Workpam commited on
Commit
6b3639a
·
verified ·
1 Parent(s): 36d795f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +58 -39
app.py CHANGED
@@ -1,69 +1,88 @@
1
  # app.py
2
  import gradio as gr
 
 
3
  from PIL import Image
4
- from huggingface_hub import InferenceApi
5
 
6
- # Inference client for the public Florence‑2‑large model
7
- inference = InferenceApi(repo_id="microsoft/Florence-2-large")
 
 
 
 
 
 
8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9
  def generate_alt_text(files, max_new_tokens):
10
  captions = []
11
  for f in files or []:
12
- # 1) Read image bytes
13
- img_bytes = open(f.name, "rb").read()
14
-
15
- # 2) Call the hosted Florence endpoint
16
- # Use `parameters=` and drop `request_format`
17
- result = inference(
18
- inputs=img_bytes,
19
- parameters={
20
- "max_new_tokens": int(max_new_tokens),
21
- "num_beams": 3,
22
- "do_sample": False
23
- }
 
 
24
  )
25
-
26
- # 3) Extract and clean up the generated text
27
- # The API returns either a string or list of dicts
28
- if isinstance(result, list) and "generated_text" in result[0]:
29
- text = result[0]["generated_text"]
30
- else:
31
- text = result if isinstance(result, str) else ""
32
- text = text.strip().capitalize()
33
- captions.append(text)
34
-
35
  return "\n\n".join(captions)
36
 
37
  # --- Gradio UI ---
38
  with gr.Blocks(title="Florence‑2‑large Alt‑Text Generator") as demo:
39
- gr.Markdown("## Florence‑2‑large Alt‑Text Generator")
40
- gr.Markdown("Powered by the Hugging Face Inference API—no tokens or local weights needed.")
41
 
42
- with gr.Row():
43
- file_input = gr.File(label="Upload Images", file_count="multiple", file_types=["image"])
44
- slider = gr.Slider(64, 4096, value=512, step=64, label="Max New Tokens")
45
- gen_btn = gr.Button("Generate Alt‑Text")
46
- clear_btn = gr.Button("Clear All")
47
 
48
- gallery = gr.Gallery(label="Image Preview", columns=4)
49
- output = gr.Textbox(label="Generated Alt‑Text", lines=6, show_copy_button=True)
50
 
51
- # Preview immediately
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 captions
59
  gen_btn.click(
60
  fn=generate_alt_text,
61
  inputs=[file_input, slider],
62
  outputs=output,
63
  queue=True
64
  )
65
-
66
- # Clear all
67
  clear_btn.click(
68
  fn=lambda: ([], 512, [], ""),
69
  outputs=[file_input, slider, gallery, output]
 
1
  # app.py
2
  import gradio as gr
3
+ import torch
4
+ import torch.nn as nn
5
  from PIL import Image
6
+ from transformers import AutoProcessor, AutoModelForCausalLM
7
 
8
+ # --- Monkey‑patch to bypass SDPA check ---
9
+ # Whenever a model asks for `_supports_sdpa`, return False instead of raising.
10
+ _orig_getattr = nn.Module.__getattr__
11
+ def _patched_getattr(self, name):
12
+ if name == "_supports_sdpa":
13
+ return False
14
+ return _orig_getattr(self, name)
15
+ nn.Module.__getattr__ = _patched_getattr
16
 
17
+ # --- Device & Model Setup ---
18
+ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
19
+ torch_dtype = torch.float16 if device.type == "cuda" else torch.float32
20
+
21
+ model = AutoModelForCausalLM.from_pretrained(
22
+ "microsoft/Florence-2-large",
23
+ torch_dtype=torch_dtype,
24
+ trust_remote_code=True
25
+ ).to(device)
26
+
27
+ processor = AutoProcessor.from_pretrained(
28
+ "microsoft/Florence-2-large",
29
+ trust_remote_code=True
30
+ )
31
+
32
+ PROMPT = "<OD>"
33
+
34
+ # --- Alt‑Text Generation Function ---
35
  def generate_alt_text(files, max_new_tokens):
36
  captions = []
37
  for f in files or []:
38
+ img = Image.open(f.name).convert("RGB")
39
+ # Preprocess
40
+ inputs = processor(
41
+ text=PROMPT,
42
+ images=img,
43
+ return_tensors="pt"
44
+ ).to(device, torch_dtype)
45
+ # Generate
46
+ outputs = model.generate(
47
+ input_ids=inputs["input_ids"],
48
+ pixel_values=inputs["pixel_values"],
49
+ max_new_tokens=int(max_new_tokens),
50
+ num_beams=3,
51
+ do_sample=False
52
  )
53
+ # Decode & post-process
54
+ raw = processor.batch_decode(outputs, skip_special_tokens=False)[0]
55
+ parsed = processor.post_process_generation(
56
+ raw, task=PROMPT, image_size=(img.width, img.height)
57
+ )
58
+ cap = parsed.get(PROMPT, "").strip().capitalize()
59
+ captions.append(cap)
 
 
 
60
  return "\n\n".join(captions)
61
 
62
  # --- Gradio UI ---
63
  with gr.Blocks(title="Florence‑2‑large Alt‑Text Generator") as demo:
64
+ gr.Markdown("# Florence‑2‑large Alt‑Text Generator")
65
+ gr.Markdown("Upload images to get `<OD>`‑style descriptions from Florence‑2‑large.")
66
 
67
+ file_input = gr.File(label="Images", file_count="multiple", file_types=["image"])
68
+ slider = gr.Slider(64, 4096, step=64, value=512, label="Max New Tokens")
69
+ gen_btn = gr.Button("Generate Alt‑Text")
70
+ clear_btn = gr.Button("Clear All")
 
71
 
72
+ gallery = gr.Gallery(label="Preview", columns=4)
73
+ output = gr.Textbox(label="Generated Alt‑Text", lines=5, show_copy_button=True)
74
 
 
75
  file_input.change(
76
  fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
77
  inputs=file_input,
78
  outputs=gallery
79
  )
 
 
80
  gen_btn.click(
81
  fn=generate_alt_text,
82
  inputs=[file_input, slider],
83
  outputs=output,
84
  queue=True
85
  )
 
 
86
  clear_btn.click(
87
  fn=lambda: ([], 512, [], ""),
88
  outputs=[file_input, slider, gallery, output]