Initial Release
Browse files
app.py
CHANGED
@@ -12,6 +12,8 @@ print(f'gr version : {gr.__version__}')
|
|
12 |
import pickle
|
13 |
import random
|
14 |
|
|
|
|
|
15 |
# %%
|
16 |
model_name = 'trclip-vitl14-e10'
|
17 |
if not os.path.exists(model_name):
|
@@ -28,24 +30,37 @@ if not os.path.exists('TrCaption-trclip-vitl14-e10'):
|
|
28 |
|
29 |
# %%
|
30 |
|
31 |
-
def load_image_embeddings():
|
32 |
path = os.path.join('TrCaption-trclip-vitl14-e10', 'image_embeddings')
|
33 |
bs = 100_000
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
-
for i in tqdm(range(0, 3_100_000, bs), desc='Loading TrCaption Image embeddings'):
|
37 |
-
with open(os.path.join(path, f'image_em_{i}.pkl'), 'rb') as f:
|
38 |
-
embeddings.append(pickle.load(f))
|
39 |
-
return torch.cat(embeddings, dim=0)
|
40 |
|
41 |
-
def load_text_embeddings():
|
42 |
path = os.path.join('TrCaption-trclip-vitl14-e10', 'text_embeddings')
|
43 |
bs = 100_000
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
|
51 |
def load_metadata():
|
@@ -56,61 +71,64 @@ def load_metadata():
|
|
56 |
trcap_urls = metadata['image_urls']
|
57 |
return trcap_texts, trcap_urls
|
58 |
|
59 |
-
|
|
|
60 |
part = index // bs
|
61 |
idx = index % bs
|
62 |
-
with open(os.path.join('TrCaption-trclip-vitl14-e10', f'{type}_embeddings', f'{type}_em_{part*bs}.pkl'), 'rb') as f:
|
63 |
embeddings = pickle.load(f)
|
64 |
return embeddings[idx]
|
65 |
|
66 |
-
# %%
|
67 |
|
68 |
-
|
69 |
-
text_embeddings = None
|
70 |
-
|
71 |
-
#%%
|
72 |
trcap_texts, trcap_urls = load_metadata()
|
73 |
# %%
|
|
|
74 |
model_path = os.path.join(model_name, 'pytorch_model.bin')
|
75 |
trclip = Trclip(model_path, clip_model='ViT-L/14', device='cpu')
|
76 |
-
#%%
|
77 |
-
import psutil
|
78 |
-
|
79 |
-
print(f"First used memory {psutil.virtual_memory().used/float(1<<30):,.0f} GB" , )
|
80 |
# %%
|
81 |
|
|
|
|
|
|
|
82 |
def run_im(im1, use_trcap_images, text1, use_trcap_texts):
|
|
|
83 |
f_texts_embeddings = None
|
84 |
-
f_image_embeddings = None
|
85 |
-
global image_embeddings
|
86 |
-
global text_embeddings
|
87 |
ims = None
|
88 |
-
print("im2", use_trcap_images)
|
89 |
if use_trcap_images:
|
90 |
-
print('TRCaption images used')
|
91 |
-
# Images taken from TRCAPTION
|
92 |
im_paths = trcap_urls
|
93 |
-
if image_embeddings is None:
|
94 |
-
print(f"First used memory {psutil.virtual_memory().used / float(1 << 30):,.0f} GB", )
|
95 |
-
text_embeddings = None
|
96 |
-
image_embeddings = load_image_embeddings()
|
97 |
-
print(f"First used memory {psutil.virtual_memory().used / float(1 << 30):,.0f} GB", )
|
98 |
-
f_image_embeddings = image_embeddings
|
99 |
else:
|
|
|
100 |
# Images taken from user
|
101 |
im_paths = [i.name for i in im1]
|
102 |
ims = [Image.open(i) for i in im_paths]
|
103 |
if use_trcap_texts:
|
|
|
104 |
random_indexes = random.sample(range(len(trcap_texts)), 2) # MAX 2 text are allowed in image retrieval UI limit
|
105 |
f_texts_embeddings = []
|
106 |
for i in random_indexes:
|
107 |
f_texts_embeddings.append(load_spesific_tensor(i, 'text'))
|
108 |
f_texts_embeddings = torch.stack(f_texts_embeddings)
|
109 |
texts = [trcap_texts[i] for i in random_indexes]
|
|
|
110 |
else:
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
-
|
|
|
114 |
|
115 |
print(f'per_mode_indices = {per_mode_indices}\n,per_mode_probs = {per_mode_probs} ')
|
116 |
print(f'im_paths = {im_paths}')
|
@@ -122,39 +140,45 @@ def run_im(im1, use_trcap_images, text1, use_trcap_texts):
|
|
122 |
|
123 |
|
124 |
def run_text(im1, use_trcap_images, text1, use_trcap_texts):
|
125 |
-
|
126 |
f_image_embeddings = None
|
127 |
-
global image_embeddings
|
128 |
-
global text_embeddings
|
129 |
ims = None
|
130 |
if use_trcap_images:
|
|
|
131 |
random_indexes = random.sample(range(len(trcap_urls)), 2) # MAX 2 text are allowed in image retrieval UI limit
|
132 |
f_image_embeddings = []
|
133 |
for i in random_indexes:
|
134 |
f_image_embeddings.append(load_spesific_tensor(i, 'image'))
|
135 |
f_image_embeddings = torch.stack(f_image_embeddings)
|
136 |
-
print('
|
137 |
# Images taken from TRCAPTION
|
138 |
im_paths = [trcap_urls[i] for i in random_indexes]
|
|
|
|
|
139 |
else:
|
|
|
140 |
# Images taken from user
|
141 |
im_paths = [i.name for i in im1[:2]]
|
142 |
ims = [Image.open(i) for i in im_paths]
|
143 |
|
144 |
if use_trcap_texts:
|
145 |
-
if text_embeddings is None:
|
146 |
-
print(f"Used memory {psutil.virtual_memory().used / float(1 << 30):,.0f} GB", )
|
147 |
-
image_embeddings = None
|
148 |
-
print(f"Image embd deleted used memory {psutil.virtual_memory().used / float(1 << 30):,.0f} GB", )
|
149 |
-
text_embeddings = load_text_embeddings()
|
150 |
-
print(f"Text embed used memory {psutil.virtual_memory().used / float(1 << 30):,.0f} GB", )
|
151 |
-
|
152 |
-
f_texts_embeddings = text_embeddings
|
153 |
texts = trcap_texts
|
154 |
else:
|
155 |
-
texts = [i.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
|
157 |
-
|
|
|
158 |
print(per_mode_indices)
|
159 |
print(per_mode_probs)
|
160 |
return text_retrieval_visualize(per_mode_indices, per_mode_probs, im_paths, texts,
|
@@ -219,7 +243,7 @@ with gr.Blocks() as demo:
|
|
219 |
<rect x="23" y="115" width="23" height="23" fill="#AEAEAE"></rect>
|
220 |
<rect x="23" y="69" width="23" height="23" fill="black"></rect>
|
221 |
</svg>
|
222 |
-
<h1 style="font-weight:
|
223 |
Trclip Demo
|
224 |
<a
|
225 |
href="https://github.com/yusufani/TrCLIP"
|
@@ -234,21 +258,35 @@ with gr.Blocks() as demo:
|
|
234 |
Also you can use pre calculated TrCaption embeddings.
|
235 |
Number of texts = 3533312
|
236 |
Number of images = 3070976
|
|
|
|
|
|
|
237 |
|
238 |
-
|
239 |
</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
240 |
</div>
|
241 |
""")
|
242 |
|
243 |
with gr.Tabs():
|
244 |
with gr.TabItem("Use Own Images"):
|
245 |
im_input = gr.components.File(label="Image input", optional=True, file_count='multiple')
|
246 |
-
is_trcap_ims = gr.Checkbox(label="Use TRCaption Images\
|
247 |
|
248 |
with gr.Tabs():
|
249 |
with gr.TabItem("Input a text (Seperated by new line Max 2 for Image retrieval)"):
|
250 |
text_input = gr.components.Textbox(label="Text input", optional=True)
|
251 |
-
is_trcap_texts = gr.Checkbox(label="Use TrCaption Captions \
|
252 |
|
253 |
im_ret_but = gr.Button("Image Retrieval")
|
254 |
text_ret_but = gr.Button("Text Retrieval")
|
|
|
12 |
import pickle
|
13 |
import random
|
14 |
|
15 |
+
import numpy as np
|
16 |
+
|
17 |
# %%
|
18 |
model_name = 'trclip-vitl14-e10'
|
19 |
if not os.path.exists(model_name):
|
|
|
30 |
|
31 |
# %%
|
32 |
|
33 |
+
def load_image_embeddings(load_batch=True):
|
34 |
path = os.path.join('TrCaption-trclip-vitl14-e10', 'image_embeddings')
|
35 |
bs = 100_000
|
36 |
+
if load_batch:
|
37 |
+
for i in tqdm(range(0, 3_100_000, bs), desc='Loading TrCaption Image embeddings'):
|
38 |
+
with open(os.path.join(path, f'image_em_{i}.pkl'), 'rb') as f:
|
39 |
+
yield pickle.load(f)
|
40 |
+
return
|
41 |
+
|
42 |
+
else:
|
43 |
+
embeddings = []
|
44 |
+
for i in tqdm(range(0, 3_100_000, bs), desc='Loading TrCaption Image embeddings'):
|
45 |
+
with open(os.path.join(path, f'image_em_{i}.pkl'), 'rb') as f:
|
46 |
+
embeddings.append(pickle.load(f))
|
47 |
+
return torch.cat(embeddings, dim=0)
|
48 |
|
|
|
|
|
|
|
|
|
49 |
|
50 |
+
def load_text_embeddings(load_batch=True):
|
51 |
path = os.path.join('TrCaption-trclip-vitl14-e10', 'text_embeddings')
|
52 |
bs = 100_000
|
53 |
+
if load_batch:
|
54 |
+
for i in tqdm(range(0, 3_600_000, bs), desc='Loading TrCaption text embeddings'):
|
55 |
+
with open(os.path.join(path, f'text_em_{i}.pkl'), 'rb') as f:
|
56 |
+
yield pickle.load(f)
|
57 |
+
return
|
58 |
+
else:
|
59 |
+
embeddings = []
|
60 |
+
for i in tqdm(range(0, 3_600_000, bs), desc='Loading TrCaption text embeddings'):
|
61 |
+
with open(os.path.join(path, f'text_em_{i}.pkl'), 'rb') as f:
|
62 |
+
embeddings.append(pickle.load(f))
|
63 |
+
return torch.cat(embeddings, dim=0)
|
64 |
|
65 |
|
66 |
def load_metadata():
|
|
|
71 |
trcap_urls = metadata['image_urls']
|
72 |
return trcap_texts, trcap_urls
|
73 |
|
74 |
+
|
75 |
+
def load_spesific_tensor(index, type, bs=100_000):
|
76 |
part = index // bs
|
77 |
idx = index % bs
|
78 |
+
with open(os.path.join('TrCaption-trclip-vitl14-e10', f'{type}_embeddings', f'{type}_em_{part * bs}.pkl'), 'rb') as f:
|
79 |
embeddings = pickle.load(f)
|
80 |
return embeddings[idx]
|
81 |
|
|
|
82 |
|
83 |
+
# %%
|
|
|
|
|
|
|
84 |
trcap_texts, trcap_urls = load_metadata()
|
85 |
# %%
|
86 |
+
print(f'INFO : Model loading')
|
87 |
model_path = os.path.join(model_name, 'pytorch_model.bin')
|
88 |
trclip = Trclip(model_path, clip_model='ViT-L/14', device='cpu')
|
|
|
|
|
|
|
|
|
89 |
# %%
|
90 |
|
91 |
+
|
92 |
+
|
93 |
+
# %%
|
94 |
def run_im(im1, use_trcap_images, text1, use_trcap_texts):
|
95 |
+
print(f'INFO : Image retrieval starting')
|
96 |
f_texts_embeddings = None
|
|
|
|
|
|
|
97 |
ims = None
|
|
|
98 |
if use_trcap_images:
|
99 |
+
print('INFO : TRCaption images used')
|
|
|
100 |
im_paths = trcap_urls
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
else:
|
102 |
+
print('INFO : Own images used')
|
103 |
# Images taken from user
|
104 |
im_paths = [i.name for i in im1]
|
105 |
ims = [Image.open(i) for i in im_paths]
|
106 |
if use_trcap_texts:
|
107 |
+
print(f'INFO : TRCaption texts used')
|
108 |
random_indexes = random.sample(range(len(trcap_texts)), 2) # MAX 2 text are allowed in image retrieval UI limit
|
109 |
f_texts_embeddings = []
|
110 |
for i in random_indexes:
|
111 |
f_texts_embeddings.append(load_spesific_tensor(i, 'text'))
|
112 |
f_texts_embeddings = torch.stack(f_texts_embeddings)
|
113 |
texts = [trcap_texts[i] for i in random_indexes]
|
114 |
+
|
115 |
else:
|
116 |
+
print(f'INFO : Own texts used')
|
117 |
+
texts = [i.strip() for i in text1.split('\n')[:2] if i.strip() != '']
|
118 |
+
|
119 |
+
if use_trcap_images: # This means that we will iterate over batches because Huggingface space has 16 gb limit :///
|
120 |
+
per_mode_probs = []
|
121 |
+
f_texts_embeddings = f_texts_embeddings if use_trcap_texts else trclip.get_text_features(texts)
|
122 |
+
for f_image_embeddings in tqdm(load_image_embeddings(load_batch=True), desc='Running image retrieval'):
|
123 |
+
batch_probs = trclip.get_results(
|
124 |
+
text_features=f_texts_embeddings, image_features=f_image_embeddings, mode='per_text', return_probs=True)
|
125 |
+
per_mode_probs.append(batch_probs)
|
126 |
+
per_mode_probs = torch.cat(per_mode_probs, dim=1)
|
127 |
+
per_mode_probs = per_mode_probs.softmax(dim=-1).cpu().detach().numpy()
|
128 |
+
per_mode_indices = [np.argsort(prob)[::-1] for prob in per_mode_probs]
|
129 |
|
130 |
+
else:
|
131 |
+
per_mode_indices, per_mode_probs = trclip.get_results(texts=texts, images=ims, text_features=f_texts_embeddings, mode='per_text')
|
132 |
|
133 |
print(f'per_mode_indices = {per_mode_indices}\n,per_mode_probs = {per_mode_probs} ')
|
134 |
print(f'im_paths = {im_paths}')
|
|
|
140 |
|
141 |
|
142 |
def run_text(im1, use_trcap_images, text1, use_trcap_texts):
|
143 |
+
print(f'INFO : Image retrieval starting')
|
144 |
f_image_embeddings = None
|
|
|
|
|
145 |
ims = None
|
146 |
if use_trcap_images:
|
147 |
+
print('INFO : TRCaption images used')
|
148 |
random_indexes = random.sample(range(len(trcap_urls)), 2) # MAX 2 text are allowed in image retrieval UI limit
|
149 |
f_image_embeddings = []
|
150 |
for i in random_indexes:
|
151 |
f_image_embeddings.append(load_spesific_tensor(i, 'image'))
|
152 |
f_image_embeddings = torch.stack(f_image_embeddings)
|
153 |
+
print(f'f_image_embeddings = {f_image_embeddings}')
|
154 |
# Images taken from TRCAPTION
|
155 |
im_paths = [trcap_urls[i] for i in random_indexes]
|
156 |
+
print(f'im_paths = {im_paths}')
|
157 |
+
|
158 |
else:
|
159 |
+
print('INFO : Own images used')
|
160 |
# Images taken from user
|
161 |
im_paths = [i.name for i in im1[:2]]
|
162 |
ims = [Image.open(i) for i in im_paths]
|
163 |
|
164 |
if use_trcap_texts:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
texts = trcap_texts
|
166 |
else:
|
167 |
+
texts = [i.strip() for i in text1.split('\n')[:2] if i.strip() != '']
|
168 |
+
|
169 |
+
if use_trcap_texts:
|
170 |
+
f_image_embeddings = f_image_embeddings if use_trcap_images else trclip.get_image_features(ims)
|
171 |
+
per_mode_probs = []
|
172 |
+
for f_texts_embeddings in tqdm(load_text_embeddings(load_batch=True), desc='Running text retrieval'):
|
173 |
+
batch_probs = trclip.get_results(
|
174 |
+
text_features=f_texts_embeddings, image_features=f_image_embeddings, mode='per_image', return_probs=True)
|
175 |
+
per_mode_probs.append(batch_probs)
|
176 |
+
per_mode_probs = torch.cat(per_mode_probs, dim=1)
|
177 |
+
per_mode_probs = per_mode_probs.softmax(dim=-1).cpu().detach().numpy()
|
178 |
+
per_mode_indices = [np.argsort(prob)[::-1] for prob in per_mode_probs]
|
179 |
|
180 |
+
else:
|
181 |
+
per_mode_indices, per_mode_probs = trclip.get_results(texts=texts, images=ims, image_features=f_image_embeddings, mode='per_image')
|
182 |
print(per_mode_indices)
|
183 |
print(per_mode_probs)
|
184 |
return text_retrieval_visualize(per_mode_indices, per_mode_probs, im_paths, texts,
|
|
|
243 |
<rect x="23" y="115" width="23" height="23" fill="#AEAEAE"></rect>
|
244 |
<rect x="23" y="69" width="23" height="23" fill="black"></rect>
|
245 |
</svg>
|
246 |
+
<h1 style="font-weight: 1500; margin-bottom: 7px;">
|
247 |
Trclip Demo
|
248 |
<a
|
249 |
href="https://github.com/yusufani/TrCLIP"
|
|
|
258 |
Also you can use pre calculated TrCaption embeddings.
|
259 |
Number of texts = 3533312
|
260 |
Number of images = 3070976
|
261 |
+
|
262 |
+
Some images are not available in the internet because I downloaded and calculated TrCaption embeddings long time ago. Don't be suprise if you encounter with Image not found :D
|
263 |
+
|
264 |
|
265 |
+
|
266 |
</p>
|
267 |
+
<p style="margin-bottom: 10px; font-size: 75%" ><em>Huggingface Space containers has 16 gb ram. TrCaption embeddings are totaly 20 gb. </em><em>I did a lot of writing and reading to files to make this space workable. That's why<span style="background-color: #ff6600; color: #ffffff;"> <strong>it's running much slower if you're using TrCaption Embeddig</strong>s</span>.</em></p>
|
268 |
+
<div class="sc-jSFjdj sc-iCoGMd jcTaHb kMthTr">
|
269 |
+
<div class="sc-iqAclL xfxEN">
|
270 |
+
<div class="sc-bdnxRM fJdnBK sc-crzoAE DykGo">
|
271 |
+
<div class="sc-gtsrHT gfuSqG"> </div>
|
272 |
+
</div>
|
273 |
+
</div>
|
274 |
+
</div>
|
275 |
+
<div class="sc-jSFjdj sc-gKAaRy jcTaHb hydYaP">
|
276 |
+
<div class="sc-pNWdM lfZLSv"> </div>
|
277 |
+
</div>
|
278 |
</div>
|
279 |
""")
|
280 |
|
281 |
with gr.Tabs():
|
282 |
with gr.TabItem("Use Own Images"):
|
283 |
im_input = gr.components.File(label="Image input", optional=True, file_count='multiple')
|
284 |
+
is_trcap_ims = gr.Checkbox(label="Use TRCaption Images\n[Note: Random 2 sample selected in text retrieval mode )]")
|
285 |
|
286 |
with gr.Tabs():
|
287 |
with gr.TabItem("Input a text (Seperated by new line Max 2 for Image retrieval)"):
|
288 |
text_input = gr.components.Textbox(label="Text input", optional=True)
|
289 |
+
is_trcap_texts = gr.Checkbox(label="Use TrCaption Captions \n[Note: Random 2 sample selected in image retrieval mode]")
|
290 |
|
291 |
im_ret_but = gr.Button("Image Retrieval")
|
292 |
text_ret_but = gr.Button("Text Retrieval")
|