lunchbreak add

app.py

@@ -94,8 +94,8 @@ device_map='cuda')
 
 
 
-unet = UNet2DConditionModel.from_pretrained('emilianJR/epiCRealism', subfolder='unet',).to(dtype).to('cpu')
-text_encoder = CLIPTextModel.from_pretrained('emilianJR/epiCRealism', subfolder='text_encoder', 
+unet = UNet2DConditionModel.from_pretrained('rynmurdock/Sea_Claws', subfolder='unet',).to(dtype).to('cpu')
+text_encoder = CLIPTextModel.from_pretrained('rynmurdock/Sea_Claws', subfolder='text_encoder', 
 device_map='cpu').to(dtype)
 
 adapter = MotionAdapter.from_pretrained("wangfuyun/AnimateLCM")
@@ -114,7 +114,7 @@ pipe.fuse_lora()
 
 pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15_vit-G.bin", map_location='cpu')
 # This IP adapter improves outputs substantially.
-pipe.set_ip_adapter_scale(.8)
+pipe.set_ip_adapter_scale(.9)
 pipe.unet.fuse_qkv_projections()
 #pipe.enable_free_init(method="gaussian", use_fast_sampling=True)
 
@@ -156,11 +156,12 @@ def generate(in_im_embs):
 
 def get_user_emb(embs, ys):
     # handle case where every instance of calibration videos is 'Neither' or 'Like' or 'Dislike'
-    if len(list(set(ys))) <= 1:
-        embs.append(.01*torch.randn(1280))
-        embs.append(.01*torch.randn(1280))
-        ys.append(0)
-        ys.append(1)
+    if len(list(ys)) <= 7:
+        aways = [.01*torch.randn(1280) for i in range(3)]
+        embs += aways
+        awal = [0 for i in range(3)]
+        ys += awal
+        print('Fixing only one feedback class available.\n')
     
     indices = list(range(len(embs)))
     # sample only as many negatives as there are positives
@@ -177,9 +178,14 @@ def get_user_emb(embs, ys):
     if len(ys) > len(embs):
         ys.pop(-1)
     
-    feature_embs = np.array(torch.stack([embs[i].squeeze().to('cpu') for i in indices]).to('cpu'))
+    feature_embs = torch.stack([embs[i].squeeze().to('cpu') for i in indices]).to('cpu')
     #scaler = preprocessing.StandardScaler().fit(feature_embs)
     #feature_embs = scaler.transform(feature_embs)
+    
+    
+    if feature_embs.norm() != 0:
+        feature_embs = feature_embs / feature_embs.norm()
+    
     chosen_y = np.array([ys[i] for i in indices])
     
     #lin_class = Ridge(fit_intercept=False).fit(feature_embs, chosen_y)