Upload 3 files
Browse files- README.md +7 -7
- app.py +89 -0
- requirements.txt +10 -0
README.md
CHANGED
|
@@ -1,13 +1,13 @@
|
|
| 1 |
---
|
| 2 |
-
title: Fuse
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: gradio
|
| 7 |
-
sdk_version: 6.24.0
|
| 8 |
-
python_version: '3.12'
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
---
|
| 12 |
|
| 13 |
-
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Fuse-MD
|
| 3 |
+
emoji: 🔎
|
| 4 |
+
colorFrom: blue
|
| 5 |
+
colorTo: purple
|
| 6 |
sdk: gradio
|
|
|
|
|
|
|
| 7 |
app_file: app.py
|
| 8 |
pinned: false
|
| 9 |
---
|
| 10 |
|
| 11 |
+
# Fuse-MD
|
| 12 |
+
|
| 13 |
+
Research demo for multimodal misogyny-meme detection in Malayalam.
|
app.py
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Fuse-MD Hugging Face ZeroGPU demo.
|
| 2 |
+
|
| 3 |
+
The Space must also contain the ``src`` and ``api`` folders and the selected
|
| 4 |
+
checkpoint at the repository root. The large text backbone is downloaded from
|
| 5 |
+
the Hugging Face Hub during startup.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
import os
|
| 9 |
+
import sys
|
| 10 |
+
import time
|
| 11 |
+
from pathlib import Path
|
| 12 |
+
|
| 13 |
+
try:
|
| 14 |
+
import spaces
|
| 15 |
+
GPU = spaces.GPU
|
| 16 |
+
except ImportError: # pragma: no cover - local-only fallback
|
| 17 |
+
def GPU(*args, **kwargs):
|
| 18 |
+
def decorate(function):
|
| 19 |
+
return function
|
| 20 |
+
return decorate
|
| 21 |
+
|
| 22 |
+
import gradio as gr
|
| 23 |
+
from huggingface_hub import snapshot_download
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
ROOT = Path(__file__).resolve().parent
|
| 27 |
+
sys.path.insert(0, str(ROOT))
|
| 28 |
+
|
| 29 |
+
TEXT_MODEL = "VishnuPJ/MalayaLLM_7B_Base"
|
| 30 |
+
CHECKPOINT = ROOT / "custom_malayalam_llamavit_fusion_element.pth"
|
| 31 |
+
MODEL_DIR = ROOT / ".hf_model_cache" / "VishnuPJ__MalayaLLM_7B_Base"
|
| 32 |
+
|
| 33 |
+
# Download one text backbone only. Keeping a single language model reduces
|
| 34 |
+
# startup time and storage usage for the free demo.
|
| 35 |
+
if not MODEL_DIR.exists():
|
| 36 |
+
snapshot_download(repo_id=TEXT_MODEL, local_dir=MODEL_DIR)
|
| 37 |
+
|
| 38 |
+
os.environ.setdefault("FUSEMD_CHECKPOINT", str(CHECKPOINT))
|
| 39 |
+
os.environ.setdefault("FUSEMD_LOCAL_MODEL_ROOT", str(ROOT / ".hf_model_cache"))
|
| 40 |
+
os.environ.setdefault("FUSEMD_DEVICE", "cuda")
|
| 41 |
+
|
| 42 |
+
from api.api.config import settings # noqa: E402
|
| 43 |
+
from api.api.service import FuseMDService # noqa: E402
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
service = FuseMDService(settings)
|
| 47 |
+
service.load()
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
@GPU(duration=120)
|
| 51 |
+
def predict(image_path, text, threshold):
|
| 52 |
+
if image_path is None:
|
| 53 |
+
return "Please upload a meme image."
|
| 54 |
+
|
| 55 |
+
started = time.perf_counter()
|
| 56 |
+
with open(image_path, "rb") as image_file:
|
| 57 |
+
response = service.predict(
|
| 58 |
+
text=text or "",
|
| 59 |
+
image_bytes=image_file.read(),
|
| 60 |
+
image_filename=Path(image_path).name,
|
| 61 |
+
threshold=threshold,
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
elapsed = time.perf_counter() - started
|
| 65 |
+
return (
|
| 66 |
+
f"### Prediction: **{response['label']}**\n\n"
|
| 67 |
+
f"Positive probability: **{response['probability']:.1%}**\n\n"
|
| 68 |
+
f"Threshold: {response['threshold']:.2f} | "
|
| 69 |
+
f"Language: {response['language']} | Fusion: {response['fusion_method']}\n\n"
|
| 70 |
+
f"Runtime: **{elapsed:.1f} seconds**"
|
| 71 |
+
)
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
with gr.Blocks(title="Fuse-MD Demo") as demo:
|
| 75 |
+
gr.Markdown(
|
| 76 |
+
"# Fuse-MD\n"
|
| 77 |
+
"Upload a Malayalam meme and enter its transcription. The model predicts "
|
| 78 |
+
"whether it contains misogynistic content."
|
| 79 |
+
)
|
| 80 |
+
with gr.Row():
|
| 81 |
+
image = gr.Image(type="filepath", label="Meme image", width=420, height=420)
|
| 82 |
+
text = gr.Textbox(label="Meme transcription", lines=5)
|
| 83 |
+
threshold = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="Decision threshold")
|
| 84 |
+
button = gr.Button("Predict", variant="primary")
|
| 85 |
+
result = gr.Markdown()
|
| 86 |
+
button.click(predict, inputs=[image, text, threshold], outputs=result)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
demo.queue().launch()
|
requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
spaces
|
| 2 |
+
gradio
|
| 3 |
+
huggingface_hub
|
| 4 |
+
transformers
|
| 5 |
+
accelerate
|
| 6 |
+
bitsandbytes
|
| 7 |
+
timm
|
| 8 |
+
opencv-python-headless
|
| 9 |
+
fastapi
|
| 10 |
+
python-multipart
|