tune some default params

app.py

@@ -159,7 +159,7 @@ def retrieve_knowledge(image):
 
 
 @torch.inference_mode()
-def generate(state: Conversation, temperature, top_p, max_new_tokens, add_knwl, do_sampling):
+def generate(state: Conversation, temperature, top_p, max_new_tokens):
     if state.skip_next:  # This generate call is skipped due to invalid inputs
         yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 3 + knwl_unchange
         return
@@ -172,37 +172,33 @@ def generate(state: Conversation, temperature, top_p, max_new_tokens, add_knwl,
 
     # retrieve and visualize knowledge
     image = state.get_images(return_pil=True)[0]
-    if bool(add_knwl):
-        knwl_embd, knwl = retrieve_knowledge(image)
-        knwl_img, knwl_txt, idx = [None, ] * 15, ["", ] * 15, 0
-        for query_type, knwl_pos in (("whole", 1), ("five", 5), ("nine", 9)):
-            if query_type == "whole":
-                images = [image, ]
-            elif query_type == "five":
-                images = five_crop(image)
-            elif query_type == "nine":
-                images = nine_crop(image)
-            
-            for pos in range(knwl_pos):
-                try:
-                    txt = ""
-                    for k, v in knwl[query_type][pos].items():
-                        v = ", ".join([vi.replace("_", " ") for vi in v])
-                        txt += f"**[{k.upper()}]:** {v}\n\n"
-                    knwl_txt[idx] += txt
-
-                    img = images[pos]
-                    img = query_trans.transforms[0](img)
-                    img = query_trans.transforms[1](img)
-                    img = query_trans.transforms[2](img)
-                    knwl_img[idx] = img
-                except KeyError:
-                    pass
-                idx += 1
-        knwl_vis = tuple(knwl_img + knwl_txt)
-    else:
-        knwl_embd = None
-        knwl_vis = knwl_none
+    knwl_embd, knwl = retrieve_knowledge(image)
+    knwl_img, knwl_txt, idx = [None, ] * 15, ["", ] * 15, 0
+    for query_type, knwl_pos in (("whole", 1), ("five", 5), ("nine", 9)):
+        if query_type == "whole":
+            images = [image, ]
+        elif query_type == "five":
+            images = five_crop(image)
+        elif query_type == "nine":
+            images = nine_crop(image)
+        
+        for pos in range(knwl_pos):
+            try:
+                txt = ""
+                for k, v in knwl[query_type][pos].items():
+                    v = ", ".join([vi.replace("_", " ") for vi in v])
+                    txt += f"**[{k.upper()}]:** {v}\n\n"
+                knwl_txt[idx] += txt
+
+                img = images[pos]
+                img = query_trans.transforms[0](img)
+                img = query_trans.transforms[1](img)
+                img = query_trans.transforms[2](img)
+                knwl_img[idx] = img
+            except KeyError:
+                pass
+            idx += 1
+    knwl_vis = tuple(knwl_img + knwl_txt)
     yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 3 + knwl_vis
 
     # generate output
@@ -217,7 +213,7 @@ def generate(state: Conversation, temperature, top_p, max_new_tokens, add_knwl,
         target=gptk_model.generate,
         kwargs=dict(
             samples=samples,
-            use_nucleus_sampling=bool(do_sampling),
+            use_nucleus_sampling=(temperature > 0.001),
             max_length=min(int(max_new_tokens), 1024),
             top_p=float(top_p),
             temperature=float(temperature),
@@ -270,7 +266,6 @@ def build_demo():
                 gr.Examples(examples=[
                     ["examples/mona_lisa.jpg", "Discuss the historical impact and the significance of this painting in the art world."],
                     ["examples/mona_lisa_dog.jpg", "Describe this photo in detail."],
-                    ["examples/diamond_head.jpg", "What is the name of this famous sight in the photo?"],
                     ["examples/horseshoe_bend.jpg", "What are the possible reasons of the formation of this sight?"],
                 ], inputs=[imagebox, textbox])
 
@@ -286,10 +281,7 @@ def build_demo():
                     clear_btn = gr.Button(value="🗑️ Clear", interactive=False, scale=1)
 
                 with gr.Accordion("Parameters", open=True):
-                    with gr.Row():
-                        add_knwl = gr.Checkbox(value=True, interactive=True, label="Knowledge")
-                        do_sampling = gr.Checkbox(value=False, interactive=True, label="Sampling")
-                    temperature = gr.Slider(minimum=0.1, maximum=2.0, value=1.0, step=0.1, interactive=True, label="Temperature",)
+                    temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True, label="Temperature",)
                     top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0.7, step=0.1, interactive=True, label="Top P",)
                     max_output_tokens = gr.Slider(minimum=0, maximum=1024, value=512, step=64, interactive=True, label="Max output tokens",)
 
@@ -318,7 +310,7 @@ def build_demo():
             regenerate, [state], [state, chatbot, textbox, imagebox] + btn_list
         ).then(
             generate,
-            [state, temperature, top_p, max_output_tokens, add_knwl, do_sampling],
+            [state, temperature, top_p, max_output_tokens],
             [state, chatbot] + btn_list + knwl_vis
         )
 
@@ -330,7 +322,7 @@ def build_demo():
             add_text, [state, textbox, imagebox], [state, chatbot, textbox, imagebox] + btn_list
         ).then(
             generate,
-            [state, temperature, top_p, max_output_tokens, add_knwl, do_sampling],
+            [state, temperature, top_p, max_output_tokens],
             [state, chatbot] + btn_list + knwl_vis
         )
 
@@ -338,7 +330,7 @@ def build_demo():
             add_text, [state, textbox, imagebox], [state, chatbot, textbox, imagebox] + btn_list
         ).then(
             generate,
-            [state, temperature, top_p, max_output_tokens, add_knwl, do_sampling],
+            [state, temperature, top_p, max_output_tokens],
             [state, chatbot] + btn_list + knwl_vis
         )