Update app.py

app.py

@@ -9,23 +9,11 @@ torch.hub.download_url_to_file('http://images.cocodataset.org/val2017/0000000397
 torch.hub.download_url_to_file('https://huggingface.co/datasets/nielsr/textcaps-sample/resolve/main/stop_sign.png', 'stop_sign.png')
 torch.hub.download_url_to_file('https://cdn.openai.com/dall-e-2/demos/text2im/astronaut/horse/photo/0.jpg', 'astronaut.jpg')
 
-# git_processor_base = AutoProcessor.from_pretrained("microsoft/git-base-coco")
-# git_model_base = AutoModelForCausalLM.from_pretrained("microsoft/git-base-coco")
-
 git_processor_large_coco = AutoProcessor.from_pretrained("microsoft/git-large-coco")
-git_model_large_coco = AutoModelForCausalLM.from_pretrained("microsoft/git-large-coco")
-
-# git_processor_large_textcaps = AutoProcessor.from_pretrained("microsoft/git-large-r-textcaps")
-# git_model_large_textcaps = AutoModelForCausalLM.from_pretrained("microsoft/git-large-r-textcaps")
-
-# blip_processor_base = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
-# blip_model_base = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
+git_model_large_coco = AutoModelForCausalLM.from_pretrained("microsoft/git-large-coco", device_map="auto")
 
 blip_processor_large = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
-blip_model_large = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large")
-
-# blip2_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")
-# blip2_model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16)
+blip_model_large = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large", device_map="auto")
 
 blip2_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-6.7b")
 blip2_model_4_bit = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-6.7b", device_map="auto", load_in_4bit=True, torch_dtype=torch.float16)
@@ -33,33 +21,15 @@ blip2_model_4_bit = Blip2ForConditionalGeneration.from_pretrained("Salesforce/bl
 instructblip_processor = AutoProcessor.from_pretrained("Salesforce/instructblip-vicuna-7b")
 instructblip_model_4_bit = InstructBlipForConditionalGeneration.from_pretrained("Salesforce/instructblip-vicuna-7b", device_map="auto", load_in_4bit=True, torch_dtype=torch.float16)
 
-# vitgpt_processor = AutoImageProcessor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
-# vitgpt_model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
-# vitgpt_tokenizer = AutoTokenizer.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
-
-# coca_model, _, coca_transform = open_clip.create_model_and_transforms(
-#  model_name="coca_ViT-L-14",
-#  pretrained="mscoco_finetuned_laion2B-s13B-b90k"
-# )
-
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# git_model_base.to(device)
-# blip_model_base.to(device)
-git_model_large_coco.to(device)
-# git_model_large_textcaps.to(device)
-blip_model_large.to(device)
-# vitgpt_model.to(device)
-# coca_model.to(device)
-# blip2_model.to(device)
-
 def generate_caption(processor, model, image, tokenizer=None, use_float_16=False):
     inputs = processor(images=image, return_tensors="pt").to(device)
 
     if use_float_16:
         inputs = inputs.to(torch.float16)
     
-    generated_ids = model.generate(pixel_values=inputs.pixel_values, max_length=50)
+    generated_ids = model.generate(pixel_values=inputs.pixel_values, num_beams=3, max_length=20, min_length=5) 
 
     if tokenizer is not None:
         generated_caption = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
@@ -69,45 +39,28 @@ def generate_caption(processor, model, image, tokenizer=None, use_float_16=False
     return generated_caption
 
 
-def generate_caption_coca(model, transform, image):
-    im = transform(image).unsqueeze(0).to(device)
-    with torch.no_grad(), torch.cuda.amp.autocast():
-        generated = model.generate(im, seq_len=20)
-    return open_clip.decode(generated[0].detach()).split("<end_of_text>")[0].replace("<start_of_text>", "")
-
-
-def generate_caption_instructblip(processor, model, image):
+def generate_caption_blip2(processor, model, image, replace_token=False):
     prompt = "Generate a caption for the image:"
     
     inputs = processor(images=image, text=prompt, return_tensors="pt").to(device=device, torch_dtype=torch.float16)
 
     generated_ids = model.generate(pixel_values=inputs.pixel_values,
-                                   num_beams=5, max_length=256, min_length=1, top_p=0.9, repetition_penalty=1.5, length_penalty=1.0, temperature=1)
-    generated_ids[generated_ids == 0] = 2 # TODO remove once https://github.com/huggingface/transformers/pull/24492 is merged
+                                   num_beams=5, max_length=50, min_length=1, top_p=0.9, repetition_penalty=1.5, length_penalty=1.0, temperature=1)
+    if replace_token:
+        # TODO remove once https://github.com/huggingface/transformers/pull/24492 is merged
+        generated_ids[generated_ids == 0] = 2 
     
     return processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
 
 
 def generate_captions(image):
-    # caption_git_base = generate_caption(git_processor_base, git_model_base, image)
-
     caption_git_large_coco = generate_caption(git_processor_large_coco, git_model_large_coco, image)
 
-    # caption_git_large_textcaps = generate_caption(git_processor_large_textcaps, git_model_large_textcaps, image)
-
-    # caption_blip_base = generate_caption(blip_processor_base, blip_model_base, image)
-
     caption_blip_large = generate_caption(blip_processor_large, blip_model_large, image)
 
-    # caption_vitgpt = generate_caption(vitgpt_processor, vitgpt_model, image, vitgpt_tokenizer)
-
-    # caption_coca = generate_caption_coca(coca_model, coca_transform, image)
-
-    # caption_blip2 = generate_caption(blip2_processor, blip2_model, image, use_float_16=True).strip()
-
-    caption_blip2_8_bit = generate_caption(blip2_processor, blip2_model_8_bit, image, use_float_16=True).strip()
+    caption_blip2_8_bit = generate_caption_blip2(blip2_processor, blip2_model_8_bit, image).strip()
 
-    caption_instructblip_4_bit = generate_caption_instructblip(instructblip_processor, instructblip_model_4_bit, image)
+    caption_instructblip_4_bit = generate_caption_blip2(instructblip_processor, instructblip_model_4_bit, image)
 
     return caption_git_large_coco, caption_blip_large, caption_blip2_8_bit, caption_instructblip_4_bit