Update app.py
Browse files
app.py
CHANGED
|
@@ -1,38 +1,39 @@
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import torch
|
| 3 |
from PIL import Image
|
| 4 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
| 5 |
|
| 6 |
# --- Device & Model Setup ---
|
| 7 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 8 |
-
torch_dtype = torch.float16 if
|
| 9 |
|
| 10 |
-
# Load and patch config to disable SDPA support checks
|
| 11 |
-
config = AutoConfig.from_pretrained(
|
| 12 |
-
"microsoft/Florence-2-large",
|
| 13 |
-
trust_remote_code=True
|
| 14 |
-
)
|
| 15 |
-
setattr(config, '_supports_sdpa', False)
|
| 16 |
-
|
| 17 |
-
# Load the patched model and processor
|
| 18 |
model = AutoModelForCausalLM.from_pretrained(
|
| 19 |
"microsoft/Florence-2-large",
|
| 20 |
-
config=config,
|
| 21 |
torch_dtype=torch_dtype,
|
| 22 |
trust_remote_code=True
|
| 23 |
).to(device)
|
|
|
|
| 24 |
processor = AutoProcessor.from_pretrained(
|
| 25 |
"microsoft/Florence-2-large",
|
| 26 |
trust_remote_code=True
|
| 27 |
)
|
| 28 |
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
def generate_alt_text(files, max_new_tokens):
|
| 31 |
captions = []
|
| 32 |
for f in files or []:
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
inputs = processor(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
generated_ids = model.generate(
|
| 37 |
input_ids=inputs["input_ids"],
|
| 38 |
pixel_values=inputs["pixel_values"],
|
|
@@ -40,35 +41,55 @@ def generate_alt_text(files, max_new_tokens):
|
|
| 40 |
num_beams=3,
|
| 41 |
do_sample=False
|
| 42 |
)
|
| 43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
parsed = processor.post_process_generation(
|
| 45 |
generated_text,
|
| 46 |
-
task=
|
| 47 |
-
image_size=(
|
| 48 |
)
|
| 49 |
-
caption = parsed.get(
|
| 50 |
captions.append(caption)
|
|
|
|
| 51 |
return "\n\n".join(captions)
|
| 52 |
|
|
|
|
| 53 |
# --- Gradio Interface ---
|
| 54 |
with gr.Blocks(title="Florence‑2‑large Alt‑Text Generator") as demo:
|
| 55 |
gr.Markdown("# Florence‑2‑large Alt‑Text Generator")
|
| 56 |
-
gr.Markdown("
|
| 57 |
|
| 58 |
with gr.Row():
|
| 59 |
-
file_input = gr.File(
|
| 60 |
-
|
| 61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
clear_btn = gr.Button("Clear All")
|
| 63 |
|
| 64 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 65 |
-
output
|
| 66 |
|
|
|
|
| 67 |
file_input.change(
|
| 68 |
fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
|
| 69 |
-
inputs=file_input,
|
|
|
|
| 70 |
)
|
| 71 |
|
|
|
|
| 72 |
run_btn.click(
|
| 73 |
fn=generate_alt_text,
|
| 74 |
inputs=[file_input, slider],
|
|
@@ -76,6 +97,7 @@ with gr.Blocks(title="Florence‑2‑large Alt‑Text Generator") as demo:
|
|
| 76 |
queue=True
|
| 77 |
)
|
| 78 |
|
|
|
|
| 79 |
clear_btn.click(
|
| 80 |
fn=lambda: ([], 512, [], ""),
|
| 81 |
outputs=[file_input, slider, gallery, output]
|
|
|
|
| 1 |
+
# app.py
|
| 2 |
import gradio as gr
|
| 3 |
import torch
|
| 4 |
from PIL import Image
|
| 5 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
| 6 |
|
| 7 |
# --- Device & Model Setup ---
|
| 8 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
| 9 |
+
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
| 10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
model = AutoModelForCausalLM.from_pretrained(
|
| 12 |
"microsoft/Florence-2-large",
|
|
|
|
| 13 |
torch_dtype=torch_dtype,
|
| 14 |
trust_remote_code=True
|
| 15 |
).to(device)
|
| 16 |
+
|
| 17 |
processor = AutoProcessor.from_pretrained(
|
| 18 |
"microsoft/Florence-2-large",
|
| 19 |
trust_remote_code=True
|
| 20 |
)
|
| 21 |
|
| 22 |
+
PROMPT = "<OD>"
|
| 23 |
+
|
| 24 |
+
# --- Alt‑Text Generation Function ---
|
| 25 |
def generate_alt_text(files, max_new_tokens):
|
| 26 |
captions = []
|
| 27 |
for f in files or []:
|
| 28 |
+
# 1) Load and preprocess
|
| 29 |
+
img = Image.open(f.name).convert("RGB")
|
| 30 |
+
inputs = processor(
|
| 31 |
+
text=PROMPT,
|
| 32 |
+
images=img,
|
| 33 |
+
return_tensors="pt"
|
| 34 |
+
).to(device, torch_dtype)
|
| 35 |
+
|
| 36 |
+
# 2) Generate
|
| 37 |
generated_ids = model.generate(
|
| 38 |
input_ids=inputs["input_ids"],
|
| 39 |
pixel_values=inputs["pixel_values"],
|
|
|
|
| 41 |
num_beams=3,
|
| 42 |
do_sample=False
|
| 43 |
)
|
| 44 |
+
|
| 45 |
+
# 3) Decode & post‑process
|
| 46 |
+
generated_text = processor.batch_decode(
|
| 47 |
+
generated_ids,
|
| 48 |
+
skip_special_tokens=False
|
| 49 |
+
)[0]
|
| 50 |
parsed = processor.post_process_generation(
|
| 51 |
generated_text,
|
| 52 |
+
task=PROMPT,
|
| 53 |
+
image_size=(img.width, img.height)
|
| 54 |
)
|
| 55 |
+
caption = parsed.get(PROMPT, "").strip().capitalize()
|
| 56 |
captions.append(caption)
|
| 57 |
+
|
| 58 |
return "\n\n".join(captions)
|
| 59 |
|
| 60 |
+
|
| 61 |
# --- Gradio Interface ---
|
| 62 |
with gr.Blocks(title="Florence‑2‑large Alt‑Text Generator") as demo:
|
| 63 |
gr.Markdown("# Florence‑2‑large Alt‑Text Generator")
|
| 64 |
+
gr.Markdown("Drop/upload images below and get detailed `<OD>` descriptions.")
|
| 65 |
|
| 66 |
with gr.Row():
|
| 67 |
+
file_input = gr.File(
|
| 68 |
+
label="Upload Images",
|
| 69 |
+
file_count="multiple",
|
| 70 |
+
file_types=["image"]
|
| 71 |
+
)
|
| 72 |
+
slider = gr.Slider(
|
| 73 |
+
minimum=64,
|
| 74 |
+
maximum=4096,
|
| 75 |
+
step=64,
|
| 76 |
+
value=512,
|
| 77 |
+
label="Max New Tokens"
|
| 78 |
+
)
|
| 79 |
+
run_btn = gr.Button("Generate Alt‑Text")
|
| 80 |
clear_btn = gr.Button("Clear All")
|
| 81 |
|
| 82 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 83 |
+
output = gr.Textbox(label="Generated Alt‑Text", lines=5, show_copy_button=True)
|
| 84 |
|
| 85 |
+
# Show previews immediately
|
| 86 |
file_input.change(
|
| 87 |
fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
|
| 88 |
+
inputs=file_input,
|
| 89 |
+
outputs=gallery
|
| 90 |
)
|
| 91 |
|
| 92 |
+
# Generate captions on click
|
| 93 |
run_btn.click(
|
| 94 |
fn=generate_alt_text,
|
| 95 |
inputs=[file_input, slider],
|
|
|
|
| 97 |
queue=True
|
| 98 |
)
|
| 99 |
|
| 100 |
+
# Clear everything
|
| 101 |
clear_btn.click(
|
| 102 |
fn=lambda: ([], 512, [], ""),
|
| 103 |
outputs=[file_input, slider, gallery, output]
|