Update app.py

app.py

@@ -50,8 +50,8 @@ def clip_optimized_latent(text, seed, iterations=25, lr=1e-2):
     params = [torch.nn.Parameter(latent_vector[i], requires_grad=True) for i in range(len(latent_vector))]
     optimizer = Adam(params, lr=lr, betas=(0.9, 0.999))
 
-    with torch.no_grad():
-        text_features = clip_model.encode_text(text_input)
+    #with torch.no_grad():
+    #    text_features = clip_model.encode_text(text_input)
         
     #pbar = tqdm(range(iterations), dynamic_ncols=True)
     
@@ -67,10 +67,11 @@ def clip_optimized_latent(text, seed, iterations=25, lr=1e-2):
         #image = clip_preprocess(Image.fromarray((image_np * 255).astype(np.uint8))).unsqueeze(0).to(device)
         
         # Extract features from the image
-        image_features = clip_model.encode_image(image)
+        #image_features = clip_model.encode_image(image)
 
         # Calculate the loss and backpropagate
-        loss = -torch.cosine_similarity(text_features, image_features).mean()
+        loss = 1 - clip_model(image, text_input)[0] / 100
+        #loss = -torch.cosine_similarity(text_features, image_features).mean()
         loss.backward()
         optimizer.step()
         
@@ -189,6 +190,8 @@ def load_model():
 clip_model, inst = load_model()
 model = inst.model
 
+similarity = 1 - clip_model(image, text)[0] / 100
+
 
 path_to_components = get_or_compute(config, inst)
 comps = np.load(path_to_components)
@@ -227,6 +230,8 @@ if 'seed1' not in st.session_state and 'seed2' not in st.session_state:
     st.session_state['seed2'] = random.randint(1, 1000)
 seed1 = st.sidebar.number_input("Seed 1", value= st.session_state['seed1'])
 seed2 = st.sidebar.number_input("Seed 2", value= st.session_state['seed2'])
+st.session_state['seed1'] = seed1
+st.session_state['seed2'] = seed1
 iters = st.sidebar.number_input("Iterations for CLIP Optimization", value = 50)
 submit_button = st.sidebar.button("Submit")
 content = st.sidebar.slider("Structural Composition", min_value=0.0, max_value=1.0, value=0.5)