Spaces:
Running
on
Zero
Running
on
Zero
roychao19477
commited on
Commit
·
1728184
1
Parent(s):
a45a351
Upload model
Browse files
app_v1.py
ADDED
@@ -0,0 +1,358 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import shlex
|
2 |
+
import subprocess
|
3 |
+
import spaces
|
4 |
+
import torch
|
5 |
+
import os
|
6 |
+
import shutil
|
7 |
+
import glob
|
8 |
+
import gradio as gr
|
9 |
+
|
10 |
+
# install packages for mamba
|
11 |
+
def install_mamba():
|
12 |
+
subprocess.run(shlex.split("pip install https://github.com/state-spaces/mamba/releases/download/v2.2.2/mamba_ssm-2.2.2+cu122torch2.3cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"))
|
13 |
+
|
14 |
+
def clone_github():
|
15 |
+
subprocess.run([
|
16 |
+
"git", "clone",
|
17 |
+
f"https://RoyChao19477:{os.environ['GITHUB_TOKEN']}@github.com/RoyChao19477/for_HF_AVSEMamba.git",
|
18 |
+
])
|
19 |
+
# move all files except README.md
|
20 |
+
for item in glob.glob("for_HF_AVSEMamba/*"):
|
21 |
+
if os.path.basename(item) != "README.md":
|
22 |
+
if os.path.isdir(item):
|
23 |
+
shutil.move(item, ".")
|
24 |
+
else:
|
25 |
+
shutil.move(item, os.path.join(".", os.path.basename(item)))
|
26 |
+
|
27 |
+
#shutil.rmtree("tmp_repo")
|
28 |
+
#subprocess.run(["ls"], check=True)
|
29 |
+
|
30 |
+
install_mamba()
|
31 |
+
clone_github()
|
32 |
+
|
33 |
+
ABOUT = """
|
34 |
+
# SEMamba: Speech Enhancement
|
35 |
+
A Mamba-based model that denoises real-world audio.
|
36 |
+
Upload or record a noisy clip and click **Enhance** to hear + see its spectrogram.
|
37 |
+
"""
|
38 |
+
|
39 |
+
|
40 |
+
import torch
|
41 |
+
import ffmpeg
|
42 |
+
import torchaudio
|
43 |
+
import torchaudio.transforms as T
|
44 |
+
import yaml
|
45 |
+
import librosa
|
46 |
+
import librosa.display
|
47 |
+
import matplotlib
|
48 |
+
import numpy as np
|
49 |
+
import soundfile as sf
|
50 |
+
import matplotlib.pyplot as plt
|
51 |
+
from models.stfts import mag_phase_stft, mag_phase_istft
|
52 |
+
from models.generator import SEMamba
|
53 |
+
from models.pcs400 import cal_pcs
|
54 |
+
from ultralytics import YOLO
|
55 |
+
import supervision as sv
|
56 |
+
|
57 |
+
import gradio as gr
|
58 |
+
import cv2
|
59 |
+
import os
|
60 |
+
import tempfile
|
61 |
+
from ultralytics import YOLO
|
62 |
+
from moviepy import ImageSequenceClip
|
63 |
+
from scipy.io import wavfile
|
64 |
+
from avse_code import run_avse
|
65 |
+
|
66 |
+
# Load face detector
|
67 |
+
model = YOLO("yolov8n-face.pt").cuda() # assumes CUDA available
|
68 |
+
|
69 |
+
|
70 |
+
from decord import VideoReader, cpu
|
71 |
+
from model import AVSEModule
|
72 |
+
from config import sampling_rate
|
73 |
+
import spaces
|
74 |
+
|
75 |
+
# Load model once globally
|
76 |
+
#ckpt_path = "ckpts/ep215_0906.oat.ckpt"
|
77 |
+
#model = AVSEModule.load_from_checkpoint(ckpt_path)
|
78 |
+
avse_model = AVSEModule()
|
79 |
+
#avse_state_dict = torch.load("ckpts/ep215_0906.oat.ckpt")
|
80 |
+
avse_state_dict = torch.load("ckpts/ep220_0908.oat.ckpt")
|
81 |
+
avse_model.load_state_dict(avse_state_dict, strict=True)
|
82 |
+
avse_model.to("cuda")
|
83 |
+
avse_model.eval()
|
84 |
+
|
85 |
+
@spaces.GPU
|
86 |
+
def run_avse_inference(video_path, audio_path):
|
87 |
+
estimated = run_avse(video_path, audio_path)
|
88 |
+
# Load audio
|
89 |
+
#noisy, _ = sf.read(audio_path, dtype='float32') # (N, )
|
90 |
+
#noisy = torch.tensor(noisy).unsqueeze(0) # (1, N)
|
91 |
+
noisy = wavfile.read(audio_path)[1].astype(np.float32) / (2 ** 15)
|
92 |
+
|
93 |
+
# Norm.
|
94 |
+
#noisy = noisy * (0.8 / np.max(np.abs(noisy)))
|
95 |
+
|
96 |
+
# Load grayscale video
|
97 |
+
vr = VideoReader(video_path, ctx=cpu(0))
|
98 |
+
frames = vr.get_batch(list(range(len(vr)))).asnumpy()
|
99 |
+
bg_frames = np.array([
|
100 |
+
cv2.cvtColor(frames[i], cv2.COLOR_RGB2GRAY) for i in range(len(frames))
|
101 |
+
]).astype(np.float32)
|
102 |
+
bg_frames /= 255.0
|
103 |
+
|
104 |
+
|
105 |
+
# Combine into input dict (match what model.enhance expects)
|
106 |
+
data = {
|
107 |
+
"noisy_audio": noisy,
|
108 |
+
"video_frames": bg_frames[np.newaxis, ...]
|
109 |
+
}
|
110 |
+
|
111 |
+
with torch.no_grad():
|
112 |
+
estimated = avse_model.enhance(data).reshape(-1)
|
113 |
+
|
114 |
+
# Save result
|
115 |
+
tmp_wav = audio_path.replace(".wav", "_enhanced.wav")
|
116 |
+
sf.write(tmp_wav, estimated, samplerate=sampling_rate)
|
117 |
+
|
118 |
+
return tmp_wav
|
119 |
+
|
120 |
+
|
121 |
+
def extract_resampled_audio(video_path, target_sr=16000):
|
122 |
+
# Step 1: extract audio via torchaudio
|
123 |
+
# (moviepy will still extract it to wav temp file)
|
124 |
+
tmp_audio_path = tempfile.mktemp(suffix=".wav")
|
125 |
+
subprocess.run(["ffmpeg", "-y", "-i", video_path, "-vn", "-acodec", "pcm_s16le", "-ar", "44100", tmp_audio_path])
|
126 |
+
|
127 |
+
# Step 2: Load and resample
|
128 |
+
waveform, sr = torchaudio.load(tmp_audio_path)
|
129 |
+
if sr != target_sr:
|
130 |
+
resampler = T.Resample(orig_freq=sr, new_freq=target_sr)
|
131 |
+
waveform = resampler(waveform)
|
132 |
+
|
133 |
+
# Step 3: Save resampled audio
|
134 |
+
resampled_audio_path = tempfile.mktemp(suffix="_16k.wav")
|
135 |
+
torchaudio.save(resampled_audio_path, waveform, sample_rate=target_sr)
|
136 |
+
return resampled_audio_path
|
137 |
+
|
138 |
+
@spaces.GPU
|
139 |
+
def extract_faces(video_file):
|
140 |
+
cap = cv2.VideoCapture(video_file)
|
141 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
142 |
+
frames = []
|
143 |
+
|
144 |
+
while True:
|
145 |
+
ret, frame = cap.read()
|
146 |
+
if not ret:
|
147 |
+
break
|
148 |
+
|
149 |
+
# Inference
|
150 |
+
results = model(frame, verbose=False)[0]
|
151 |
+
for box in results.boxes:
|
152 |
+
# version 1
|
153 |
+
# x1, y1, x2, y2 = map(int, box.xyxy[0])
|
154 |
+
|
155 |
+
# version 2
|
156 |
+
h, w, _ = frame.shape
|
157 |
+
x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
|
158 |
+
pad_ratio = 0.5 # 30% padding
|
159 |
+
|
160 |
+
dx = (x2 - x1) * pad_ratio
|
161 |
+
dy = (y2 - y1) * pad_ratio
|
162 |
+
|
163 |
+
x1 = int(max(0, x1 - dx))
|
164 |
+
y1 = int(max(0, y1 - dy))
|
165 |
+
x2 = int(min(w, x2 + dx))
|
166 |
+
y2 = int(min(h, y2 + dy))
|
167 |
+
# Added for v3
|
168 |
+
shift_down = int(0.1 * (y2 - y1))
|
169 |
+
y1 = int(min(max(0, y1 + shift_down), h))
|
170 |
+
y2 = int(min(max(0, y2 + shift_down), h))
|
171 |
+
face_crop = frame[y1:y2, x1:x2]
|
172 |
+
if face_crop.size != 0:
|
173 |
+
resized = cv2.resize(face_crop, (224, 224))
|
174 |
+
frames.append(resized)
|
175 |
+
|
176 |
+
#h_crop, w_crop = face_crop.shape[:2]
|
177 |
+
#side = min(h_crop, w_crop)
|
178 |
+
#start_y = (h_crop - side) // 2
|
179 |
+
#start_x = (w_crop - side) // 2
|
180 |
+
#square_crop = face_crop[start_y:start_y+side, start_x:start_x+side]
|
181 |
+
#resized = cv2.resize(square_crop, (224, 224))
|
182 |
+
#frames.append(resized)
|
183 |
+
|
184 |
+
break # only one face per frame
|
185 |
+
|
186 |
+
cap.release()
|
187 |
+
|
188 |
+
# Save as video
|
189 |
+
tmpdir = tempfile.mkdtemp()
|
190 |
+
output_path = os.path.join(tmpdir, "face_only_video.mp4")
|
191 |
+
#clip = ImageSequenceClip([cv2.cvtColor(f, cv2.COLOR_BGR2RGB) for f in frames], fps=25)
|
192 |
+
#clip = ImageSequenceClip([cv2.cvtColor(f, cv2.COLOR_BGR2RGB) for f in frames], fps=fps)
|
193 |
+
clip = ImageSequenceClip(
|
194 |
+
[cv2.cvtColor(cv2.resize(f, (224, 224)), cv2.COLOR_BGR2RGB) for f in frames],
|
195 |
+
fps=fps
|
196 |
+
)
|
197 |
+
clip.write_videofile(output_path, codec="libx264", audio=False, fps=25)
|
198 |
+
|
199 |
+
# Save audio from original, resampled to 16kHz
|
200 |
+
audio_path = os.path.join(tmpdir, "audio_16k.wav")
|
201 |
+
|
202 |
+
# Extract audio using ffmpeg-python (more robust than moviepy)
|
203 |
+
ffmpeg.input(video_file).output(
|
204 |
+
audio_path,
|
205 |
+
ar=16000, # resample to 16k
|
206 |
+
ac=1, # mono
|
207 |
+
format='wav',
|
208 |
+
vn=None # no video
|
209 |
+
).run(overwrite_output=True)
|
210 |
+
|
211 |
+
|
212 |
+
|
213 |
+
|
214 |
+
# ------------------------------- #
|
215 |
+
# AVSE models
|
216 |
+
|
217 |
+
enhanced_audio_path = run_avse_inference(output_path, audio_path)
|
218 |
+
|
219 |
+
|
220 |
+
return output_path, enhanced_audio_path
|
221 |
+
#return output_path, audio_path
|
222 |
+
|
223 |
+
iface = gr.Interface(
|
224 |
+
fn=extract_faces,
|
225 |
+
inputs=gr.Video(label="Upload or record your video"),
|
226 |
+
outputs=[
|
227 |
+
gr.Video(label="Detected Face Only Video"),
|
228 |
+
#gr.Audio(label="Extracted Audio (16kHz)", type="filepath"),
|
229 |
+
gr.Audio(label="Enhanced Audio", type="filepath")
|
230 |
+
],
|
231 |
+
title="Face Detector",
|
232 |
+
description="Upload or record a video. We'll crop face regions and return a face-only video and its 16kHz audio."
|
233 |
+
)
|
234 |
+
|
235 |
+
iface.launch()
|
236 |
+
|
237 |
+
|
238 |
+
|
239 |
+
ckpt = "ckpts/SEMamba_advanced.pth"
|
240 |
+
cfg_f = "recipes/SEMamba_advanced.yaml"
|
241 |
+
|
242 |
+
# load config
|
243 |
+
with open(cfg_f, 'r') as f:
|
244 |
+
cfg = yaml.safe_load(f)
|
245 |
+
|
246 |
+
|
247 |
+
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
248 |
+
device = "cuda"
|
249 |
+
model = SEMamba(cfg).to(device)
|
250 |
+
#sdict = torch.load(ckpt, map_location=device)
|
251 |
+
#model.load_state_dict(sdict["generator"])
|
252 |
+
#model.eval()
|
253 |
+
|
254 |
+
@spaces.GPU
|
255 |
+
def enhance(filepath, model_name):
|
256 |
+
# Load model based on selection
|
257 |
+
ckpt_path = {
|
258 |
+
"VCTK-Demand": "ckpts/SEMamba_advanced.pth",
|
259 |
+
"VCTK+DNS": "ckpts/vd.pth"
|
260 |
+
}[model_name]
|
261 |
+
|
262 |
+
print("Loading:", ckpt_path)
|
263 |
+
model.load_state_dict(torch.load(ckpt_path, map_location=device)["generator"])
|
264 |
+
model.eval()
|
265 |
+
with torch.no_grad():
|
266 |
+
# load & resample
|
267 |
+
wav, orig_sr = librosa.load(filepath, sr=None)
|
268 |
+
noisy_wav = wav.copy()
|
269 |
+
if orig_sr != 16000:
|
270 |
+
wav = librosa.resample(wav, orig_sr=orig_sr, target_sr=16000)
|
271 |
+
x = torch.from_numpy(wav).float().to(device)
|
272 |
+
norm = torch.sqrt(len(x)/torch.sum(x**2))
|
273 |
+
#x = (x * norm).unsqueeze(0)
|
274 |
+
x = (x * norm)
|
275 |
+
|
276 |
+
# split into 4s segments (64000 samples)
|
277 |
+
segment_len = 4 * 16000
|
278 |
+
chunks = x.split(segment_len)
|
279 |
+
enhanced_chunks = []
|
280 |
+
|
281 |
+
for chunk in chunks:
|
282 |
+
if len(chunk) < segment_len:
|
283 |
+
#pad = torch.zeros(segment_len - len(chunk), device=chunk.device)
|
284 |
+
pad = (torch.randn(segment_len - len(chunk), device=chunk.device) * 1e-4)
|
285 |
+
chunk = torch.cat([chunk, pad])
|
286 |
+
chunk = chunk.unsqueeze(0)
|
287 |
+
|
288 |
+
amp, pha, _ = mag_phase_stft(chunk, 400, 100, 400, 0.3)
|
289 |
+
amp2, pha2, _ = model(amp, pha)
|
290 |
+
out = mag_phase_istft(amp2, pha2, 400, 100, 400, 0.3)
|
291 |
+
out = (out / norm).squeeze(0)
|
292 |
+
enhanced_chunks.append(out)
|
293 |
+
|
294 |
+
out = torch.cat(enhanced_chunks)[:len(x)].cpu().numpy() # trim padding
|
295 |
+
|
296 |
+
# back to original rate
|
297 |
+
if orig_sr != 16000:
|
298 |
+
out = librosa.resample(out, orig_sr=16000, target_sr=orig_sr)
|
299 |
+
|
300 |
+
# Normalize
|
301 |
+
peak = np.max(np.abs(out))
|
302 |
+
if peak > 0.05:
|
303 |
+
out = out / peak * 0.85
|
304 |
+
|
305 |
+
# write file
|
306 |
+
sf.write("enhanced.wav", out, orig_sr)
|
307 |
+
|
308 |
+
# spectrograms
|
309 |
+
fig, axs = plt.subplots(1, 2, figsize=(16, 4))
|
310 |
+
|
311 |
+
# noisy
|
312 |
+
D_noisy = librosa.stft(noisy_wav, n_fft=512, hop_length=256)
|
313 |
+
S_noisy = librosa.amplitude_to_db(np.abs(D_noisy), ref=np.max)
|
314 |
+
librosa.display.specshow(S_noisy, sr=orig_sr, hop_length=256, x_axis="time", y_axis="hz", ax=axs[0], vmax=0)
|
315 |
+
axs[0].set_title("Noisy Spectrogram")
|
316 |
+
|
317 |
+
# enhanced
|
318 |
+
D_clean = librosa.stft(out, n_fft=512, hop_length=256)
|
319 |
+
S_clean = librosa.amplitude_to_db(np.abs(D_clean), ref=np.max)
|
320 |
+
librosa.display.specshow(S_clean, sr=orig_sr, hop_length=256, x_axis="time", y_axis="hz", ax=axs[1], vmax=0)
|
321 |
+
#librosa.display.specshow(S_clean, sr=16000, hop_length=512, x_axis="time", y_axis="hz", ax=axs[1], vmax=0)
|
322 |
+
axs[1].set_title("Enhanced Spectrogram")
|
323 |
+
|
324 |
+
plt.tight_layout()
|
325 |
+
|
326 |
+
return "enhanced.wav", fig
|
327 |
+
|
328 |
+
#with gr.Blocks() as demo:
|
329 |
+
# gr.Markdown(ABOUT)
|
330 |
+
# input_audio = gr.Audio(label="Input Audio", type="filepath", interactive=True)
|
331 |
+
# enhance_btn = gr.Button("Enhance")
|
332 |
+
# output_audio = gr.Audio(label="Enhanced Audio", type="filepath")
|
333 |
+
# plot_output = gr.Plot(label="Spectrograms")
|
334 |
+
#
|
335 |
+
# enhance_btn.click(fn=enhance, inputs=input_audio, outputs=[output_audio, plot_output])
|
336 |
+
#
|
337 |
+
#demo.queue().launch()
|
338 |
+
|
339 |
+
with gr.Blocks() as demo:
|
340 |
+
gr.Markdown(ABOUT)
|
341 |
+
input_audio = gr.Audio(label="Input Audio", type="filepath", interactive=True)
|
342 |
+
model_choice = gr.Radio(
|
343 |
+
label="Choose Model (The use of VCTK+DNS is recommended)",
|
344 |
+
choices=["VCTK-Demand", "VCTK+DNS"],
|
345 |
+
value="VCTK-Demand"
|
346 |
+
)
|
347 |
+
enhance_btn = gr.Button("Enhance")
|
348 |
+
output_audio = gr.Audio(label="Enhanced Audio", type="filepath")
|
349 |
+
plot_output = gr.Plot(label="Spectrograms")
|
350 |
+
|
351 |
+
enhance_btn.click(
|
352 |
+
fn=enhance,
|
353 |
+
inputs=[input_audio, model_choice],
|
354 |
+
outputs=[output_audio, plot_output]
|
355 |
+
)
|
356 |
+
gr.Markdown("**Note**: The current models are trained on 16kHz audio. Therefore, any input audio not sampled at 16kHz will be automatically resampled before enhancement.")
|
357 |
+
|
358 |
+
demo.queue().launch()
|