Fixes to image resolution

app.py

@@ -115,14 +115,13 @@ def generate_for_prompt(input_text, state, ret_scale_factor, num_words, temperat
         elif type(p) == dict:
             # Decide whether to generate or retrieve.
             if p['decision'] is not None and p['decision'][0] == 'gen':
-                image = p['gen'][0][0].resize((224, 224))
+                image = p['gen'][0][0]#.resize((224, 224))
                 filename = save_image_to_local(image)
-                response += f'<img src="./file={filename}" style="display: inline-block;"><p style="font-size: 12px; color: #555;">(Generated)</p>'
+                response += f'<img src="./file={filename}" style="display: inline-block;"><p style="font-size: 12px; color: #555; margin-top: 0;">(Generated)</p>'
             else:
-                image = p['ret'][0][0].resize((224, 224))
+                image = p['ret'][0][0]#.resize((224, 224))
                 filename = save_image_to_local(image)
-                response += f'<img src="./file={filename}" style="display: inline-block;"><p style="font-size: 12px; color: #555;">(Retrieved)</p>'
-
+                response += f'<img src="./file={filename}" style="display: inline-block;"><p style="font-size: 12px; color: #555; margin-top: 0;">(Retrieved)</p>'
 
     chat_history = model_inputs + \
         [' '.join([s for s in model_outputs if type(s) == str]) + '\n']
@@ -180,7 +179,7 @@ with gr.Blocks(css=css) as demo:
                         share_button = gr.Button("🤗 Share to Community (opens new window)", elem_id="share-btn")
 
         with gr.Column(scale=0.3, min_width=400):
-            ret_scale_factor = gr.Slider(minimum=0.0, maximum=3.0, value=1.0, step=0.1, interactive=True,
+            ret_scale_factor = gr.Slider(minimum=0.0, maximum=3.0, value=1.2, step=0.1, interactive=True,
                                          label="Frequency multiplier for returning images (higher means more frequent)")
             # max_ret_images = gr.Number(
             #     minimum=0, maximum=3, value=2, precision=1, interactive=True, label="Max images to return")

gill/models.py

@@ -878,10 +878,10 @@ def load_gill(embeddings_dir: str, model_args_path: str, model_ckpt_path: str, d
   model = GILL(tokenizer, args, path_array=path_array, emb_matrix=emb_matrix,
                load_sd=not debug, num_gen_images=1, decision_model_path=decision_model_path)
   model = model.eval()
-  if not debug:
-    model = model.bfloat16()
-    model = model.cuda()
+  if torch.cuda.is_available():
+    model = model.bfloat16().cuda()
 
+  if not debug:
     # Load pretrained linear mappings and [IMG] embeddings.
     checkpoint = torch.load(model_ckpt_path)
     state_dict = {}