.

app.py

@@ -162,14 +162,14 @@ def predict(
         
         
 # sample a .8 of rated embeddings for some stochasticity, or at least two embeddings.
-def get_coeff(embs, ys):
-    n_to_choose = max(int(len(embs)*.8), 2)
-    indices = random.sample(range(len(embs)), n_to_choose)
+def get_coeff(embs_local, ys):
+    n_to_choose = max(int(len(embs_local)*.8), 2)
+    indices = random.sample(range(len(embs_local)), n_to_choose)
 
     # we may have just encountered a rare multi-threading diffusers issue (https://github.com/huggingface/diffusers/issues/5749);
     # this ends up adding a rating but losing an embedding, it seems.
     # let's take off a rating if so to continue without indexing errors.
-    if len(ys) > len(embs):
+    if len(ys) > len(embs_local):
         print('ys are longer than embs; popping latest rating')
         ys.pop(-1)
     
@@ -186,9 +186,10 @@ def get_coeff(embs, ys):
         if has_0 and has_1:
             break
     
-    feature_embs = np.array(torch.cat([embs[i].to('cpu') for i in indices]).to('cpu'))
+    feature_embs = np.array(torch.cat([embs_local[i].to('cpu') for i in indices]).to('cpu'))
     scaler = preprocessing.StandardScaler().fit(feature_embs)
     feature_embs = scaler.transform(feature_embs)
+    print(len(feature_embs), len(ys))
     
     lin_class = LinearSVC(max_iter=50000, dual='auto', class_weight='balanced').fit(feature_embs, np.array([ys[i] for i in indices]))
     lin_class.coef_ = torch.tensor(lin_class.coef_, dtype=torch.double)
@@ -232,20 +233,19 @@ def next_image(embs, img_embs, ys, calibrate_prompts):
             w = 1.4# if len(embs) % 2 == 0 else 0
             
             prompt= '' if glob_idx % 2 == 0 else rng_prompt
-            
             prompt, _ = generate(prompt, in_embs=im_s)
             print(prompt)
             im_emb = autoencoder.embed(prompt)
             embs.append(im_emb)
             
             learn_emb = get_coeff(img_embs, ys)
-            
             img_emb = w * learn_emb.to(dtype=torch.float16)
             image, img_emb = predict(prompt, im_emb=img_emb)
             img_embs.append(img_emb)
 
             if len(embs) > 20:
                 embs.pop(0)
+                img_embs.pop(0)
                 ys.pop(0)
             return image, embs, img_embs, ys, calibrate_prompts