Update app.py
Browse files
app.py
CHANGED
|
@@ -1,104 +1,59 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
-
import
|
| 3 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
| 4 |
from PIL import Image
|
| 5 |
|
| 6 |
-
# ---
|
| 7 |
-
|
| 8 |
-
torch_dtype = torch.float16 if device == "cuda" else torch.float32
|
| 9 |
-
MODEL_NAME = "microsoft/Florence-2-large"
|
| 10 |
|
| 11 |
-
#
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
|
|
|
| 15 |
)
|
| 16 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 17 |
-
MODEL_NAME,
|
| 18 |
-
trust_remote_code=True,
|
| 19 |
-
torch_dtype=torch_dtype
|
| 20 |
-
).to(device)
|
| 21 |
|
| 22 |
-
# ---
|
| 23 |
def generate_alt_text(files, max_new_tokens):
|
| 24 |
captions = []
|
| 25 |
for f in files or []:
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
return_tensors="pt"
|
| 35 |
-
).to(device)
|
| 36 |
-
# Generate tokens
|
| 37 |
-
generated_ids = model.generate(
|
| 38 |
-
input_ids=inputs["input_ids"],
|
| 39 |
-
pixel_values=inputs["pixel_values"],
|
| 40 |
-
max_new_tokens=int(max_new_tokens),
|
| 41 |
-
num_beams=3,
|
| 42 |
-
do_sample=False
|
| 43 |
)
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
generated_ids,
|
| 47 |
-
skip_special_tokens=True
|
| 48 |
-
)[0]
|
| 49 |
-
# Post-process for caption task
|
| 50 |
-
result = processor.post_process_generation(
|
| 51 |
-
generated_text,
|
| 52 |
-
task=prompt,
|
| 53 |
-
image_size=(image.width, image.height)
|
| 54 |
-
)
|
| 55 |
-
caption = result.get(prompt, "").strip().capitalize()
|
| 56 |
-
captions.append(caption)
|
| 57 |
return "\n\n".join(captions)
|
| 58 |
|
| 59 |
-
# --- Gradio UI ---
|
| 60 |
-
with gr.Blocks(title="
|
| 61 |
-
gr.Markdown("#
|
| 62 |
-
gr.Markdown("Using
|
| 63 |
|
| 64 |
with gr.Row():
|
| 65 |
-
file_input = gr.File(
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
)
|
| 70 |
-
max_new_tokens = gr.Slider(
|
| 71 |
-
minimum=16,
|
| 72 |
-
maximum=512,
|
| 73 |
-
step=16,
|
| 74 |
-
value=128,
|
| 75 |
-
label="Max New Tokens"
|
| 76 |
-
)
|
| 77 |
-
generate_btn = gr.Button("Generate Alt Text")
|
| 78 |
-
clear_btn = gr.Button("Clear All")
|
| 79 |
|
| 80 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 81 |
-
output
|
| 82 |
|
| 83 |
-
#
|
| 84 |
file_input.change(
|
| 85 |
fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
|
| 86 |
inputs=file_input,
|
| 87 |
outputs=gallery
|
| 88 |
)
|
| 89 |
|
| 90 |
-
#
|
| 91 |
-
|
| 92 |
-
fn=generate_alt_text,
|
| 93 |
-
inputs=[file_input, max_new_tokens],
|
| 94 |
-
outputs=output,
|
| 95 |
-
queue=True
|
| 96 |
-
)
|
| 97 |
|
| 98 |
-
#
|
| 99 |
-
clear_btn.click(
|
| 100 |
-
fn=lambda: ([], 128, [], ""),
|
| 101 |
-
outputs=[file_input, max_new_tokens, gallery, output]
|
| 102 |
-
)
|
| 103 |
|
| 104 |
demo.launch()
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
+
from transformers import pipeline
|
|
|
|
| 3 |
from PIL import Image
|
| 4 |
|
| 5 |
+
# --- 1) Pick a lightweight, public model ---
|
| 6 |
+
MODEL_NAME = "nlpconnect/vit-gpt2-image-captioning"
|
|
|
|
|
|
|
| 7 |
|
| 8 |
+
# --- 2) Instantiate the image-to-text pipeline ---
|
| 9 |
+
# We’ll pass all generation arguments inside generate_kwargs
|
| 10 |
+
captioner = pipeline(
|
| 11 |
+
"image-to-text",
|
| 12 |
+
model=MODEL_NAME
|
| 13 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
+
# --- 3) Define your generation function ---
|
| 16 |
def generate_alt_text(files, max_new_tokens):
|
| 17 |
captions = []
|
| 18 |
for f in files or []:
|
| 19 |
+
img = Image.open(f.name).convert("RGB")
|
| 20 |
+
# send everything via generate_kwargs to bypass sanitize checks
|
| 21 |
+
res = captioner(
|
| 22 |
+
img,
|
| 23 |
+
generate_kwargs={
|
| 24 |
+
"max_new_tokens": int(max_new_tokens),
|
| 25 |
+
"num_beams": 4
|
| 26 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
)
|
| 28 |
+
text = res[0]["generated_text"].strip().capitalize()
|
| 29 |
+
captions.append(text)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
return "\n\n".join(captions)
|
| 31 |
|
| 32 |
+
# --- 4) Build the Gradio UI ---
|
| 33 |
+
with gr.Blocks(title="Vit‑GPT2 Image Captioning (Stable)") as demo:
|
| 34 |
+
gr.Markdown("# Vit‑GPT2 Image Captioning")
|
| 35 |
+
gr.Markdown(f"Using **{MODEL_NAME}**, passing params via `generate_kwargs`")
|
| 36 |
|
| 37 |
with gr.Row():
|
| 38 |
+
file_input = gr.File(label="Upload Images", file_count="multiple", file_types=["image"])
|
| 39 |
+
slider = gr.Slider(16, 128, value=64, step=16, label="Max New Tokens")
|
| 40 |
+
go_btn = gr.Button("Generate Alt‑Text")
|
| 41 |
+
clear_btn = gr.Button("Clear All")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 44 |
+
output = gr.Textbox(label="Generated Alt‑Text", lines=10, show_copy_button=True)
|
| 45 |
|
| 46 |
+
# preview images
|
| 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 |
+
# generate captions
|
| 54 |
+
go_btn.click(generate_alt_text, [file_input, slider], output, queue=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
+
# clear everything
|
| 57 |
+
clear_btn.click(lambda: ([], 64, [], ""), outputs=[file_input, slider, gallery, output])
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
demo.launch()
|