Update app.py

app.py

@@ -260,6 +260,7 @@ def captioning(img):
             "Describe this image with a caption to be used for image generation."
     )
     inputsa = processor5(images=img, text=cap_prompt, return_tensors="pt").to('cuda')
+    '''
     generated_ids = model5.generate(
         **inputsa,
         do_sample=False,
@@ -271,6 +272,19 @@ def captioning(img):
         length_penalty=1.0,
         temperature=1,
     )
+    '''
+    generated_ids = model5.generate(
+        **inputsa,
+        do_sample=text_decoding_method == "Nucleus sampling",
+        num_beams=1,
+        max_length=128,
+        min_length=64,
+        top_p=0.9,
+        repetition_penalty=1.0,
+        length_penalty=2.0,
+        temperature=0.5,
+    )
+            
 
     generated_text = processor5.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
     generated_text = generated_text.replace(cap_prompt, "").strip() #Or could try .split(prompt, 1)[-1].strip()
@@ -282,14 +296,14 @@ def captioning(img):
         #with torch.no_grad():
         generated_ids = model5.generate(
             **inputs,
-            do_sample=False,
-            num_beams=5,
+            do_sample=text_decoding_method == "Nucleus sampling",
+            num_beams=1,
             max_length=64,
             #min_length=16,
-            top_p=0.9,
-            repetition_penalty=1.5,
-            length_penalty=1.0,
-            temperature=1,
+            top_p=0.1,
+            repetition_penalty=1.0,
+            length_penalty=2.0,
+            temperature=0.5,
         )
         generated_text = processor5.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
         response_text = generated_text.replace(prompt, "").strip() #Or could try .split(prompt, 1)[-1].strip()
@@ -302,14 +316,14 @@ def captioning(img):
     ).to('cuda')
     generated_ids = model5.generate(
             **inputf,
-            do_sample=False,
-            num_beams=5,
+            do_sample=text_decoding_method == "Nucleus sampling",
+            num_beams=1,
             max_length=96,
             min_length=64,
-            top_p=0.9,
-            repetition_penalty=1.5,
+            top_p=0.1,
+            repetition_penalty=1.0,
             length_penalty=1.0,
-            temperature=1,
+            temperature=0.5,
     )
     generated_texta = processor5.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
     response_text = generated_texta.replace(generated_text, "").strip()