Workpam commited on
Commit
8970b7b
·
verified ·
1 Parent(s): 524733a

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +23 -28
app.py CHANGED
@@ -3,68 +3,63 @@ from transformers import pipeline
3
  from PIL import Image
4
  import torch
5
 
6
- # --- Setup BLIP-2 Base Model Pipeline ---
7
- # Uses Salesforce/blip2-flan-t5-base for more accurate captions
8
- MODEL_NAME = "Salesforce/blip2-flan-t5-base"
9
- DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
10
-
11
- # Initialize pipeline with trust_remote_code (required for BLIP-2)
12
  captioner = pipeline(
13
  "image-to-text",
14
- model=MODEL_NAME,
15
- torch_dtype=(torch.float16 if DEVICE=="cuda" else torch.float32),
16
  device_map="auto",
17
- trust_remote_code=True
18
  )
19
 
20
  # --- Caption Generation Function ---
21
- def generate_alt_text(files, max_tokens):
22
  captions = []
23
  for f in files or []:
24
  img = Image.open(f.name).convert("RGB")
25
- # Pass generation args via generate_kwargs
26
- result = captioner(
27
  img,
28
  generate_kwargs={
29
- "max_new_tokens": int(max_tokens),
30
- "num_beams": 5,
31
- "no_repeat_ngram_size": 3
32
  }
33
  )
34
- text = result[0]["generated_text"].strip().capitalize()
35
  captions.append(text)
36
  return "\n\n".join(captions)
37
 
38
- # --- Gradio Interface ---
39
- with gr.Blocks(title="BLIP-2 Base Alt Text Generator") as demo:
40
- gr.Markdown("# BLIP-2 (FLAN-T5 Base) Alt Text Generator")
41
- gr.Markdown(f"Model: **{MODEL_NAME}** for improved accuracy")
42
 
43
  with gr.Row():
44
  file_input = gr.File(label="Upload Images", file_count="multiple", file_types=["image"])
45
- max_tokens = gr.Slider(16, 256, value=64, step=16, label="Max New Tokens")
46
- generate_btn = gr.Button("Generate Alt Text")
47
- clear_btn = gr.Button("Clear All")
48
 
49
  gallery = gr.Gallery(label="Image Preview", columns=4)
50
- output = gr.Textbox(label="Generated Alt Texts", lines=10, show_copy_button=True)
51
 
52
  file_input.change(
53
  fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
54
  inputs=file_input,
55
- outputs=gallery
56
  )
57
 
58
- generate_btn.click(
59
  fn=generate_alt_text,
60
- inputs=[file_input, max_tokens],
61
  outputs=output,
62
  queue=True
63
  )
64
 
65
  clear_btn.click(
66
  fn=lambda: ([], 64, [], ""),
67
- outputs=[file_input, max_tokens, gallery, output]
68
  )
69
 
70
  demo.launch()
 
3
  from PIL import Image
4
  import torch
5
 
6
+ # --- Device & Model Setup ---
7
+ device = "cuda" if torch.cuda.is_available() else "cpu"
 
 
 
 
8
  captioner = pipeline(
9
  "image-to-text",
10
+ model="Salesforce/blip2-flan-t5-xl",
11
+ torch_dtype=torch.float16 if device=="cuda" else torch.float32,
12
  device_map="auto",
13
+ trust_remote_code=True, # required for BLIP-2
14
  )
15
 
16
  # --- Caption Generation Function ---
17
+ def generate_alt_text(files, max_new_tokens):
18
  captions = []
19
  for f in files or []:
20
  img = Image.open(f.name).convert("RGB")
21
+ # Funnel all generation args through generate_kwargs
22
+ outputs = captioner(
23
  img,
24
  generate_kwargs={
25
+ "max_new_tokens": int(max_new_tokens),
26
+ "num_beams": 4
 
27
  }
28
  )
29
+ text = outputs[0]["generated_text"].strip().capitalize()
30
  captions.append(text)
31
  return "\n\n".join(captions)
32
 
33
+ # --- Gradio App ---
34
+ with gr.Blocks(title="BLIP‑2 XL Alt‑Text Generator") as demo:
35
+ gr.Markdown("# BLIP‑2 XL Alt‑Text Generator")
36
+ gr.Markdown("Model: **Salesforce/blip2‑flan‑t5‑xl**\n\nUses `generate_kwargs` to set `max_new_tokens` & `num_beams`.")
37
 
38
  with gr.Row():
39
  file_input = gr.File(label="Upload Images", file_count="multiple", file_types=["image"])
40
+ slider = gr.Slider(16, 256, value=64, step=16, label="Max New Tokens")
41
+ run_btn = gr.Button("Generate Alt‑Text")
42
+ clear_btn = gr.Button("Clear All")
43
 
44
  gallery = gr.Gallery(label="Image Preview", columns=4)
45
+ output = gr.Textbox(label="Generated Alt‑Text", lines=10, show_copy_button=True)
46
 
47
  file_input.change(
48
  fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
49
  inputs=file_input,
50
+ outputs=gallery,
51
  )
52
 
53
+ run_btn.click(
54
  fn=generate_alt_text,
55
+ inputs=[file_input, slider],
56
  outputs=output,
57
  queue=True
58
  )
59
 
60
  clear_btn.click(
61
  fn=lambda: ([], 64, [], ""),
62
+ outputs=[file_input, slider, gallery, output]
63
  )
64
 
65
  demo.launch()