Spaces:
Sleeping
Sleeping
Commit
Β·
d83af99
1
Parent(s):
07d8d5d
misc
Browse files- app.py +26 -47
- image_init/10o.png +2 -2
- image_init/1o.png +2 -2
- image_init/2o.png +2 -2
- image_init/3o.png +2 -2
- image_init/4o.png +2 -2
- image_init/5o.png +2 -2
- image_init/7o.png +2 -2
- image_init/9o.png +2 -2
- requirements.txt +2 -1
app.py
CHANGED
@@ -3,7 +3,7 @@ import gradio as gr
|
|
3 |
import random
|
4 |
import time
|
5 |
import torch
|
6 |
-
|
7 |
|
8 |
import config
|
9 |
from model import get_model_and_tokenizer
|
@@ -137,8 +137,8 @@ def background_next_image():
|
|
137 |
if len(unrated_from_user) >= 10:
|
138 |
continue
|
139 |
|
140 |
-
if len(rated_rows) <
|
141 |
-
continue
|
142 |
|
143 |
global glob_idx
|
144 |
glob_idx += 1
|
@@ -170,11 +170,13 @@ def background_next_image():
|
|
170 |
|
171 |
def pluck_img(user_id):
|
172 |
rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, None) is not None for i in prevs_df.iterrows()]]
|
|
|
173 |
ems = rated_rows['embeddings'].to_list()
|
174 |
ys = [i[user_id][0] for i in rated_rows['user:rating'].to_list()]
|
175 |
user_emb = get_user_emb(ems, ys)
|
176 |
|
177 |
not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
|
|
|
178 |
while len(not_rated_rows) == 0:
|
179 |
not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
|
180 |
time.sleep(.1)
|
@@ -182,8 +184,10 @@ def pluck_img(user_id):
|
|
182 |
|
183 |
unrated_from_user = not_rated_rows[[i[1]['from_user_id'] == user_id for i in not_rated_rows.iterrows()]]
|
184 |
if len(unrated_from_user) > 0:
|
|
|
185 |
# NOTE the way I've setup pandas here is so gdm horrible. TODO overhaul
|
186 |
-
img = unrated_from_user['paths'].to_list()[
|
|
|
187 |
|
188 |
best_sim = -10000000
|
189 |
for i in not_rated_rows.iterrows():
|
@@ -201,7 +205,7 @@ def next_image(calibrate_prompts, user_id):
|
|
201 |
if len(calibrate_prompts) > 0:
|
202 |
cal_video = calibrate_prompts.pop(0)
|
203 |
image = prevs_df[prevs_df['paths'] == cal_video]['paths'].to_list()[0]
|
204 |
-
return image, calibrate_prompts
|
205 |
else:
|
206 |
image = pluck_img(user_id)
|
207 |
return image, calibrate_prompts
|
@@ -211,9 +215,6 @@ def next_image(calibrate_prompts, user_id):
|
|
211 |
|
212 |
|
213 |
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
def start(_, calibrate_prompts, user_id, request: gr.Request):
|
218 |
user_id = int(str(time.time())[-7:].replace('.', ''))
|
219 |
image, calibrate_prompts = next_image(calibrate_prompts, user_id)
|
@@ -227,19 +228,17 @@ def start(_, calibrate_prompts, user_id, request: gr.Request):
|
|
227 |
image,
|
228 |
calibrate_prompts,
|
229 |
user_id,
|
230 |
-
|
231 |
]
|
232 |
|
233 |
|
234 |
def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
|
235 |
global prevs_df
|
236 |
|
237 |
-
|
238 |
if choice == 'π':
|
239 |
choice = [1, 1]
|
240 |
elif choice == 'Neither (Space)':
|
241 |
-
img, calibrate_prompts
|
242 |
-
return img, calibrate_prompts
|
243 |
elif choice == 'π':
|
244 |
choice = [0, 0]
|
245 |
elif choice == 'π Style':
|
@@ -251,7 +250,6 @@ def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
|
|
251 |
|
252 |
# if we detected NSFW, leave that area of latent space regardless of how they rated chosen.
|
253 |
# TODO skip allowing rating & just continue
|
254 |
-
|
255 |
if img is None:
|
256 |
print('NSFW -- choice is disliked')
|
257 |
choice = [0, 0]
|
@@ -260,8 +258,10 @@ def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
|
|
260 |
# if it's still in the dataframe, add the choice
|
261 |
if len(prevs_df.loc[row_mask, 'user:rating']) > 0:
|
262 |
prevs_df.loc[row_mask, 'user:rating'][0][user_id] = choice
|
263 |
-
print(row_mask, prevs_df.loc[row_mask, 'latest_user_to_rate'], [user_id])
|
264 |
prevs_df.loc[row_mask, 'latest_user_to_rate'] = [user_id]
|
|
|
|
|
|
|
265 |
img, calibrate_prompts = next_image(calibrate_prompts, user_id)
|
266 |
return img, calibrate_prompts
|
267 |
|
@@ -330,19 +330,7 @@ Explore the latent space without text prompts based on your preferences. [rynmur
|
|
330 |
''', elem_id="description")
|
331 |
user_id = gr.State()
|
332 |
# calibration videos -- this is a misnomer now :D
|
333 |
-
calibrate_prompts = [
|
334 |
-
'./5o.png',
|
335 |
-
'./2o.png',
|
336 |
-
'./6o.png',
|
337 |
-
'./7o.png',
|
338 |
-
'./1o.png',
|
339 |
-
'./8o.png',
|
340 |
-
'./3o.png',
|
341 |
-
'./4o.png',
|
342 |
-
'./10o.png',
|
343 |
-
'./9o.png',
|
344 |
-
]
|
345 |
-
calibrate_prompts = gr.State(['image_init/'+c for c in calibrate_prompts])
|
346 |
def l():
|
347 |
return None
|
348 |
|
@@ -424,34 +412,25 @@ def encode_space(x):
|
|
424 |
im_emb = model.prior_pipe.image_encoder(im)["image_embeds"]
|
425 |
return im_emb.detach().to('cpu').to(torch.float32)
|
426 |
|
|
|
|
|
|
|
|
|
|
|
427 |
# prep our calibration videos
|
428 |
-
m_calibrate =
|
429 |
-
|
430 |
-
('
|
431 |
-
('./3o.png', 'describe the scene: memento mori'),
|
432 |
-
('./4o.png', 'describe the scene: a green plate with anespresso'),
|
433 |
-
('./5o.png', '5 '),
|
434 |
-
('./6o.png', '6 '),
|
435 |
-
('./7o.png', '7 '),
|
436 |
-
('./8o.png', '8 '),
|
437 |
-
('./9o.png', '9 '),
|
438 |
-
('./10o.png', '10 '),
|
439 |
-
]
|
440 |
-
m_calibrate = [('image_init/'+c[0], c[1]) for c in m_calibrate]
|
441 |
-
for im, txt in m_calibrate:
|
442 |
-
tmp_df = pd.DataFrame(columns=['paths', 'embeddings', 'ips', 'user:rating', 'text', 'gemb'])
|
443 |
tmp_df['paths'] = [im]
|
444 |
image = Image.open(im).convert('RGB')
|
445 |
im_emb = encode_space(image)
|
446 |
|
447 |
tmp_df['embeddings'] = [im_emb.detach().to('cpu')]
|
448 |
tmp_df['user:rating'] = [{' ': ' '}]
|
449 |
-
tmp_df['text'] = [
|
|
|
|
|
450 |
prevs_df = pd.concat((prevs_df, tmp_df))
|
451 |
|
452 |
glob_idx = 0
|
453 |
demo.launch(share=True,)
|
454 |
-
|
455 |
-
|
456 |
-
# TODO interface is shifted -- auto-resize images to all be the same.
|
457 |
-
|
|
|
3 |
import random
|
4 |
import time
|
5 |
import torch
|
6 |
+
import glob
|
7 |
|
8 |
import config
|
9 |
from model import get_model_and_tokenizer
|
|
|
137 |
if len(unrated_from_user) >= 10:
|
138 |
continue
|
139 |
|
140 |
+
if len(rated_rows) < 4:
|
141 |
+
continue
|
142 |
|
143 |
global glob_idx
|
144 |
glob_idx += 1
|
|
|
170 |
|
171 |
def pluck_img(user_id):
|
172 |
rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, None) is not None for i in prevs_df.iterrows()]]
|
173 |
+
print(rated_rows)
|
174 |
ems = rated_rows['embeddings'].to_list()
|
175 |
ys = [i[user_id][0] for i in rated_rows['user:rating'].to_list()]
|
176 |
user_emb = get_user_emb(ems, ys)
|
177 |
|
178 |
not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
|
179 |
+
print(not_rated_rows)
|
180 |
while len(not_rated_rows) == 0:
|
181 |
not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
|
182 |
time.sleep(.1)
|
|
|
184 |
|
185 |
unrated_from_user = not_rated_rows[[i[1]['from_user_id'] == user_id for i in not_rated_rows.iterrows()]]
|
186 |
if len(unrated_from_user) > 0:
|
187 |
+
print(unrated_from_user)
|
188 |
# NOTE the way I've setup pandas here is so gdm horrible. TODO overhaul
|
189 |
+
img = unrated_from_user['paths'].to_list()[-1]
|
190 |
+
return img
|
191 |
|
192 |
best_sim = -10000000
|
193 |
for i in not_rated_rows.iterrows():
|
|
|
205 |
if len(calibrate_prompts) > 0:
|
206 |
cal_video = calibrate_prompts.pop(0)
|
207 |
image = prevs_df[prevs_df['paths'] == cal_video]['paths'].to_list()[0]
|
208 |
+
return image, calibrate_prompts
|
209 |
else:
|
210 |
image = pluck_img(user_id)
|
211 |
return image, calibrate_prompts
|
|
|
215 |
|
216 |
|
217 |
|
|
|
|
|
|
|
218 |
def start(_, calibrate_prompts, user_id, request: gr.Request):
|
219 |
user_id = int(str(time.time())[-7:].replace('.', ''))
|
220 |
image, calibrate_prompts = next_image(calibrate_prompts, user_id)
|
|
|
228 |
image,
|
229 |
calibrate_prompts,
|
230 |
user_id,
|
|
|
231 |
]
|
232 |
|
233 |
|
234 |
def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
|
235 |
global prevs_df
|
236 |
|
|
|
237 |
if choice == 'π':
|
238 |
choice = [1, 1]
|
239 |
elif choice == 'Neither (Space)':
|
240 |
+
img, calibrate_prompts = next_image(calibrate_prompts, user_id)
|
241 |
+
return img, calibrate_prompts
|
242 |
elif choice == 'π':
|
243 |
choice = [0, 0]
|
244 |
elif choice == 'π Style':
|
|
|
250 |
|
251 |
# if we detected NSFW, leave that area of latent space regardless of how they rated chosen.
|
252 |
# TODO skip allowing rating & just continue
|
|
|
253 |
if img is None:
|
254 |
print('NSFW -- choice is disliked')
|
255 |
choice = [0, 0]
|
|
|
258 |
# if it's still in the dataframe, add the choice
|
259 |
if len(prevs_df.loc[row_mask, 'user:rating']) > 0:
|
260 |
prevs_df.loc[row_mask, 'user:rating'][0][user_id] = choice
|
|
|
261 |
prevs_df.loc[row_mask, 'latest_user_to_rate'] = [user_id]
|
262 |
+
else:
|
263 |
+
print('Image apparently removed', img)
|
264 |
+
breakpoint()
|
265 |
img, calibrate_prompts = next_image(calibrate_prompts, user_id)
|
266 |
return img, calibrate_prompts
|
267 |
|
|
|
330 |
''', elem_id="description")
|
331 |
user_id = gr.State()
|
332 |
# calibration videos -- this is a misnomer now :D
|
333 |
+
calibrate_prompts = gr.State( [l for l in random.sample(glob.glob('image_init/*'), k=8)] )
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
334 |
def l():
|
335 |
return None
|
336 |
|
|
|
412 |
im_emb = model.prior_pipe.image_encoder(im)["image_embeds"]
|
413 |
return im_emb.detach().to('cpu').to(torch.float32)
|
414 |
|
415 |
+
# NOTE:
|
416 |
+
# media is moved into a random tmp folder so we need to parse filenames carefully.
|
417 |
+
# do not have any cases where a file name is the same or could be `in` another filename
|
418 |
+
# you also can't use jpegs lmao
|
419 |
+
|
420 |
# prep our calibration videos
|
421 |
+
m_calibrate = glob.glob('image_init/*')
|
422 |
+
for im in m_calibrate:
|
423 |
+
tmp_df = pd.DataFrame(columns=['paths', 'embeddings', 'ips', 'user:rating', 'text', 'gemb', 'from_user_id'])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
424 |
tmp_df['paths'] = [im]
|
425 |
image = Image.open(im).convert('RGB')
|
426 |
im_emb = encode_space(image)
|
427 |
|
428 |
tmp_df['embeddings'] = [im_emb.detach().to('cpu')]
|
429 |
tmp_df['user:rating'] = [{' ': ' '}]
|
430 |
+
tmp_df['text'] = ['']
|
431 |
+
# tmp_df['from_user_id'] = [0]
|
432 |
+
# tmp_df['latest_user_to_rate'] = [0]
|
433 |
prevs_df = pd.concat((prevs_df, tmp_df))
|
434 |
|
435 |
glob_idx = 0
|
436 |
demo.launch(share=True,)
|
|
|
|
|
|
|
|
image_init/10o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/1o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/2o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/3o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/4o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/5o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/7o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
image_init/9o.png
CHANGED
![]() |
Git LFS Details
|
![]() |
Git LFS Details
|
requirements.txt
CHANGED
@@ -14,4 +14,5 @@ peft
|
|
14 |
imageio
|
15 |
apscheduler
|
16 |
pandas
|
17 |
-
av
|
|
|
|
14 |
imageio
|
15 |
apscheduler
|
16 |
pandas
|
17 |
+
av
|
18 |
+
glob2
|