Update app.py

app.py

@@ -1,5 +1,5 @@
 import gradio as gr
-from transformers import AutoProcessor, AutoTokenizer, AutoImageProcessor, AutoModelForCausalLM, BlipForConditionalGeneration, Blip2ForConditionalGeneration, VisionEncoderDecoderModel
+from transformers import AutoProcessor, AutoTokenizer, AutoImageProcessor, AutoModelForCausalLM, BlipForConditionalGeneration, Blip2ForConditionalGeneration, VisionEncoderDecoderModel, InstructBlipForConditionalGeneration
 import torch
 import open_clip
 
@@ -15,7 +15,7 @@ torch.hub.download_url_to_file('https://cdn.openai.com/dall-e-2/demos/text2im/as
 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_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")
@@ -27,17 +27,20 @@ blip_model_large = BlipForConditionalGeneration.from_pretrained("Salesforce/blip
 # blip2_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")
 # blip2_model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16)
 
-blip2_processor_8_bit = AutoProcessor.from_pretrained("Salesforce/blip2-opt-6.7b")
-blip2_model_8_bit = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-6.7b", device_map="auto", load_in_8bit=True)
+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)
+
+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"
-)
+# 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"
 
@@ -47,7 +50,7 @@ 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)
+# coca_model.to(device)
 # blip2_model.to(device)
 
 def generate_caption(processor, model, image, tokenizer=None, use_float_16=False):
@@ -73,6 +76,18 @@ def generate_caption_coca(model, transform, image):
     return open_clip.decode(generated[0].detach()).split("<end_of_text>")[0].replace("<start_of_text>", "")
 
 
+def generate_caption_instructblip(processor, model, image):
+    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
+    
+    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)
 
@@ -86,20 +101,22 @@ def generate_captions(image):
 
     # caption_vitgpt = generate_caption(vitgpt_processor, vitgpt_model, image, vitgpt_tokenizer)
 
-    caption_coca = generate_caption_coca(coca_model, coca_transform, image)
+    # 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_8_bit, blip2_model_8_bit, 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_instructblip_4_bit = generate_caption_instructblip(instructblip_processor, instructblip_model_4_bit, image)
 
-    return caption_git_large_coco, caption_blip_large, caption_coca, caption_blip2_8_bit
+    return caption_git_large_coco, caption_blip_large, caption_blip2_8_bit, caption_instructblip_4_bit
 
    
 examples = [["cats.jpg"], ["stop_sign.png"], ["astronaut.jpg"]]
-outputs = [gr.outputs.Textbox(label="Caption generated by GIT-large fine-tuned on COCO"), gr.outputs.Textbox(label="Caption generated by BLIP-large"), gr.outputs.Textbox(label="Caption generated by CoCa"), gr.outputs.Textbox(label="Caption generated by BLIP-2 OPT 6.7b")] 
+outputs = [gr.outputs.Textbox(label="Caption generated by GIT-large fine-tuned on COCO"), gr.outputs.Textbox(label="Caption generated by BLIP-large"), gr.outputs.Textbox(label="Caption generated by BLIP-2 OPT 6.7b"), gr.outputs.Textbox(label="Caption generated by InstructBLIP"), ] 
 
 title = "Interactive demo: comparing image captioning models"
-description = "Gradio Demo to compare GIT, BLIP, CoCa, and BLIP-2, 4 state-of-the-art vision+language models. To use it, simply upload your image and click 'submit', or click one of the examples to load them. Read more at the links below."
+description = "Gradio Demo to compare GIT, BLIP, BLIP-2 and InstructBLIP, 4 state-of-the-art vision+language models. To use it, simply upload your image and click 'submit', or click one of the examples to load them. Read more at the links below."
 article = "<p style='text-align: center'><a href='https://huggingface.co/docs/transformers/main/model_doc/blip' target='_blank'>BLIP docs</a> | <a href='https://huggingface.co/docs/transformers/main/model_doc/git' target='_blank'>GIT docs</a></p>"
 
 interface = gr.Interface(fn=generate_captions,