Delete test_windowed_streamer.py with huggingface_hub
Browse files- test_windowed_streamer.py +0 -320
test_windowed_streamer.py
DELETED
|
@@ -1,320 +0,0 @@
|
|
| 1 |
-
"""
|
| 2 |
-
Tests MioCodecWindowedStreamer against a mock model matching the real
|
| 3 |
-
MioCodecModel's decoder interface. Verifies:
|
| 4 |
-
- no shape errors through the full windowed pipeline
|
| 5 |
-
- first chunk (no left context) and last chunk (no right lookahead)
|
| 6 |
-
are handled correctly as genuine edge cases, not crashes
|
| 7 |
-
- variable chunk_size (>1, including small values like 2) works
|
| 8 |
-
- boundary discontinuity is LOWER than the causal-only chunk_streamer's
|
| 9 |
-
boundaries (this is the actual point of the redesign — real lookahead
|
| 10 |
-
context should produce smoother joins than causal-only + hard cut)
|
| 11 |
-
- total released audio covers all pushed tokens exactly once (no gaps,
|
| 12 |
-
no duplicated frames)
|
| 13 |
-
|
| 14 |
-
Run: python test_windowed_streamer.py
|
| 15 |
-
"""
|
| 16 |
-
import torch
|
| 17 |
-
import torch.nn as nn
|
| 18 |
-
import torch.nn.functional as F
|
| 19 |
-
|
| 20 |
-
from streaming_istft_head import ISTFTHead as StreamingISTFTHead
|
| 21 |
-
from windowed_streamer import WindowedStreamConfig, MioCodecWindowedStreamer, set_transformer_causal
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
class MockNonCausalTransformer(nn.Module):
|
| 25 |
-
"""Mimics Transformer(causal=False)'s call signature. Bidirectional
|
| 26 |
-
(unlike test_chunk_streamer's MockCausalTransformer, this one lets a
|
| 27 |
-
position's output depend on ALL positions in the input, including
|
| 28 |
-
later ones, via a mean-pool + per-position mix -- simulates real
|
| 29 |
-
bidirectional attention's key property for boundary-smoothness testing:
|
| 30 |
-
every output position is influenced by the whole window, not just the
|
| 31 |
-
causal prefix.)"""
|
| 32 |
-
|
| 33 |
-
def __init__(self, dim):
|
| 34 |
-
super().__init__()
|
| 35 |
-
self.dim = dim
|
| 36 |
-
self.proj = nn.Linear(dim, dim)
|
| 37 |
-
self.mix = nn.Linear(dim, dim)
|
| 38 |
-
self.causal = False # settable by set_transformer_causal, mimicking real Attention.causal
|
| 39 |
-
|
| 40 |
-
@property
|
| 41 |
-
def layers(self):
|
| 42 |
-
# Return fake layer objects so set_transformer_causal can toggle
|
| 43 |
-
# something, and reflect it back into self.causal for forward().
|
| 44 |
-
outer = self
|
| 45 |
-
|
| 46 |
-
class _Layer:
|
| 47 |
-
attention = type("A", (), {"causal": outer.causal})()
|
| 48 |
-
|
| 49 |
-
return [_Layer()]
|
| 50 |
-
|
| 51 |
-
def forward(self, x, condition=None, **kwargs):
|
| 52 |
-
out = self.proj(x)
|
| 53 |
-
# simulate genuine bidirectional mixing: every position sees the
|
| 54 |
-
# mean of the whole window (a crude stand-in for attention seeing
|
| 55 |
-
# future context when causal=False)
|
| 56 |
-
context = self.mix(x.mean(dim=1, keepdim=True))
|
| 57 |
-
out = out + context
|
| 58 |
-
if condition is not None:
|
| 59 |
-
out = out + condition
|
| 60 |
-
return out
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
class MockConfig:
|
| 64 |
-
sample_rate = 24000
|
| 65 |
-
hop_length = 256
|
| 66 |
-
n_fft = 1024
|
| 67 |
-
istft_padding = "same"
|
| 68 |
-
wave_interpolation_mode = "linear"
|
| 69 |
-
wave_upsample_factor = 4
|
| 70 |
-
wave_decoder_dim = 32
|
| 71 |
-
wave_resnet_num_blocks = 2
|
| 72 |
-
wave_resnet_kernel_size = 3
|
| 73 |
-
wave_resnet_num_groups = 8
|
| 74 |
-
wave_resnet_dropout = 0.0
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
class MockResNetBlock(nn.Module):
|
| 78 |
-
"""Original (non-causal, symmetric-padded) ResNetBlock, as used by the
|
| 79 |
-
windowed streamer (unlike CausalResNetStack in causal_resnet.py)."""
|
| 80 |
-
|
| 81 |
-
def __init__(self, channels, kernel_size=3, num_groups=8, dropout=0.0):
|
| 82 |
-
super().__init__()
|
| 83 |
-
padding = (kernel_size - 1) // 2
|
| 84 |
-
self.norm1 = nn.GroupNorm(num_groups, channels, eps=1e-6)
|
| 85 |
-
self.conv1 = nn.Conv1d(channels, channels, kernel_size, padding=padding)
|
| 86 |
-
self.norm2 = nn.GroupNorm(num_groups, channels, eps=1e-6)
|
| 87 |
-
self.conv2 = nn.Conv1d(channels, channels, kernel_size, padding=padding)
|
| 88 |
-
self.dropout = nn.Dropout(dropout)
|
| 89 |
-
|
| 90 |
-
def forward(self, x):
|
| 91 |
-
r = x
|
| 92 |
-
x = F.silu(self.norm1(x))
|
| 93 |
-
x = self.conv1(x)
|
| 94 |
-
x = F.silu(self.norm2(x))
|
| 95 |
-
x = self.dropout(x)
|
| 96 |
-
x = self.conv2(x)
|
| 97 |
-
return x + r
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
class MockResNetStack(nn.Module):
|
| 101 |
-
def __init__(self, channels, num_blocks=2, kernel_size=3, num_groups=8, dropout=0.0):
|
| 102 |
-
super().__init__()
|
| 103 |
-
self.blocks = nn.ModuleList([
|
| 104 |
-
MockResNetBlock(channels, kernel_size, num_groups, dropout) for _ in range(num_blocks)
|
| 105 |
-
])
|
| 106 |
-
|
| 107 |
-
def forward(self, x):
|
| 108 |
-
for b in self.blocks:
|
| 109 |
-
x = b(x)
|
| 110 |
-
return x
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
class MockMioCodecModel(nn.Module):
|
| 114 |
-
def __init__(self):
|
| 115 |
-
super().__init__()
|
| 116 |
-
self.config = MockConfig()
|
| 117 |
-
dim = self.config.wave_decoder_dim
|
| 118 |
-
|
| 119 |
-
self.wave_prenet = MockNonCausalTransformer(dim)
|
| 120 |
-
self.wave_conv_upsample = nn.ConvTranspose1d(
|
| 121 |
-
dim, dim, kernel_size=self.config.wave_upsample_factor, stride=self.config.wave_upsample_factor
|
| 122 |
-
)
|
| 123 |
-
self.wave_prior_net = MockResNetStack(dim, self.config.wave_resnet_num_blocks)
|
| 124 |
-
self.wave_decoder = MockNonCausalTransformer(dim)
|
| 125 |
-
self.wave_post_net = MockResNetStack(dim, self.config.wave_resnet_num_blocks)
|
| 126 |
-
self.wave_upsampler = None
|
| 127 |
-
self.istft_head = StreamingISTFTHead(
|
| 128 |
-
dim=dim, n_fft=self.config.n_fft, hop_length=self.config.hop_length, padding=self.config.istft_padding,
|
| 129 |
-
)
|
| 130 |
-
|
| 131 |
-
def _calculate_target_stft_length(self, audio_length: int) -> int:
|
| 132 |
-
return audio_length // self.config.hop_length
|
| 133 |
-
|
| 134 |
-
def forward_wave(self, content_embedding, global_embedding, stft_length):
|
| 135 |
-
"""
|
| 136 |
-
Mirrors the real MioCodecModel.forward_wave exactly (see model.py),
|
| 137 |
-
so this mock can stand in for the "teacher" (full-utterance,
|
| 138 |
-
unbounded-context) path in alignment tests, not just the windowed
|
| 139 |
-
streaming path.
|
| 140 |
-
"""
|
| 141 |
-
local_latent = self.wave_prenet(content_embedding)
|
| 142 |
-
if self.wave_conv_upsample is not None:
|
| 143 |
-
local_latent = self.wave_conv_upsample(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 144 |
-
local_latent = F.interpolate(
|
| 145 |
-
local_latent.transpose(1, 2), size=stft_length, mode=self.config.wave_interpolation_mode
|
| 146 |
-
).transpose(1, 2)
|
| 147 |
-
local_latent = self.wave_prior_net(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 148 |
-
local_latent = self.wave_decoder(local_latent, condition=global_embedding.unsqueeze(1))
|
| 149 |
-
local_latent = self.wave_post_net(local_latent.transpose(1, 2)).transpose(1, 2)
|
| 150 |
-
if self.wave_upsampler is not None:
|
| 151 |
-
local_latent = self.wave_upsampler(local_latent.transpose(1, 2))
|
| 152 |
-
return self.istft_head(local_latent)
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
def test_first_and_last_chunk_edge_cases():
|
| 156 |
-
"""First chunk must have zero left-context; last chunk must have zero right-lookahead. Not crashes, correct behavior."""
|
| 157 |
-
torch.manual_seed(0)
|
| 158 |
-
model = MockMioCodecModel().eval()
|
| 159 |
-
dim = model.config.wave_decoder_dim
|
| 160 |
-
global_embedding = torch.randn(dim)
|
| 161 |
-
|
| 162 |
-
total_tokens = 45 # not a clean multiple of chunk_size(20)+lookahead(10)
|
| 163 |
-
all_tokens = torch.randn(total_tokens, dim)
|
| 164 |
-
|
| 165 |
-
streamer = MioCodecWindowedStreamer(
|
| 166 |
-
model, global_embedding,
|
| 167 |
-
config=WindowedStreamConfig(chunk_size=20, left_context_tokens=10, right_lookahead_tokens=10),
|
| 168 |
-
)
|
| 169 |
-
|
| 170 |
-
with torch.no_grad():
|
| 171 |
-
outputs = []
|
| 172 |
-
for i in range(total_tokens):
|
| 173 |
-
out = streamer.push_token(all_tokens[i])
|
| 174 |
-
if out is not None:
|
| 175 |
-
outputs.append(out)
|
| 176 |
-
final = streamer.finish()
|
| 177 |
-
if final is not None:
|
| 178 |
-
outputs.append(final)
|
| 179 |
-
|
| 180 |
-
assert len(outputs) >= 1, "Expected at least one chunk released"
|
| 181 |
-
for o in outputs:
|
| 182 |
-
assert torch.isfinite(o).all(), "Non-finite audio in output"
|
| 183 |
-
assert o.numel() > 0, "Empty audio chunk released"
|
| 184 |
-
print(f"PASS: first/last chunk edge cases handled, {len(outputs)} chunks released, all finite.")
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
def test_variable_chunk_sizes():
|
| 188 |
-
"""chunk_size must work for any value > 1, including small values like 2."""
|
| 189 |
-
torch.manual_seed(1)
|
| 190 |
-
model = MockMioCodecModel().eval()
|
| 191 |
-
dim = model.config.wave_decoder_dim
|
| 192 |
-
global_embedding = torch.randn(dim)
|
| 193 |
-
|
| 194 |
-
for chunk_size in [2, 3, 10, 20, 25, 50]:
|
| 195 |
-
streamer = MioCodecWindowedStreamer(
|
| 196 |
-
model, global_embedding,
|
| 197 |
-
config=WindowedStreamConfig(chunk_size=chunk_size, left_context_tokens=10, right_lookahead_tokens=10),
|
| 198 |
-
)
|
| 199 |
-
tokens = torch.randn(chunk_size * 3 + 5, dim)
|
| 200 |
-
with torch.no_grad():
|
| 201 |
-
outputs = [streamer.push_token(tokens[i]) for i in range(tokens.shape[0])]
|
| 202 |
-
outputs = [o for o in outputs if o is not None]
|
| 203 |
-
final = streamer.finish()
|
| 204 |
-
if final is not None:
|
| 205 |
-
outputs.append(final)
|
| 206 |
-
assert len(outputs) >= 1, f"chunk_size={chunk_size}: no output released"
|
| 207 |
-
for o in outputs:
|
| 208 |
-
assert torch.isfinite(o).all(), f"chunk_size={chunk_size}: non-finite audio"
|
| 209 |
-
print(f" chunk_size={chunk_size}: {len(outputs)} chunks, all finite -- OK")
|
| 210 |
-
print("PASS: variable chunk sizes (2 to 50) all work.")
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
def test_min_chunk_size_rejects_1():
|
| 214 |
-
try:
|
| 215 |
-
WindowedStreamConfig(chunk_size=1)
|
| 216 |
-
raise AssertionError("chunk_size=1 should have raised ValueError")
|
| 217 |
-
except ValueError as e:
|
| 218 |
-
print(f"chunk_size=1 correctly rejected: {e}")
|
| 219 |
-
WindowedStreamConfig(chunk_size=2) # minimum valid value
|
| 220 |
-
print("PASS: chunk_size > 1 constraint enforced.")
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
def test_boundary_smoothness_vs_causal_only():
|
| 224 |
-
"""
|
| 225 |
-
Compares boundary jump magnitude between windowed (bidirectional,
|
| 226 |
-
real lookahead) streaming and a zero-lookahead degenerate case
|
| 227 |
-
(same code path, right_lookahead_tokens=0), to isolate the effect of
|
| 228 |
-
the lookahead itself.
|
| 229 |
-
"""
|
| 230 |
-
torch.manual_seed(2)
|
| 231 |
-
model = MockMioCodecModel().eval()
|
| 232 |
-
dim = model.config.wave_decoder_dim
|
| 233 |
-
global_embedding = torch.randn(dim)
|
| 234 |
-
total_tokens = 60
|
| 235 |
-
all_tokens = torch.randn(total_tokens, dim)
|
| 236 |
-
|
| 237 |
-
streamer_windowed = MioCodecWindowedStreamer(
|
| 238 |
-
model, global_embedding,
|
| 239 |
-
config=WindowedStreamConfig(chunk_size=15, left_context_tokens=10, right_lookahead_tokens=10),
|
| 240 |
-
)
|
| 241 |
-
with torch.no_grad():
|
| 242 |
-
chunks_windowed = []
|
| 243 |
-
for i in range(total_tokens):
|
| 244 |
-
out = streamer_windowed.push_token(all_tokens[i])
|
| 245 |
-
if out is not None:
|
| 246 |
-
chunks_windowed.append(out)
|
| 247 |
-
final = streamer_windowed.finish()
|
| 248 |
-
if final is not None:
|
| 249 |
-
chunks_windowed.append(final)
|
| 250 |
-
|
| 251 |
-
streamer_no_lookahead = MioCodecWindowedStreamer(
|
| 252 |
-
model, global_embedding,
|
| 253 |
-
config=WindowedStreamConfig(chunk_size=15, left_context_tokens=10, right_lookahead_tokens=0),
|
| 254 |
-
)
|
| 255 |
-
with torch.no_grad():
|
| 256 |
-
chunks_no_lookahead = []
|
| 257 |
-
for i in range(total_tokens):
|
| 258 |
-
out = streamer_no_lookahead.push_token(all_tokens[i])
|
| 259 |
-
if out is not None:
|
| 260 |
-
chunks_no_lookahead.append(out)
|
| 261 |
-
final = streamer_no_lookahead.finish()
|
| 262 |
-
if final is not None:
|
| 263 |
-
chunks_no_lookahead.append(final)
|
| 264 |
-
|
| 265 |
-
def boundary_jump_stats(chunks):
|
| 266 |
-
audio = torch.cat(chunks, dim=0)
|
| 267 |
-
diffs = (audio[1:] - audio[:-1]).abs()
|
| 268 |
-
typical = diffs.median().item()
|
| 269 |
-
boundary_indices = []
|
| 270 |
-
pos = 0
|
| 271 |
-
for c in chunks[:-1]:
|
| 272 |
-
pos += c.shape[0]
|
| 273 |
-
if 0 < pos < diffs.shape[0]:
|
| 274 |
-
boundary_indices.append(pos)
|
| 275 |
-
jumps = [diffs[i].item() for i in boundary_indices]
|
| 276 |
-
ratios = [j / (typical + 1e-8) for j in jumps]
|
| 277 |
-
return ratios
|
| 278 |
-
|
| 279 |
-
ratios_with_lookahead = boundary_jump_stats(chunks_windowed)
|
| 280 |
-
ratios_without_lookahead = boundary_jump_stats(chunks_no_lookahead)
|
| 281 |
-
|
| 282 |
-
print(f" boundary jump ratios WITH lookahead: {[f'{r:.2f}' for r in ratios_with_lookahead]}")
|
| 283 |
-
print(f" boundary jump ratios WITHOUT lookahead: {[f'{r:.2f}' for r in ratios_without_lookahead]}")
|
| 284 |
-
print(" (lower ratio = smoother join; this is a structural sanity check on random")
|
| 285 |
-
print(" weights, not a quality claim -- real weights needed for actual quality eval)")
|
| 286 |
-
print("PASS: both configurations run without error; compare ratios above.")
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
def test_exact_coverage_no_gaps_no_dupes():
|
| 290 |
-
"""Total tokens released as audio should account for every pushed token exactly once."""
|
| 291 |
-
torch.manual_seed(3)
|
| 292 |
-
model = MockMioCodecModel().eval()
|
| 293 |
-
dim = model.config.wave_decoder_dim
|
| 294 |
-
global_embedding = torch.randn(dim)
|
| 295 |
-
|
| 296 |
-
total_tokens = 77
|
| 297 |
-
all_tokens = torch.randn(total_tokens, dim)
|
| 298 |
-
streamer = MioCodecWindowedStreamer(
|
| 299 |
-
model, global_embedding,
|
| 300 |
-
config=WindowedStreamConfig(chunk_size=20, left_context_tokens=10, right_lookahead_tokens=10),
|
| 301 |
-
)
|
| 302 |
-
with torch.no_grad():
|
| 303 |
-
for i in range(total_tokens):
|
| 304 |
-
streamer.push_token(all_tokens[i])
|
| 305 |
-
streamer.finish()
|
| 306 |
-
|
| 307 |
-
assert streamer._released_count == total_tokens, (
|
| 308 |
-
f"released_count={streamer._released_count} != total_tokens={total_tokens} "
|
| 309 |
-
f"-- some tokens were never released as audio, or released twice."
|
| 310 |
-
)
|
| 311 |
-
print(f"PASS: exactly {total_tokens} tokens accounted for, no gaps or duplicates in release bookkeeping.")
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
if __name__ == "__main__":
|
| 315 |
-
test_first_and_last_chunk_edge_cases()
|
| 316 |
-
test_variable_chunk_sizes()
|
| 317 |
-
test_min_chunk_size_rejects_1()
|
| 318 |
-
test_boundary_smoothness_vs_causal_only()
|
| 319 |
-
test_exact_coverage_no_gaps_no_dupes()
|
| 320 |
-
print("\nALL WINDOWED STREAMER TESTS PASSED")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|