misc

app.py

@@ -3,7 +3,7 @@ import gradio as gr
 import random
 import time
 import torch
-
+import glob
 
 import config
 from model import get_model_and_tokenizer
@@ -137,8 +137,8 @@ def background_next_image():
             if len(unrated_from_user) >= 10:
                 continue
             
-            if len(rated_rows) < 5:
-                continue            
+            if len(rated_rows) < 4:
+                continue
 
             global glob_idx
             glob_idx += 1
@@ -170,11 +170,13 @@ def background_next_image():
     
 def pluck_img(user_id):
     rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, None) is not None for i in prevs_df.iterrows()]]
+    print(rated_rows)
     ems = rated_rows['embeddings'].to_list()
     ys = [i[user_id][0] for i in rated_rows['user:rating'].to_list()]
     user_emb = get_user_emb(ems, ys)
 
     not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
+    print(not_rated_rows)
     while len(not_rated_rows) == 0:
         not_rated_rows = prevs_df[[i[1]['user:rating'].get(user_id, 'gone') == 'gone' for i in prevs_df.iterrows()]]
         time.sleep(.1)
@@ -182,8 +184,10 @@ def pluck_img(user_id):
 
     unrated_from_user = not_rated_rows[[i[1]['from_user_id'] == user_id for i in not_rated_rows.iterrows()]]
     if len(unrated_from_user) > 0:
+        print(unrated_from_user)
         # NOTE the way I've setup pandas here is so gdm horrible. TODO overhaul
-        img = unrated_from_user['paths'].to_list()[0]
+        img = unrated_from_user['paths'].to_list()[-1]
+        return img
 
     best_sim = -10000000
     for i in not_rated_rows.iterrows():
@@ -201,7 +205,7 @@ def next_image(calibrate_prompts, user_id):
         if len(calibrate_prompts) > 0:
             cal_video = calibrate_prompts.pop(0)
             image = prevs_df[prevs_df['paths'] == cal_video]['paths'].to_list()[0]
-            return image, calibrate_prompts, 
+            return image, calibrate_prompts
         else:
             image = pluck_img(user_id)
             return image, calibrate_prompts
@@ -211,9 +215,6 @@ def next_image(calibrate_prompts, user_id):
 
 
 
-
-
-
 def start(_, calibrate_prompts, user_id, request: gr.Request):
     user_id = int(str(time.time())[-7:].replace('.', ''))
     image, calibrate_prompts = next_image(calibrate_prompts, user_id)
@@ -227,19 +228,17 @@ def start(_, calibrate_prompts, user_id, request: gr.Request):
             image,
             calibrate_prompts,
             user_id,
-            
             ]
 
 
 def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
     global prevs_df
     
-    
     if choice == '👍':
         choice = [1, 1]
     elif choice == 'Neither (Space)':
-        img, calibrate_prompts,  = next_image(calibrate_prompts, user_id)
-        return img, calibrate_prompts, 
+        img, calibrate_prompts = next_image(calibrate_prompts, user_id)
+        return img, calibrate_prompts
     elif choice == '👎':
         choice = [0, 0]
     elif choice == '👍 Style':
@@ -251,7 +250,6 @@ def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
     
     # if we detected NSFW, leave that area of latent space regardless of how they rated chosen.
     # TODO skip allowing rating & just continue
-
     if img is None:
         print('NSFW -- choice is disliked')
         choice = [0, 0]
@@ -260,8 +258,10 @@ def choose(img, choice, calibrate_prompts, user_id, request: gr.Request):
     # if it's still in the dataframe, add the choice
     if len(prevs_df.loc[row_mask, 'user:rating']) > 0:
         prevs_df.loc[row_mask, 'user:rating'][0][user_id] = choice
-        print(row_mask, prevs_df.loc[row_mask, 'latest_user_to_rate'], [user_id])
         prevs_df.loc[row_mask, 'latest_user_to_rate'] = [user_id]
+    else:
+        print('Image apparently removed', img)
+        breakpoint()
     img, calibrate_prompts = next_image(calibrate_prompts, user_id)
     return img, calibrate_prompts
 
@@ -330,19 +330,7 @@ Explore the latent space without text prompts based on your preferences. [rynmur
     ''', elem_id="description")
     user_id = gr.State()
     # calibration videos -- this is a misnomer now :D
-    calibrate_prompts = [
-    './5o.png',
-    './2o.png',
-    './6o.png',
-    './7o.png',
-    './1o.png',
-    './8o.png',
-    './3o.png',
-    './4o.png',
-    './10o.png',
-    './9o.png',
-    ]
-    calibrate_prompts = gr.State(['image_init/'+c for c in calibrate_prompts])
+    calibrate_prompts = gr.State( [l for l in random.sample(glob.glob('image_init/*'), k=8)] )
     def l():
         return None
 
@@ -424,34 +412,25 @@ def encode_space(x):
     im_emb = model.prior_pipe.image_encoder(im)["image_embeds"]
     return im_emb.detach().to('cpu').to(torch.float32)
 
+# NOTE:
+# media is moved into a random tmp folder so we need to parse filenames carefully.
+# do not have any cases where a file name is the same or could be `in` another filename
+# you also can't use jpegs lmao
+
 # prep our calibration videos
-m_calibrate = [ # DO NOT NAME THESE PNGs JUST NUMBERS! apparently we assign images by number
-    ('./1o.png', 'describe the scene: omens in the suburbs'),
-    ('./2o.png', 'describe the scene: geometric abstract art of a windmill'),
-    ('./3o.png', 'describe the scene: memento mori'),
-    ('./4o.png', 'describe the scene: a green plate with anespresso'),
-    ('./5o.png', '5 '),
-    ('./6o.png', '6 '),
-    ('./7o.png', '7 '),
-    ('./8o.png', '8 '),
-    ('./9o.png', '9 '),
-    ('./10o.png', '10 '),
-    ]
-m_calibrate = [('image_init/'+c[0], c[1]) for c in m_calibrate]
-for im, txt in m_calibrate:
-    tmp_df = pd.DataFrame(columns=['paths', 'embeddings', 'ips', 'user:rating', 'text', 'gemb'])
+m_calibrate = glob.glob('image_init/*')
+for im in m_calibrate:
+    tmp_df = pd.DataFrame(columns=['paths', 'embeddings', 'ips', 'user:rating', 'text', 'gemb', 'from_user_id'])
     tmp_df['paths'] = [im]
     image = Image.open(im).convert('RGB')
     im_emb = encode_space(image)
     
     tmp_df['embeddings'] = [im_emb.detach().to('cpu')]
     tmp_df['user:rating'] = [{' ': ' '}]
-    tmp_df['text'] = [txt]
+    tmp_df['text'] = ['']
+    # tmp_df['from_user_id'] = [0]
+    # tmp_df['latest_user_to_rate'] = [0]
     prevs_df = pd.concat((prevs_df, tmp_df))
 
 glob_idx = 0
 demo.launch(share=True,)
-
-
-# TODO interface is shifted -- auto-resize images to all be the same.
-

requirements.txt

@@ -14,4 +14,5 @@ peft
 imageio
 apscheduler
 pandas
-av
+av
+glob2