Update app.py
Browse files
app.py
CHANGED
|
@@ -1,63 +1,153 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
-
from transformers import pipeline
|
| 3 |
-
from PIL import Image
|
| 4 |
import torch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
trust_remote_code=True,
|
| 12 |
-
|
| 13 |
-
torch_dtype=torch.float16 if device == "cuda" else torch.float32
|
| 14 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
def generate_alt_text(files, max_new_tokens):
|
| 18 |
captions = []
|
| 19 |
for f in files or []:
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
)
|
| 28 |
-
|
| 29 |
-
|
|
|
|
|
|
|
|
|
|
| 30 |
return "\n\n".join(captions)
|
| 31 |
|
| 32 |
# --- Gradio UI ---
|
| 33 |
-
with gr.Blocks(title="
|
| 34 |
-
gr.Markdown("#
|
| 35 |
-
gr.Markdown("Model
|
| 36 |
|
| 37 |
with gr.Row():
|
| 38 |
-
|
| 39 |
-
slider
|
| 40 |
-
run_btn
|
| 41 |
-
clear_btn
|
| 42 |
|
| 43 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 44 |
-
output = gr.Textbox(label="Generated Alt
|
| 45 |
|
| 46 |
-
|
| 47 |
-
fn=lambda
|
| 48 |
-
inputs=
|
| 49 |
)
|
| 50 |
-
|
| 51 |
run_btn.click(
|
| 52 |
fn=generate_alt_text,
|
| 53 |
-
inputs=[
|
| 54 |
outputs=output,
|
| 55 |
queue=True
|
| 56 |
)
|
| 57 |
-
|
| 58 |
clear_btn.click(
|
| 59 |
-
fn=lambda: ([],
|
| 60 |
-
outputs=[
|
| 61 |
)
|
| 62 |
|
| 63 |
demo.launch()
|
|
|
|
| 1 |
import gradio as gr
|
|
|
|
|
|
|
| 2 |
import torch
|
| 3 |
+
from PIL import Image
|
| 4 |
+
from torchvision.transforms import Compose, Lambda, Resize, ToTensor, Normalize
|
| 5 |
+
from torchvision.transforms.functional import InterpolationMode
|
| 6 |
+
from transformers import AutoModel, AutoTokenizer
|
| 7 |
+
|
| 8 |
+
# --- Configuration ---
|
| 9 |
+
MODEL_NAME = "5CD-AI/Vintern-1B-v3_5"
|
| 10 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 11 |
+
DTYPE = torch.bfloat16 if DEVICE == "cuda" else torch.float32
|
| 12 |
+
|
| 13 |
+
# Image normalization constants
|
| 14 |
+
IMAGENET_MEAN = (0.485, 0.456, 0.406)
|
| 15 |
+
IMAGENET_STD = (0.229, 0.224, 0.225)
|
| 16 |
+
|
| 17 |
+
# --- Preprocessing Helpers ---
|
| 18 |
+
def build_transform(input_size):
|
| 19 |
+
return Compose([
|
| 20 |
+
Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
|
| 21 |
+
Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
|
| 22 |
+
ToTensor(),
|
| 23 |
+
Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
|
| 24 |
+
])
|
| 25 |
+
|
| 26 |
+
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
| 27 |
+
best_diff = float("inf")
|
| 28 |
+
best = (1, 1)
|
| 29 |
+
area = width * height
|
| 30 |
+
for (i, j) in target_ratios:
|
| 31 |
+
targ = i / j
|
| 32 |
+
diff = abs(aspect_ratio - targ)
|
| 33 |
+
if diff < best_diff or (diff == best_diff and area > 0.5 * image_size * image_size * i * j):
|
| 34 |
+
best_diff = diff
|
| 35 |
+
best = (i, j)
|
| 36 |
+
return best
|
| 37 |
+
|
| 38 |
+
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
|
| 39 |
+
w, h = image.size
|
| 40 |
+
ar = w / h
|
| 41 |
+
# generate all possible splits between min_num and max_num
|
| 42 |
+
ratios = sorted(
|
| 43 |
+
[(i, j)
|
| 44 |
+
for n in range(min_num, max_num+1)
|
| 45 |
+
for i in range(1, n+1)
|
| 46 |
+
for j in range(1, n+1)
|
| 47 |
+
if i*j <= max_num and i*j >= min_num],
|
| 48 |
+
key=lambda x: x[0]*x[1]
|
| 49 |
+
)
|
| 50 |
+
cols, rows = find_closest_aspect_ratio(ar, ratios, w, h, image_size)
|
| 51 |
+
target_w, target_h = cols * image_size, rows * image_size
|
| 52 |
+
resized = image.resize((target_w, target_h), resample=Image.BICUBIC)
|
| 53 |
|
| 54 |
+
tiles = []
|
| 55 |
+
num_cols = cols
|
| 56 |
+
for idx in range(cols * rows):
|
| 57 |
+
row_idx = idx // num_cols
|
| 58 |
+
col_idx = idx % num_cols
|
| 59 |
+
left = col_idx * image_size
|
| 60 |
+
upper = row_idx * image_size
|
| 61 |
+
right = left + image_size
|
| 62 |
+
lower = upper + image_size
|
| 63 |
+
tiles.append(resized.crop((left, upper, right, lower)))
|
| 64 |
+
|
| 65 |
+
# optionally add a centered thumbnail
|
| 66 |
+
if use_thumbnail and len(tiles) != 1:
|
| 67 |
+
thumb = image.resize((image_size, image_size), resample=Image.BICUBIC)
|
| 68 |
+
tiles.append(thumb)
|
| 69 |
+
|
| 70 |
+
return tiles
|
| 71 |
+
|
| 72 |
+
def load_image(path, input_size=448, max_num=6):
|
| 73 |
+
img = Image.open(path).convert("RGB")
|
| 74 |
+
tiles = dynamic_preprocess(img, max_num=max_num, image_size=input_size, use_thumbnail=True)
|
| 75 |
+
transform = build_transform(input_size)
|
| 76 |
+
pixel_values = torch.stack([transform(t) for t in tiles])
|
| 77 |
+
return pixel_values # shape: (num_tiles, 3, input_size, input_size)
|
| 78 |
+
|
| 79 |
+
# --- Model & Tokenizer ---
|
| 80 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 81 |
+
MODEL_NAME,
|
| 82 |
trust_remote_code=True,
|
| 83 |
+
use_fast=False
|
|
|
|
| 84 |
)
|
| 85 |
+
model = AutoModel.from_pretrained(
|
| 86 |
+
MODEL_NAME,
|
| 87 |
+
torch_dtype=DTYPE,
|
| 88 |
+
low_cpu_mem_usage=True,
|
| 89 |
+
trust_remote_code=True
|
| 90 |
+
).to(DEVICE).eval()
|
| 91 |
|
| 92 |
+
# Default generation settings
|
| 93 |
+
DEFAULT_GEN = {
|
| 94 |
+
"max_new_tokens": 128,
|
| 95 |
+
"do_sample": False,
|
| 96 |
+
"num_beams": 4,
|
| 97 |
+
"repetition_penalty": 1.2
|
| 98 |
+
}
|
| 99 |
+
|
| 100 |
+
# --- Alt-Text Generation ---
|
| 101 |
def generate_alt_text(files, max_new_tokens):
|
| 102 |
captions = []
|
| 103 |
for f in files or []:
|
| 104 |
+
# 1) preprocess image into tiles
|
| 105 |
+
pixels = load_image(f.name, max_num=6).to(DEVICE, DTYPE)
|
| 106 |
+
# 2) prompt the model
|
| 107 |
+
prompt = "<image>\nGenerate a concise, descriptive alt text for the image."
|
| 108 |
+
# 3) chat-style generation
|
| 109 |
+
response, _ = model.chat(
|
| 110 |
+
tokenizer,
|
| 111 |
+
pixels,
|
| 112 |
+
prompt,
|
| 113 |
+
{**DEFAULT_GEN, "max_new_tokens": int(max_new_tokens)},
|
| 114 |
+
history=None,
|
| 115 |
+
return_history=True
|
| 116 |
)
|
| 117 |
+
# 4) collect and clean
|
| 118 |
+
cap = response.strip()
|
| 119 |
+
if cap:
|
| 120 |
+
cap = cap[0].upper() + cap[1:]
|
| 121 |
+
captions.append(cap)
|
| 122 |
return "\n\n".join(captions)
|
| 123 |
|
| 124 |
# --- Gradio UI ---
|
| 125 |
+
with gr.Blocks(title="Vintern-1B-v3.5 Alt Text Generator") as demo:
|
| 126 |
+
gr.Markdown("# Vintern‑1B‑v3.5 Alt Text Generator")
|
| 127 |
+
gr.Markdown("Model fine‑tuned for Vietnamese OCR & document tasks, repurposed for alt‑text.")
|
| 128 |
|
| 129 |
with gr.Row():
|
| 130 |
+
file_input = gr.File(label="Upload Images", file_count="multiple", file_types=["image"])
|
| 131 |
+
slider = gr.Slider(16, 512, value=128, step=16, label="Max New Tokens")
|
| 132 |
+
run_btn = gr.Button("Generate Alt Text")
|
| 133 |
+
clear_btn = gr.Button("Clear All")
|
| 134 |
|
| 135 |
gallery = gr.Gallery(label="Image Preview", columns=4)
|
| 136 |
+
output = gr.Textbox(label="Generated Alt Texts", lines=10, show_copy_button=True)
|
| 137 |
|
| 138 |
+
file_input.change(
|
| 139 |
+
fn=lambda files: [Image.open(f.name).convert("RGB") for f in (files or [])],
|
| 140 |
+
inputs=file_input, outputs=gallery
|
| 141 |
)
|
|
|
|
| 142 |
run_btn.click(
|
| 143 |
fn=generate_alt_text,
|
| 144 |
+
inputs=[file_input, slider],
|
| 145 |
outputs=output,
|
| 146 |
queue=True
|
| 147 |
)
|
|
|
|
| 148 |
clear_btn.click(
|
| 149 |
+
fn=lambda: ([], 128, [], ""),
|
| 150 |
+
outputs=[file_input, slider, gallery, output]
|
| 151 |
)
|
| 152 |
|
| 153 |
demo.launch()
|