Spaces:
Running on Zero
Running on Zero
Build both stages once in the worker, and declare the backbone's license
Browse filesBuilds both stages once, in the worker, rather than per call, moves the beat checkpoint link and downloads into the worker, declares CC-BY-NC-4.0, the license of the MERT backbone, and says so in the README, and declares 24 kHz mono. It also imports spaces plainly, since it is always installed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- README.md +4 -2
- app.py +44 -63
- inference.py +26 -10
README.md
CHANGED
|
@@ -10,18 +10,20 @@ preload_from_hub:
|
|
| 10 |
- teamup-tech/FST-AI-Music-Detection-checkpoints
|
| 11 |
app_file: app.py
|
| 12 |
pinned: false
|
| 13 |
-
license:
|
| 14 |
short_description: Detects AI-generated music from structure and MERT
|
| 15 |
tags:
|
| 16 |
- category:analysis
|
| 17 |
- subcategory:music-analysis
|
| 18 |
- input:audio
|
| 19 |
- output:file/txt
|
|
|
|
|
|
|
| 20 |
- music
|
| 21 |
---
|
| 22 |
|
| 23 |
Detects AI-generated music from structure and MERT.
|
| 24 |
|
| 25 |
-
Deploys [FST-AI-Music-Detection](https://github.com/Mippia/FST-AI-Music-Detection) for HARP. Upstream states its license in its README rather than in a `LICENSE` file,
|
| 26 |
|
| 27 |
A HARP deployment. See https://github.com/TEAMuP-dev/HARP
|
|
|
|
| 10 |
- teamup-tech/FST-AI-Music-Detection-checkpoints
|
| 11 |
app_file: app.py
|
| 12 |
pinned: false
|
| 13 |
+
license: cc-by-nc-4.0
|
| 14 |
short_description: Detects AI-generated music from structure and MERT
|
| 15 |
tags:
|
| 16 |
- category:analysis
|
| 17 |
- subcategory:music-analysis
|
| 18 |
- input:audio
|
| 19 |
- output:file/txt
|
| 20 |
+
- sample-rate:24000
|
| 21 |
+
- channels:mono
|
| 22 |
- music
|
| 23 |
---
|
| 24 |
|
| 25 |
Detects AI-generated music from structure and MERT.
|
| 26 |
|
| 27 |
+
Deploys [FST-AI-Music-Detection](https://github.com/Mippia/FST-AI-Music-Detection) for HARP. Upstream states its license in its README rather than in a `LICENSE` file: Mippia/FST-AI-Music-Detection licenses the code GPL, and leaves the checkpoints unstated. The model runs on the MERT-v1-95M backbone, whose weights are CC-BY-NC-4.0, so the `license` field above declares that, the more restrictive of the two.
|
| 28 |
|
| 29 |
A HARP deployment. See https://github.com/TEAMuP-dev/HARP
|
app.py
CHANGED
|
@@ -1,21 +1,8 @@
|
|
|
|
|
|
|
|
| 1 |
import sys
|
| 2 |
sys.stdout.reconfigure(line_buffering=True)
|
| 3 |
|
| 4 |
-
try:
|
| 5 |
-
import spaces
|
| 6 |
-
except ImportError:
|
| 7 |
-
# keep @spaces.GPU usable as a no-op; ZeroGPU requires this exact name.
|
| 8 |
-
class spaces:
|
| 9 |
-
class GPU:
|
| 10 |
-
def __init__(self, func=None, duration=60):
|
| 11 |
-
self.func = func
|
| 12 |
-
|
| 13 |
-
def __call__(self, *args, **kwargs):
|
| 14 |
-
if self.func is not None:
|
| 15 |
-
return self.func(*args, **kwargs)
|
| 16 |
-
func = args[0]
|
| 17 |
-
return func
|
| 18 |
-
|
| 19 |
import gradio as gr
|
| 20 |
import torch
|
| 21 |
import librosa
|
|
@@ -79,83 +66,77 @@ torchaudio.info = _patched_info
|
|
| 79 |
|
| 80 |
import contextlib
|
| 81 |
import tempfile
|
| 82 |
-
import threading
|
| 83 |
from pathlib import Path
|
| 84 |
|
| 85 |
from huggingface_hub import hf_hub_download
|
| 86 |
from pyharp import ModelCard, build_endpoint
|
| 87 |
-
from pyharp.tags import Subcategory
|
| 88 |
-
|
| 89 |
-
from inference import inference
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
|
|
|
|
|
|
| 102 |
|
|
|
|
| 103 |
CHECKPOINT_REPO = "teamup-tech/FST-AI-Music-Detection-checkpoints"
|
| 104 |
CHECKPOINT_DIR = Path("checkpoints")
|
| 105 |
|
| 106 |
-
checkpoints_loading = True
|
| 107 |
checkpoint_error = None
|
| 108 |
|
| 109 |
|
| 110 |
-
def
|
| 111 |
-
"""Fetch the Stage-1 and Stage-2 checkpoints
|
| 112 |
-
|
| 113 |
-
global
|
| 114 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
CHECKPOINT_DIR.mkdir(exist_ok=True)
|
| 116 |
for filename in ("Stage-1.ckpt", "Stage-2.ckpt"):
|
| 117 |
hf_hub_download(repo_id=CHECKPOINT_REPO, filename=filename, local_dir=str(CHECKPOINT_DIR),
|
| 118 |
revision="801f36e6103034bbd0d20fcbfa6a1a1ad5e919d3")
|
|
|
|
| 119 |
print("Checkpoints ready.")
|
| 120 |
except Exception as e:
|
| 121 |
checkpoint_error = str(e)
|
| 122 |
-
print(f"Checkpoint
|
| 123 |
-
finally:
|
| 124 |
-
checkpoints_loading = False
|
| 125 |
-
|
| 126 |
|
| 127 |
-
_loader = threading.Thread(target=download_checkpoints, daemon=True)
|
| 128 |
-
_loader.start()
|
| 129 |
|
| 130 |
-
# pyharp
|
| 131 |
-
# __main__)
|
| 132 |
-
#
|
| 133 |
-
#
|
| 134 |
if __name__ != "__main__":
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
model_card = ModelCard(
|
| 139 |
-
name="Fusion Segment Transformer",
|
| 140 |
-
description=(
|
| 141 |
-
"Detects AI-generated music by fusing MERT audio embeddings with a "
|
| 142 |
-
"structural, downbeat-segmented view of the track. Submitted to ICASSP "
|
| 143 |
-
"2026. Repository: https://github.com/Mippia/FST-AI-Music-Detection "
|
| 144 |
-
"Paper: https://arxiv.org/abs/2601.13647"
|
| 145 |
-
),
|
| 146 |
-
author="Yumin Kim, Seonghyeon Go",
|
| 147 |
-
tags=[Subcategory.MUSIC_ANALYSIS, 'music'],
|
| 148 |
-
)
|
| 149 |
|
| 150 |
|
| 151 |
def process_fn(audio_path: str) -> str:
|
| 152 |
"""Runs the two-stage detector on an uploaded track and writes the verdict to a text file."""
|
| 153 |
# Checked before the GPU is reserved, so a request that cannot run spends none of the
|
| 154 |
# caller's ZeroGPU allowance
|
| 155 |
-
if checkpoints_loading:
|
| 156 |
-
raise gr.Error("Model checkpoints are still downloading, please wait a moment and try again.")
|
| 157 |
if checkpoint_error:
|
| 158 |
-
raise gr.Error(f"Failed to
|
| 159 |
if not audio_path:
|
| 160 |
raise gr.Error("Provide the music to test.")
|
| 161 |
try:
|
|
|
|
| 1 |
+
import spaces # imported first, as ZeroGPU requires
|
| 2 |
+
|
| 3 |
import sys
|
| 4 |
sys.stdout.reconfigure(line_buffering=True)
|
| 5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
import gradio as gr
|
| 7 |
import torch
|
| 8 |
import librosa
|
|
|
|
| 66 |
|
| 67 |
import contextlib
|
| 68 |
import tempfile
|
|
|
|
| 69 |
from pathlib import Path
|
| 70 |
|
| 71 |
from huggingface_hub import hf_hub_download
|
| 72 |
from pyharp import ModelCard, build_endpoint
|
| 73 |
+
from pyharp.tags import Channels, SampleRate, Subcategory
|
| 74 |
+
|
| 75 |
+
from inference import inference, load_models
|
| 76 |
+
|
| 77 |
+
model_card = ModelCard(
|
| 78 |
+
name="Fusion Segment Transformer",
|
| 79 |
+
description=(
|
| 80 |
+
"Detects AI-generated music by fusing MERT audio embeddings with a "
|
| 81 |
+
"structural, downbeat-segmented view of the track. Submitted to ICASSP "
|
| 82 |
+
"2026. The track is mixed to mono and resampled to 24 kHz internally. The code is "
|
| 83 |
+
"GPL, but the MERT-v1-95M backbone it runs on is CC-BY-NC-4.0, so results cannot be "
|
| 84 |
+
"used commercially. Repository: https://github.com/Mippia/FST-AI-Music-Detection "
|
| 85 |
+
"Paper: https://arxiv.org/abs/2601.13647"
|
| 86 |
+
),
|
| 87 |
+
author="Yumin Kim, Seonghyeon Go",
|
| 88 |
+
tags=[Subcategory.MUSIC_ANALYSIS, SampleRate(24000), Channels(1), 'music'],
|
| 89 |
+
)
|
| 90 |
|
| 91 |
+
DEVICE = "cuda:0" if torch.cuda.is_available() else "cpu"
|
| 92 |
CHECKPOINT_REPO = "teamup-tech/FST-AI-Music-Detection-checkpoints"
|
| 93 |
CHECKPOINT_DIR = Path("checkpoints")
|
| 94 |
|
|
|
|
| 95 |
checkpoint_error = None
|
| 96 |
|
| 97 |
|
| 98 |
+
def load_checkpoints():
|
| 99 |
+
"""Fetch the Stage-1 and Stage-2 checkpoints, where inference.py looks for them, and build
|
| 100 |
+
both stages from them, so that no request downloads or builds anything."""
|
| 101 |
+
global checkpoint_error
|
| 102 |
try:
|
| 103 |
+
# beat_this downloads its checkpoint from the JKU cloud on first use, which has failed
|
| 104 |
+
# both inside a GPU call (no name resolution) and at startup (503). The same file is
|
| 105 |
+
# mirrored on the Hub and linked where torch.hub looks before downloading.
|
| 106 |
+
beat_checkpoint = Path(torch.hub.get_dir()) / "checkpoints" / "beat_this-final0.ckpt"
|
| 107 |
+
if not beat_checkpoint.exists():
|
| 108 |
+
beat_checkpoint.parent.mkdir(parents=True, exist_ok=True)
|
| 109 |
+
with contextlib.suppress(FileExistsError):
|
| 110 |
+
beat_checkpoint.symlink_to(hf_hub_download(
|
| 111 |
+
"teamup-tech/Beat-This-weights", "final0.ckpt",
|
| 112 |
+
revision="4afbe376535aaab71b76ae2ca2d7c16b08877321",
|
| 113 |
+
))
|
| 114 |
+
|
| 115 |
CHECKPOINT_DIR.mkdir(exist_ok=True)
|
| 116 |
for filename in ("Stage-1.ckpt", "Stage-2.ckpt"):
|
| 117 |
hf_hub_download(repo_id=CHECKPOINT_REPO, filename=filename, local_dir=str(CHECKPOINT_DIR),
|
| 118 |
revision="801f36e6103034bbd0d20fcbfa6a1a1ad5e919d3")
|
| 119 |
+
load_models(DEVICE)
|
| 120 |
print("Checkpoints ready.")
|
| 121 |
except Exception as e:
|
| 122 |
checkpoint_error = str(e)
|
| 123 |
+
print(f"Checkpoint error: {e}")
|
|
|
|
|
|
|
|
|
|
| 124 |
|
|
|
|
|
|
|
| 125 |
|
| 126 |
+
# pyharp runs process_fn in a worker, which imports this file as __mp_main__ (rather than
|
| 127 |
+
# __main__) and starts with the server, so everything is fetched and built there alone, on its
|
| 128 |
+
# main thread, where ZeroGPU's emulation records each placement, before process_fn can reserve a
|
| 129 |
+
# GPU. A failure is kept rather than raised, so that each request can report it.
|
| 130 |
if __name__ != "__main__":
|
| 131 |
+
load_checkpoints()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 132 |
|
| 133 |
|
| 134 |
def process_fn(audio_path: str) -> str:
|
| 135 |
"""Runs the two-stage detector on an uploaded track and writes the verdict to a text file."""
|
| 136 |
# Checked before the GPU is reserved, so a request that cannot run spends none of the
|
| 137 |
# caller's ZeroGPU allowance
|
|
|
|
|
|
|
| 138 |
if checkpoint_error:
|
| 139 |
+
raise gr.Error(f"Failed to load the model: {checkpoint_error}")
|
| 140 |
if not audio_path:
|
| 141 |
raise gr.Error("Provide the music to test.")
|
| 142 |
try:
|
inference.py
CHANGED
|
@@ -193,24 +193,40 @@ def get_model(model_type, device):
|
|
| 193 |
model.eval()
|
| 194 |
return model, embed_dim
|
| 195 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 196 |
def inference(audio_path):
|
| 197 |
# device ์ค์ ์ ๋ช
ํํ ํ๊ธฐ
|
| 198 |
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
|
| 199 |
-
|
| 200 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 201 |
segments, padding_mask = load_audio(audio_path, sr=24000)
|
| 202 |
segments = segments.to(device).to(torch.float32)
|
| 203 |
padding_mask = padding_mask.to(device).unsqueeze(0)
|
| 204 |
logits, embedding = backbone_model(segments.squeeze(1))
|
| 205 |
|
| 206 |
-
# ๋ชจ๋ธ ๋ก๋ํ ๋๋ map_location ์ถ๊ฐ
|
| 207 |
-
# pointing ckpt to files loaded on teamup-tech's HF account
|
| 208 |
-
model = MusicAudioClassifier.load_from_checkpoint(
|
| 209 |
-
checkpoint_path='checkpoints/Stage-2.ckpt',
|
| 210 |
-
input_dim=input_dim,
|
| 211 |
-
map_location=device # ์ด ๋ถ๋ถ ์ถ๊ฐ
|
| 212 |
-
)
|
| 213 |
-
|
| 214 |
# Run inference
|
| 215 |
print(f"Segments shape: {segments.shape}")
|
| 216 |
print("Running inference...")
|
|
|
|
| 193 |
model.eval()
|
| 194 |
return model, embed_dim
|
| 195 |
|
| 196 |
+
_MODELS = {}
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
def load_models(device):
|
| 200 |
+
"""
|
| 201 |
+
Build both stages once, on the CPU, and place them on the device, so that no request builds
|
| 202 |
+
them. The Space calls this in its worker at startup.
|
| 203 |
+
"""
|
| 204 |
+
# pointing ckpt to files loaded on teamup-tech's HF account
|
| 205 |
+
backbone_model = MERT_AudioCNN.load_from_checkpoint(
|
| 206 |
+
'checkpoints/Stage-1.ckpt', map_location='cpu'
|
| 207 |
+
).to(device).eval()
|
| 208 |
+
classifier = MusicAudioClassifier.load_from_checkpoint(
|
| 209 |
+
checkpoint_path='checkpoints/Stage-2.ckpt',
|
| 210 |
+
input_dim=768,
|
| 211 |
+
map_location='cpu',
|
| 212 |
+
).to(device).eval()
|
| 213 |
+
_MODELS.update(backbone=backbone_model, classifier=classifier)
|
| 214 |
+
|
| 215 |
+
|
| 216 |
def inference(audio_path):
|
| 217 |
# device ์ค์ ์ ๋ช
ํํ ํ๊ธฐ
|
| 218 |
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
|
| 219 |
+
|
| 220 |
+
if not _MODELS:
|
| 221 |
+
load_models(device)
|
| 222 |
+
backbone_model = _MODELS["backbone"]
|
| 223 |
+
model = _MODELS["classifier"]
|
| 224 |
+
|
| 225 |
segments, padding_mask = load_audio(audio_path, sr=24000)
|
| 226 |
segments = segments.to(device).to(torch.float32)
|
| 227 |
padding_mask = padding_mask.to(device).unsqueeze(0)
|
| 228 |
logits, embedding = backbone_model(segments.squeeze(1))
|
| 229 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 230 |
# Run inference
|
| 231 |
print(f"Segments shape: {segments.shape}")
|
| 232 |
print("Running inference...")
|