Caption images and send to CSV file

.gitignore

@@ -162,4 +162,6 @@ cython_debug/
 # venv folder
 env/
 
-node_modules/
+node_modules/
+
+csv/

app.py

@@ -2,11 +2,16 @@ import gradio as gr
 from transformers import Blip2Processor, Blip2ForConditionalGeneration
 import torch
 import PIL.Image
+from pathlib import Path
+import tempfile
+import os
+import os.path
 
 BLIP_MODEL_ID = "Salesforce/blip2-opt-2.7b"
+CAPTION_CSV_DIR=os.path.join(os.getcwd(), 'csv')
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-if device != 'cuda':
+if device.type != 'cuda':
     print(f"You are using {device}. This is much slower than using "
           "a CUDA-enabled GPU.")
 
@@ -14,30 +19,37 @@ if device != 'cuda':
 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)
 
-
-def caption_images(images: list[PIL.Image.Image], prompt: str) -> dict[str, str]:
+def captions_images_to_csv(images: list[PIL.Image.Image]) -> str:
+    caption_map = caption_images(images)
+    lines = [k + "," + v + "\n" for k, v in caption_map.items()]
+    
+    with tempfile.NamedTemporaryFile('w', dir=CAPTION_CSV_DIR, delete=False, suffix=".csv") as f:    
+        f.writelines(lines)
+        return f.name
+    
+    
+def caption_images(images: list[PIL.Image.Image]) -> 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)
+    generated_ids = blip_model.generate(**inputs, max_new_tokens=20)
     
     results = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)
     
-    return str(dict(zip([image.name for image in images], results)))
+    return dict(zip([Path(image.name).stem for image in images], [result.replace("\n", "") for result in 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()
+        image_caption_output = gr.File()
         
 
-    caption_images_button.click(caption_images, inputs=[images_input, blip_prompt], outputs=image_caption_label)
+    caption_images_button.click(captions_images_to_csv, inputs=[images_input], outputs=image_caption_output)
 
 
 print("Launching Gradio")