orig_img_dims[1]:
|
409 |
-
scale_factor = MAX_IMG_WIDTH / orig_img_dims[0]
|
410 |
-
scaled_dims = [MAX_IMG_WIDTH, int(orig_img_dims[1] * scale_factor)]
|
411 |
-
else:
|
412 |
-
scale_factor = MAX_IMG_HEIGHT / orig_img_dims[1]
|
413 |
-
scaled_dims = [int(orig_img_dims[0] * scale_factor), MAX_IMG_HEIGHT]
|
414 |
-
|
415 |
-
if tile_behavior == "width":
|
416 |
-
vis_image = Image.new("RGB", (len(vis_images) * img_dim, img_dim))
|
417 |
-
for x, v_img in enumerate(vis_images):
|
418 |
-
vis_image.paste(v_img, (x * img_dim, 0))
|
419 |
-
activations_image = vis_image.resize(scaled_dims)
|
420 |
-
|
421 |
-
elif tile_behavior == "height":
|
422 |
-
vis_image = Image.new("RGB", (img_dim, len(vis_images) * img_dim))
|
423 |
-
for y, v_img in enumerate(vis_images):
|
424 |
-
vis_image.paste(v_img, (0, y * img_dim))
|
425 |
-
activations_image = vis_image.resize(scaled_dims)
|
426 |
-
|
427 |
-
else:
|
428 |
-
activations_image = vis_images[0].resize(scaled_dims)
|
429 |
-
|
430 |
-
return activations_image, image_features, np.mean(image_similarities)
|
431 |
-
|
432 |
-
|
433 |
-
def visualize_gradcam(image):
|
434 |
-
if "search_field_value" not in st.session_state:
|
435 |
-
return
|
436 |
-
|
437 |
-
header_cols = st.columns([80, 20], vertical_alignment="bottom")
|
438 |
-
with header_cols[0]:
|
439 |
-
st.title("Image + query activation gradients")
|
440 |
-
with header_cols[1]:
|
441 |
-
if st.button("Close"):
|
442 |
-
st.rerun()
|
443 |
-
|
444 |
-
if st.session_state.active_model == "M-CLIP (multilingual ViT)":
|
445 |
-
img_dim = 240
|
446 |
-
image_model = st.session_state.ml_image_model
|
447 |
-
# Sometimes used for token importance viz
|
448 |
-
tokenized_text = st.session_state.ml_tokenizer.tokenize(
|
449 |
-
st.session_state.search_field_value
|
450 |
-
)
|
451 |
-
elif st.session_state.active_model == "Legacy (multilingual ResNet)":
|
452 |
-
img_dim = 288
|
453 |
-
image_model = st.session_state.rn_image_model
|
454 |
-
# Sometimes used for token importance viz
|
455 |
-
tokenized_text = st.session_state.rn_tokenizer.tokenize(
|
456 |
-
st.session_state.search_field_value
|
457 |
-
)
|
458 |
-
else: # J-CLIP
|
459 |
-
img_dim = 224
|
460 |
-
image_model = st.session_state.ja_image_model
|
461 |
-
# Sometimes used for token importance viz
|
462 |
-
tokenized_text = st.session_state.ja_tokenizer.tokenize(
|
463 |
-
st.session_state.search_field_value
|
464 |
-
)
|
465 |
-
|
466 |
-
st.image(image)
|
467 |
-
|
468 |
-
with st.spinner("Calculating..."):
|
469 |
-
# info_text = st.text("Calculating activation regions...")
|
470 |
-
|
471 |
-
activations_image, image_features, similarity_score = get_overlay_vis(
|
472 |
-
image, img_dim, image_model
|
473 |
-
)
|
474 |
-
|
475 |
-
st.markdown(
|
476 |
-
f"**Query text:** {st.session_state.search_field_value} | **Approx. image relevance:** {round(similarity_score.item(), 3)}"
|
477 |
-
)
|
478 |
-
|
479 |
-
st.image(activations_image)
|
480 |
-
|
481 |
-
# image_io = BytesIO()
|
482 |
-
# activations_image.save(image_io, "PNG")
|
483 |
-
# dataurl = "data:image/png;base64," + b64encode(image_io.getvalue()).decode(
|
484 |
-
# "ascii"
|
485 |
-
# )
|
486 |
-
|
487 |
-
# st.html(
|
488 |
-
# f"""<div style="display: flex; flex-direction: column; align-items: center;">
|
489 |
-
# <img src="{dataurl}" />
|
490 |
-
# </div>"""
|
491 |
-
# )
|
492 |
-
|
493 |
-
tokenized_text = [
|
494 |
-
tok.replace("▁", "").replace("#", "") for tok in tokenized_text if tok != "▁"
|
495 |
-
]
|
496 |
-
tokenized_text = [
|
497 |
-
tok
|
498 |
-
for tok in tokenized_text
|
499 |
-
if tok
|
500 |
-
not in ["s", "ed", "a", "the", "an", "ing", "て", "に", "の", "は", "と", "た"]
|
501 |
-
]
|
502 |
-
|
503 |
-
if (
|
504 |
-
len(tokenized_text) > 1
|
505 |
-
and len(tokenized_text) < 25
|
506 |
-
and st.button(
|
507 |
-
"Calculate text importance (may take some time)",
|
508 |
-
)
|
509 |
-
):
|
510 |
-
scores_per_token = {}
|
511 |
-
|
512 |
-
progress_text = f"Processing {len(tokenized_text)} text tokens"
|
513 |
-
progress_bar = st.progress(0.0, text=progress_text)
|
514 |
-
|
515 |
-
for t, tok in enumerate(tokenized_text):
|
516 |
-
token = tok
|
517 |
-
|
518 |
-
for img_feats in image_features:
|
519 |
-
if st.session_state.active_model == "Legacy (multilingual ResNet)":
|
520 |
-
word_rel = rn_perword_relevance(
|
521 |
-
img_feats,
|
522 |
-
st.session_state.search_field_value,
|
523 |
-
image_model,
|
524 |
-
tokenize,
|
525 |
-
GradCAM,
|
526 |
-
st.session_state.device,
|
527 |
-
token,
|
528 |
-
data_only=True,
|
529 |
-
img_dim=img_dim,
|
530 |
-
)
|
531 |
-
else:
|
532 |
-
word_rel = vit_perword_relevance(
|
533 |
-
img_feats,
|
534 |
-
st.session_state.search_field_value,
|
535 |
-
image_model,
|
536 |
-
tokenize,
|
537 |
-
st.session_state.device,
|
538 |
-
token,
|
539 |
-
img_dim=img_dim,
|
540 |
-
)
|
541 |
-
avg_score = np.mean(word_rel)
|
542 |
-
if avg_score == 0 or np.isnan(avg_score):
|
543 |
-
continue
|
544 |
-
|
545 |
-
if token not in scores_per_token:
|
546 |
-
scores_per_token[token] = [1 / avg_score]
|
547 |
-
else:
|
548 |
-
scores_per_token[token].append(1 / avg_score)
|
549 |
-
|
550 |
-
progress_bar.progress(
|
551 |
-
(t + 1) / len(tokenized_text),
|
552 |
-
text=f"Processing token {t+1} of {len(tokenized_text)}",
|
553 |
-
)
|
554 |
-
progress_bar.empty()
|
555 |
-
|
556 |
-
avg_scores_per_token = [
|
557 |
-
np.mean(scores_per_token[tok]) for tok in list(scores_per_token.keys())
|
558 |
-
]
|
559 |
-
|
560 |
-
normed_scores = torch.softmax(torch.tensor(avg_scores_per_token), dim=0)
|
561 |
-
|
562 |
-
token_scores = [f"{round(score.item() * 100, 3)}%" for score in normed_scores]
|
563 |
-
st.session_state.text_table_df = pd.DataFrame(
|
564 |
-
{"token": list(scores_per_token.keys()), "importance": token_scores}
|
565 |
-
)
|
566 |
-
|
567 |
-
st.markdown("**Importance of each text token to relevance score**")
|
568 |
-
st.table(st.session_state.text_table_df)
|
569 |
-
|
570 |
-
|
571 |
-
@st.dialog(" ", width="large")
|
572 |
-
def image_modal(image):
|
573 |
-
visualize_gradcam(image)
|
574 |
-
|
575 |
-
|
576 |
-
def vis_known_image(vis_image_id):
|
577 |
-
image_url = st.session_state.images_info.loc[vis_image_id]["image_url"]
|
578 |
-
image_response = requests.get(image_url)
|
579 |
-
image = Image.open(BytesIO(image_response.content), formats=["JPEG", "GIF", "PNG"])
|
580 |
-
image = image.convert("RGB")
|
581 |
-
|
582 |
-
image_modal(image)
|
583 |
-
|
584 |
-
|
585 |
-
def vis_uploaded_image():
|
586 |
-
uploaded_file = st.session_state.uploaded_image
|
587 |
-
if uploaded_file is not None:
|
588 |
-
# To read file as bytes:
|
589 |
-
bytes_data = uploaded_file.getvalue()
|
590 |
-
image = Image.open(BytesIO(bytes_data), formats=["JPEG", "GIF", "PNG"])
|
591 |
-
image = image.convert("RGB")
|
592 |
-
|
593 |
-
image_modal(image)
|
594 |
-
|
595 |
-
|
596 |
-
def format_vision_mode(mode_stub):
|
597 |
-
return mode_stub.capitalize()
|
598 |
-
|
599 |
-
|
600 |
-
st.title("Explore Japanese visual aesthetics with CLIP models")
|
601 |
-
|
602 |
-
st.markdown(
|
603 |
-
"""
|
604 |
-
<style>
|
605 |
-
[data-testid=stImageCaption] {
|
606 |
-
padding: 0 0 0 0;
|
607 |
-
}
|
608 |
-
[data-testid=stVerticalBlockBorderWrapper] {
|
609 |
-
line-height: 1.2;
|
610 |
-
}
|
611 |
-
[data-testid=stVerticalBlock] {
|
612 |
-
gap: .75rem;
|
613 |
-
}
|
614 |
-
[data-testid=baseButton-secondary] {
|
615 |
-
min-height: 1rem;
|
616 |
-
padding: 0 0.75rem;
|
617 |
-
margin: 0 0 1rem 0;
|
618 |
-
}
|
619 |
-
div[aria-label="dialog"]>button[aria-label="Close"] {
|
620 |
-
display: none;
|
621 |
-
}
|
622 |
-
[data-testid=stFullScreenFrame] {
|
623 |
-
display: flex;
|
624 |
-
flex-direction: column;
|
625 |
-
align-items: center;
|
626 |
-
}
|
627 |
-
</style>
|
628 |
-
""",
|
629 |
-
unsafe_allow_html=True,
|
630 |
-
)
|
631 |
-
|
632 |
-
search_row = st.columns([45, 8, 8, 10, 1, 8, 20], vertical_alignment="center")
|
633 |
-
with search_row[0]:
|
634 |
-
search_field = st.text_input(
|
635 |
-
label="search",
|
636 |
-
label_visibility="collapsed",
|
637 |
-
placeholder="Type something, or click a suggested search below.",
|
638 |
-
on_change=string_search,
|
639 |
-
key="search_field_value",
|
640 |
-
)
|
641 |
-
with search_row[1]:
|
642 |
-
st.button(
|
643 |
-
"Search", on_click=string_search, use_container_width=True, type="primary"
|
644 |
-
)
|
645 |
-
with search_row[2]:
|
646 |
-
st.markdown("**Vision mode:**")
|
647 |
-
with search_row[3]:
|
648 |
-
st.selectbox(
|
649 |
-
"Vision mode",
|
650 |
-
options=["tiled", "stretched", "cropped"],
|
651 |
-
key="vision_mode",
|
652 |
-
help="How to consider images that aren't square",
|
653 |
-
on_change=load_image_features,
|
654 |
-
format_func=format_vision_mode,
|
655 |
-
label_visibility="collapsed",
|
656 |
-
)
|
657 |
-
with search_row[4]:
|
658 |
-
st.empty()
|
659 |
-
with search_row[5]:
|
660 |
-
st.markdown("**CLIP model:**")
|
661 |
-
with search_row[6]:
|
662 |
-
st.selectbox(
|
663 |
-
"CLIP Model:",
|
664 |
-
options=[
|
665 |
-
"M-CLIP (multilingual ViT)",
|
666 |
-
"J-CLIP (日本語 ViT)",
|
667 |
-
"Legacy (multilingual ResNet)",
|
668 |
-
],
|
669 |
-
key="active_model",
|
670 |
-
on_change=string_search,
|
671 |
-
label_visibility="collapsed",
|
672 |
-
)
|
673 |
-
|
674 |
-
canned_searches = st.columns([12, 22, 22, 22, 22], vertical_alignment="top")
|
675 |
-
with canned_searches[0]:
|
676 |
-
st.markdown("**Suggested searches:**")
|
677 |
-
if st.session_state.active_model == "J-CLIP (日本語 ViT)":
|
678 |
-
with canned_searches[1]:
|
679 |
-
st.button(
|
680 |
-
"間",
|
681 |
-
on_click=clip_search,
|
682 |
-
args=["間"],
|
683 |
-
use_container_width=True,
|
684 |
-
)
|
685 |
-
with canned_searches[2]:
|
686 |
-
st.button("奥", on_click=clip_search, args=["奥"], use_container_width=True)
|
687 |
-
with canned_searches[3]:
|
688 |
-
st.button("山", on_click=clip_search, args=["山"], use_container_width=True)
|
689 |
-
with canned_searches[4]:
|
690 |
-
st.button(
|
691 |
-
"花に酔えり 羽織着て刀 さす女",
|
692 |
-
on_click=clip_search,
|
693 |
-
args=["花に酔えり 羽織着て刀 さす女"],
|
694 |
-
use_container_width=True,
|
695 |
-
)
|
696 |
-
else:
|
697 |
-
with canned_searches[1]:
|
698 |
-
st.button(
|
699 |
-
"negative space",
|
700 |
-
on_click=clip_search,
|
701 |
-
args=["negative space"],
|
702 |
-
use_container_width=True,
|
703 |
-
)
|
704 |
-
with canned_searches[2]:
|
705 |
-
st.button("間", on_click=clip_search, args=["間"], use_container_width=True)
|
706 |
-
with canned_searches[3]:
|
707 |
-
st.button("음각", on_click=clip_search, args=["음각"], use_container_width=True)
|
708 |
-
with canned_searches[4]:
|
709 |
-
st.button(
|
710 |
-
"αρνητικός χώρος",
|
711 |
-
on_click=clip_search,
|
712 |
-
args=["αρνητικός χώρος"],
|
713 |
-
use_container_width=True,
|
714 |
-
)
|
715 |
-
|
716 |
-
controls = st.columns([25, 25, 20, 35], gap="large", vertical_alignment="center")
|
717 |
-
with controls[0]:
|
718 |
-
im_per_pg = st.columns([30, 70], vertical_alignment="center")
|
719 |
-
with im_per_pg[0]:
|
720 |
-
st.markdown("**Images/page:**")
|
721 |
-
with im_per_pg[1]:
|
722 |
-
batch_size = st.select_slider(
|
723 |
-
"Images/page:", range(10, 50, 10), label_visibility="collapsed"
|
724 |
-
)
|
725 |
-
with controls[1]:
|
726 |
-
im_per_row = st.columns([30, 70], vertical_alignment="center")
|
727 |
-
with im_per_row[0]:
|
728 |
-
st.markdown("**Images/row:**")
|
729 |
-
with im_per_row[1]:
|
730 |
-
row_size = st.select_slider(
|
731 |
-
"Images/row:", range(1, 6), value=5, label_visibility="collapsed"
|
732 |
-
)
|
733 |
-
num_batches = ceil(len(st.session_state.image_ids) / batch_size)
|
734 |
-
with controls[2]:
|
735 |
-
pager = st.columns([40, 60], vertical_alignment="center")
|
736 |
-
with pager[0]:
|
737 |
-
st.markdown(f"Page **{st.session_state.current_page}** of **{num_batches}** ")
|
738 |
-
with pager[1]:
|
739 |
-
st.number_input(
|
740 |
-
"Page",
|
741 |
-
min_value=1,
|
742 |
-
max_value=num_batches,
|
743 |
-
step=1,
|
744 |
-
label_visibility="collapsed",
|
745 |
-
key="current_page",
|
746 |
-
)
|
747 |
-
with controls[3]:
|
748 |
-
st.file_uploader(
|
749 |
-
"Upload an image",
|
750 |
-
type=["jpg", "jpeg", "gif", "png"],
|
751 |
-
key="uploaded_image",
|
752 |
-
label_visibility="collapsed",
|
753 |
-
on_change=vis_uploaded_image,
|
754 |
-
disabled=st.session_state.disable_uploader,
|
755 |
-
)
|
756 |
-
|
757 |
-
|
758 |
-
if len(st.session_state.search_image_ids) == 0:
|
759 |
-
batch = []
|
760 |
-
else:
|
761 |
-
batch = st.session_state.search_image_ids[
|
762 |
-
(st.session_state.current_page - 1) * batch_size : st.session_state.current_page
|
763 |
-
* batch_size
|
764 |
-
]
|
765 |
-
|
766 |
-
grid = st.columns(row_size)
|
767 |
-
col = 0
|
768 |
-
for image_id in batch:
|
769 |
-
with grid[col]:
|
770 |
-
link_text = st.session_state.images_info.loc[image_id]["permalink"].split("/")[
|
771 |
-
2
|
772 |
-
]
|
773 |
-
# st.image(
|
774 |
-
# st.session_state.images_info.loc[image_id]["image_url"],
|
775 |
-
# caption=st.session_state.images_info.loc[image_id]["caption"],
|
776 |
-
# )
|
777 |
-
st.html(
|
778 |
-
f"""<div style="display: flex; flex-direction: column; align-items: center">
|
779 |
-
<img src="{st.session_state.images_info.loc[image_id]['image_url']}" style="max-width: 100%; max-height: {MAX_IMG_HEIGHT}px" />
|
780 |
-
<div>{st.session_state.images_info.loc[image_id]['caption']} <b>[{round(st.session_state.search_image_scores[image_id], 3)}]</b></div>
|
781 |
-
</div>"""
|
782 |
-
)
|
783 |
-
st.caption(
|
784 |
-
f"""<div style="display: flex; flex-direction: column; align-items: center; position: relative; top: -12px">
|
785 |
-
<a href="{st.session_state.images_info.loc[image_id]['permalink']}">{link_text}</a>
|
786 |
-
<div>""",
|
787 |
-
unsafe_allow_html=True,
|
788 |
-
)
|
789 |
-
if not (
|
790 |
-
RUN_LITE and st.session_state.active_model == "Legacy (multilingual ResNet)"
|
791 |
-
):
|
792 |
-
st.button(
|
793 |
-
"Explain this",
|
794 |
-
on_click=vis_known_image,
|
795 |
-
args=[image_id],
|
796 |
-
use_container_width=True,
|
797 |
-
key=image_id,
|
798 |
-
)
|
799 |
-
else:
|
800 |
-
st.empty()
|
801 |
-
col = (col + 1) % row_size
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|