Spaces:
Build error
Build error
Update sonic.py
Browse files
sonic.py
CHANGED
|
@@ -1,9 +1,7 @@
|
|
| 1 |
-
import os, math
|
| 2 |
-
import torch
|
| 3 |
from PIL import Image
|
| 4 |
from omegaconf import OmegaConf
|
| 5 |
from tqdm import tqdm
|
| 6 |
-
import cv2
|
| 7 |
|
| 8 |
from diffusers import AutoencoderKLTemporalDecoder
|
| 9 |
from diffusers.schedulers import EulerDiscreteScheduler
|
|
@@ -40,15 +38,15 @@ def test(
|
|
| 40 |
face_mask = batch["face_mask"]
|
| 41 |
image_embeds = image_encoder(clip_img).image_embeds
|
| 42 |
|
| 43 |
-
audio_feature = batch["audio_feature"]
|
| 44 |
-
audio_len = int(batch["audio_len"])
|
| 45 |
step = int(config.step)
|
| 46 |
|
| 47 |
-
# ---
|
| 48 |
if audio_len < step:
|
| 49 |
step = max(1, audio_len)
|
| 50 |
|
| 51 |
-
window = 16000
|
| 52 |
audio_prompts, last_prompts = [], []
|
| 53 |
|
| 54 |
# --- window 단위 Whisper 인코딩 --------------------------------------
|
|
@@ -56,26 +54,28 @@ def test(
|
|
| 56 |
chunk = audio_feature[:, :, i : i + window]
|
| 57 |
|
| 58 |
prompt_layers = wav_enc.encoder(chunk, output_hidden_states=True).hidden_states
|
| 59 |
-
last_hidden = wav_enc.encoder(chunk).last_hidden_state.unsqueeze(-2)
|
| 60 |
|
| 61 |
-
audio_prompts.append(torch.stack(prompt_layers, dim=2))
|
| 62 |
-
last_prompts.append(last_hidden)
|
| 63 |
|
| 64 |
if len(audio_prompts) == 0:
|
| 65 |
raise ValueError("[ERROR] No speech recognised in the provided audio.")
|
| 66 |
|
| 67 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
| 68 |
-
last_prompts = torch.cat(last_prompts,
|
| 69 |
|
| 70 |
# padding 규칙
|
| 71 |
audio_prompts = torch.cat(
|
| 72 |
-
[torch.zeros_like(audio_prompts[:, :4]),
|
|
|
|
| 73 |
torch.zeros_like(audio_prompts[:, :6])], dim=1)
|
| 74 |
last_prompts = torch.cat(
|
| 75 |
-
[torch.zeros_like(last_prompts[:, :24]),
|
|
|
|
| 76 |
torch.zeros_like(last_prompts[:, :26])], dim=1)
|
| 77 |
|
| 78 |
-
# ---
|
| 79 |
total_tokens = audio_prompts.shape[1]
|
| 80 |
num_chunks = max(1, math.ceil(total_tokens / (2 * step)))
|
| 81 |
|
|
@@ -84,19 +84,25 @@ def test(
|
|
| 84 |
for i in tqdm(range(num_chunks)):
|
| 85 |
start = i * 2 * step
|
| 86 |
|
| 87 |
-
cond_clip = audio_prompts[:, start : start + 10]
|
| 88 |
-
if cond_clip.shape[
|
| 89 |
-
pad = torch.zeros_like(cond_clip[:, :
|
| 90 |
-
cond_clip = torch.cat([cond_clip, pad], dim=
|
| 91 |
|
| 92 |
-
|
| 93 |
-
bucket_clip =
|
| 94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
|
|
|
| 96 |
|
| 97 |
ref_list.append(ref_img[0])
|
| 98 |
-
audio_list.append(audio_pe(cond_clip).squeeze(0)[0])
|
| 99 |
-
uncond_list.append(audio_pe(torch.zeros_like(cond_clip)).squeeze(0)[0])
|
| 100 |
motion_buckets.append(motion[0])
|
| 101 |
|
| 102 |
# ----------------------------------------------------------------------
|
|
@@ -125,7 +131,7 @@ def test(
|
|
| 125 |
|
| 126 |
|
| 127 |
# ------------------------------------------------------------------
|
| 128 |
-
#
|
| 129 |
# ------------------------------------------------------------------
|
| 130 |
class Sonic:
|
| 131 |
config_file = os.path.join(BASE_DIR, "config/inference/sonic.yaml")
|
|
@@ -144,18 +150,18 @@ class Sonic:
|
|
| 144 |
def _load_models(self, cfg):
|
| 145 |
dtype = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16}[cfg.weight_dtype]
|
| 146 |
|
| 147 |
-
vae
|
| 148 |
sched = EulerDiscreteScheduler.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="scheduler")
|
| 149 |
-
|
| 150 |
-
unet
|
| 151 |
add_ip_adapters(unet, [32], [cfg.ip_audio_scale])
|
| 152 |
|
| 153 |
a2t = AudioProjModel(10, 5, 384, 1024, 1024, 32).to(self.device)
|
| 154 |
a2b = Audio2bucketModel(50, 1, 384, 1024, 1024, 1, 2).to(self.device)
|
| 155 |
|
| 156 |
-
unet.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.unet_checkpoint_path),
|
| 157 |
-
a2t.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2token_checkpoint_path),
|
| 158 |
-
a2b.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2bucket_checkpoint_path),
|
| 159 |
|
| 160 |
whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny")).to(self.device).eval()
|
| 161 |
whisper.requires_grad_(False)
|
|
@@ -166,11 +172,11 @@ class Sonic:
|
|
| 166 |
self.rife = RIFEModel(device=self.device)
|
| 167 |
self.rife.load_model(os.path.join(BASE_DIR, "checkpoints/RIFE/"))
|
| 168 |
|
| 169 |
-
for m in (
|
| 170 |
m.to(dtype)
|
| 171 |
|
| 172 |
-
self.pipe = SonicPipeline(unet=unet, image_encoder=
|
| 173 |
-
self.image_encoder =
|
| 174 |
self.audio2token = a2t
|
| 175 |
self.audio2bucket = a2b
|
| 176 |
self.whisper = whisper
|
|
|
|
| 1 |
+
import os, math, torch, cv2
|
|
|
|
| 2 |
from PIL import Image
|
| 3 |
from omegaconf import OmegaConf
|
| 4 |
from tqdm import tqdm
|
|
|
|
| 5 |
|
| 6 |
from diffusers import AutoencoderKLTemporalDecoder
|
| 7 |
from diffusers.schedulers import EulerDiscreteScheduler
|
|
|
|
| 38 |
face_mask = batch["face_mask"]
|
| 39 |
image_embeds = image_encoder(clip_img).image_embeds
|
| 40 |
|
| 41 |
+
audio_feature = batch["audio_feature"] # (1,80,T)
|
| 42 |
+
audio_len = int(batch["audio_len"]) # Python int
|
| 43 |
step = int(config.step)
|
| 44 |
|
| 45 |
+
# --- step 보정 (최소 1) -----------------------------------------------
|
| 46 |
if audio_len < step:
|
| 47 |
step = max(1, audio_len)
|
| 48 |
|
| 49 |
+
window = 16000 # 1초 chunk
|
| 50 |
audio_prompts, last_prompts = [], []
|
| 51 |
|
| 52 |
# --- window 단위 Whisper 인코딩 --------------------------------------
|
|
|
|
| 54 |
chunk = audio_feature[:, :, i : i + window]
|
| 55 |
|
| 56 |
prompt_layers = wav_enc.encoder(chunk, output_hidden_states=True).hidden_states
|
| 57 |
+
last_hidden = wav_enc.encoder(chunk).last_hidden_state.unsqueeze(-2) # (1,t,1,384)
|
| 58 |
|
| 59 |
+
audio_prompts.append(torch.stack(prompt_layers, dim=2)) # (1,L,12,80)
|
| 60 |
+
last_prompts.append(last_hidden) # (1,L,1,384)
|
| 61 |
|
| 62 |
if len(audio_prompts) == 0:
|
| 63 |
raise ValueError("[ERROR] No speech recognised in the provided audio.")
|
| 64 |
|
| 65 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
| 66 |
+
last_prompts = torch.cat(last_prompts, dim=1)
|
| 67 |
|
| 68 |
# padding 규칙
|
| 69 |
audio_prompts = torch.cat(
|
| 70 |
+
[torch.zeros_like(audio_prompts[:, :4]),
|
| 71 |
+
audio_prompts,
|
| 72 |
torch.zeros_like(audio_prompts[:, :6])], dim=1)
|
| 73 |
last_prompts = torch.cat(
|
| 74 |
+
[torch.zeros_like(last_prompts[:, :24]),
|
| 75 |
+
last_prompts,
|
| 76 |
torch.zeros_like(last_prompts[:, :26])], dim=1)
|
| 77 |
|
| 78 |
+
# --- 반드시 ≥1 chunk --------------------------------------------------
|
| 79 |
total_tokens = audio_prompts.shape[1]
|
| 80 |
num_chunks = max(1, math.ceil(total_tokens / (2 * step)))
|
| 81 |
|
|
|
|
| 84 |
for i in tqdm(range(num_chunks)):
|
| 85 |
start = i * 2 * step
|
| 86 |
|
| 87 |
+
cond_clip = audio_prompts[:, start : start + 10] # (1,10,12,80)
|
| 88 |
+
if cond_clip.shape[1] < 10: # 짧으면 패딩
|
| 89 |
+
pad = torch.zeros_like(cond_clip[:, : 10 - cond_clip.shape[1]])
|
| 90 |
+
cond_clip = torch.cat([cond_clip, pad], dim=1)
|
| 91 |
|
| 92 |
+
# ------------------ (★) bucket_clip 차원 맞춤 -------------------
|
| 93 |
+
bucket_clip = last_prompts[:, start : start + 50] # (1,50,1,384)
|
| 94 |
+
if bucket_clip.shape[1] < 50: # 짧으면 패딩
|
| 95 |
+
pad = torch.zeros_like(bucket_clip[:, : 50 - bucket_clip.shape[1]])
|
| 96 |
+
bucket_clip = torch.cat([bucket_clip, pad], dim=1)
|
| 97 |
+
|
| 98 |
+
bucket_clip = bucket_clip.unsqueeze(1) # → (1,1,50,1,384) ✔ 5-D
|
| 99 |
+
# -----------------------------------------------------------------
|
| 100 |
|
| 101 |
+
motion = audio2bucket(bucket_clip, image_embeds) * 16 + 16
|
| 102 |
|
| 103 |
ref_list.append(ref_img[0])
|
| 104 |
+
audio_list.append(audio_pe(cond_clip.unsqueeze(1)).squeeze(0)[0]) # (10,···)→ unsqueeze 후 4-D
|
| 105 |
+
uncond_list.append(audio_pe(torch.zeros_like(cond_clip).unsqueeze(1)).squeeze(0)[0])
|
| 106 |
motion_buckets.append(motion[0])
|
| 107 |
|
| 108 |
# ----------------------------------------------------------------------
|
|
|
|
| 131 |
|
| 132 |
|
| 133 |
# ------------------------------------------------------------------
|
| 134 |
+
# Sonic 클래스
|
| 135 |
# ------------------------------------------------------------------
|
| 136 |
class Sonic:
|
| 137 |
config_file = os.path.join(BASE_DIR, "config/inference/sonic.yaml")
|
|
|
|
| 150 |
def _load_models(self, cfg):
|
| 151 |
dtype = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16}[cfg.weight_dtype]
|
| 152 |
|
| 153 |
+
vae = AutoencoderKLTemporalDecoder.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="vae", variant="fp16")
|
| 154 |
sched = EulerDiscreteScheduler.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="scheduler")
|
| 155 |
+
imgE = CLIPVisionModelWithProjection.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="image_encoder", variant="fp16")
|
| 156 |
+
unet = UNetSpatioTemporalConditionModel.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="unet", variant="fp16")
|
| 157 |
add_ip_adapters(unet, [32], [cfg.ip_audio_scale])
|
| 158 |
|
| 159 |
a2t = AudioProjModel(10, 5, 384, 1024, 1024, 32).to(self.device)
|
| 160 |
a2b = Audio2bucketModel(50, 1, 384, 1024, 1024, 1, 2).to(self.device)
|
| 161 |
|
| 162 |
+
unet.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.unet_checkpoint_path), map_location="cpu"))
|
| 163 |
+
a2t.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2token_checkpoint_path), map_location="cpu"))
|
| 164 |
+
a2b.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2bucket_checkpoint_path), map_location="cpu"))
|
| 165 |
|
| 166 |
whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny")).to(self.device).eval()
|
| 167 |
whisper.requires_grad_(False)
|
|
|
|
| 172 |
self.rife = RIFEModel(device=self.device)
|
| 173 |
self.rife.load_model(os.path.join(BASE_DIR, "checkpoints/RIFE/"))
|
| 174 |
|
| 175 |
+
for m in (imgE, vae, unet):
|
| 176 |
m.to(dtype)
|
| 177 |
|
| 178 |
+
self.pipe = SonicPipeline(unet=unet, image_encoder=imgE, vae=vae, scheduler=sched).to(device=self.device, dtype=dtype)
|
| 179 |
+
self.image_encoder = imgE
|
| 180 |
self.audio2token = a2t
|
| 181 |
self.audio2bucket = a2b
|
| 182 |
self.whisper = whisper
|