Create app.py

app.py

@@ -0,0 +1,48 @@
+import gradio as gr
+from transformers import Blip2Processor, Blip2ForConditionalGeneration
+import torch
+import PIL.Image
+from huggingface_hub import HfApi
+
+BLIP_MODEL_ID = "Salesforce/blip2-opt-2.7b"
+
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+if device != 'cuda':
+    print(f"You are using {device}. This is much slower than using "
+          "a CUDA-enabled GPU.")
+
+
+blip_processor = Blip2Processor.from_pretrained(BLIP_MODEL_ID)
+blip_model = Blip2ForConditionalGeneration.from_pretrained(BLIP_MODEL_ID, torch_dtype=torch.float16, device_map="auto").to(device)
+
+print("Loading API")
+api = HfApi()
+
+
+def caption_images(images: list[PIL.Image.Image], prompt: str) -> dict[str, str]:
+    image_files = [PIL.Image.open(image.name).convert('RGB') for image in images]
+    inputs = blip_processor(images=image_files, return_tensors='pt').to(device, torch.float16)
+    
+    for image in image_files:
+        image.close()
+    
+    generated_ids = blip_model.generate(**inputs)
+    
+    results = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)
+    
+    return str(dict(zip([image.name for image in images], results)))
+
+
+with gr.Blocks() as demo:
+    with gr.Tab("Image Caption Bot"):
+        images_input = gr.File(file_count="multiple", file_types=["image"])
+        blip_prompt = gr.Textbox("Prompt")
+        caption_images_button = gr.Button("Submit")
+        image_caption_label = gr.Label()
+        
+
+    caption_images_button.click(caption_images, inputs=[images_input, blip_prompt], outputs=image_caption_label)
+
+
+print("Launching Gradio")
+demo.launch()