NUM_WORKERS = 5
train_loader = DataLoader(timit_train, shuffle=True, batch_size=BATCH_SIZE,
num_workers=NUM_WORKERS, drop_last=True)
val_loader = DataLoader(timit_val, batch_size=BATCH_SIZE,
num_workers=NUM_WORKERS, drop_last=True)
# some random parameters, does it look sensible?
LR = 1e-3
REDUCE_LR_PATIENCE = 5
EARLY_STOP_PATIENCE = 20
MAX_EPOCHS = 300
# the model here should be constructed in the script accordingly to the passed config (including the model type)
# most of the models accept `sample_rate` parameter for encoders, which is important (default is 16000, override)
#model = DCUNet("DCUNet-20", fix_length_mode="trim", sample_rate=SAMPLE_RATE)
model = ConvTasNet(n_src=1)
checkpoint = ModelCheckpoint(
filename='{epoch:02d}-{val_loss:.2f}',
monitor="val_loss",
mode="min",
save_top_k=5,
verbose=True
)
optimizer = optim.Adam(model.parameters(), lr=LR)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=REDUCE_LR_PATIENCE)
early_stopping = EarlyStopping(monitor='val_loss', patience=EARLY_STOP_PATIENCE)
# Probably we also need to subclass `System`, in order to log the target metrics on the validation set (PESQ/STOI)
system = System(model, optimizer, sisdr_loss_wrapper, train_loader, val_loader, scheduler)
def main(conf):
model_path = os.path.join(conf['exp_dir'], 'best_model.pth')
model = ConvTasNet.from_pretrained(model_path)
# Handle device placement
if conf['use_gpu']:
model.cuda()
model_device = next(model.parameterss()).device
test_set = LibriMix(csv_dir=conf['test_dir'],
task=conf['task'],
sample_rate=conf['sample_rate'],
n_src=conf['train_conf']['data']['n_src'],
segment=None) # Uses all segment length
# Used to reorder sources only
loss_func = PITLossWrapper(pairwise_neg_sisdr, pit_from='pw_mtx')
# Randomly choose the indexes of sentences to save.
eval_save_dir = os.path.join(conf['exp_dir'], conf['out_dir'])
ex_save_dir = os.path.join(eval_save_dir, 'examples/')
if conf['n_save_ex'] == -1:
conf['n_save_ex'] = len(test_set)
save_idx = random.sample(range(len(test_set)), conf['n_save_ex'])
series_list = []
torch.no_grad().__enter__()
for idx in tqdm(range(len(test_set))):
# Forward the network on the mixture.
mix, sources = tensors_to_device(test_set[idx], device=model_device)
est_sources = model(mix.unsqueeze(0))
loss, reordered_sources = loss_func(est_sources,
sources[None],
return_est=True)
mix_np = mix.cpu().data.numpy()
sources_np = sources.squeeze().cpu().data.numpy()
est_sources_np = reordered_sources.squeeze().cpu().data.numpy()
# For each utterance, we get a dictionary with the mixture path,
# the input and output metrics
utt_metrics = get_metrics(mix_np,
sources_np,
est_sources_np,
sample_rate=conf['sample_rate'])
utt_metrics['mix_path'] = test_set.mixture_path
series_list.append(pd.Series(utt_metrics))
# Save some examples in a folder. Wav files and metrics as text.
if idx in save_idx:
local_save_dir = os.path.join(ex_save_dir, 'ex_{}/'.format(idx))
os.makedirs(local_save_dir, exist_ok=True)
sf.write(local_save_dir + "mixture.wav", mix_np,
conf['sample_rate'])
# Loop over the sources and estimates
for src_idx, src in enumerate(sources_np):
sf.write(local_save_dir + "s{}.wav".format(src_idx), src,
conf['sample_rate'])
for src_idx, est_src in enumerate(est_sources_np):
sf.write(local_save_dir + "s{}_estimate.wav".format(src_idx),
est_src, conf['sample_rate'])
# Write local metrics to the example folder.
with open(local_save_dir + 'metrics.json', 'w') as f:
json.dump(utt_metrics, f, indent=0)
# Save all metrics to the experiment folder.
all_metrics_df = pd.DataFrame(series_list)
all_metrics_df.to_csv(os.path.join(eval_save_dir, 'all_metrics.csv'))
# Print and save summary metrics
final_results = {}
for metric_name in compute_metrics:
input_metric_name = 'input_' + metric_name
ldf = all_metrics_df[metric_name] - all_metrics_df[input_metric_name]
final_results[metric_name] = all_metrics_df[metric_name].mean()
final_results[metric_name + '_imp'] = ldf.mean()
print('Overall metrics :')
pprint(final_results)
with open(os.path.join(eval_save_dir, 'final_metrics.json'), 'w') as f:
json.dump(final_results, f, indent=0)
model_dict = torch.load(model_path, map_location='cpu')
publishable = save_publishable(os.path.join(conf['exp_dir'],
'publish_dir'),
model_dict,
metrics=final_results,
train_conf=train_conf)
total_df.sort_values(['SI-SDR', 'PESQ', 'STOI'], inplace=True)
total_df = total_df.round({'SI-SDR': 3, 'PESQ': 3, 'STOI': 3})
print(total_df)
return total_df
models = {
'input': None,
'baseline': RegressionFCNN.from_pretrained('models/baseline_model_v1.pt'),
'vae': VAE.from_pretrained('/jmain01/home/JAD007/txk02/aaa18-txk02/workspace/models/VAE.pt'),
'auto_encoder': VAE.from_pretrained('/jmain01/home/JAD007/txk02/aaa18-txk02/workspace/models/AutoEncoder.pt'),
'waveunet_v1': WaveUNet.from_pretrained('models/waveunet_model_adapt.pt'),
'dcunet_20': DCUNet.from_pretrained('models/dcunet_20_random_v2.pt'),
'dccrn': DCCRNet.from_pretrained('models/dccrn_random_v1.pt'),
'smolnet': SMoLnet.from_pretrained('models/SMoLnet.pt'),
'dprnn': DPRNNTasNet.from_pretrained('models/dprnn_model.pt'),
'conv_tasnet': ConvTasNet.from_pretrained('models/convtasnet_model.pt'),
'dptnet': DPTNet.from_pretrained('models/dptnet_model.pt'),
'demucs': Demucs.from_pretrained('models/Demucs.pt'),
}
def eval_all_and_plot(models, test_set, directory, plot_name):
results_dfs = {}
for model_name, model in models.items():
print(f'Evaluating {model_labels[model_name]}')
csv_path = f'/jmain01/home/JAD007/txk02/aaa18-txk02/DRONE_project/asteroid/notebooks/{directory}/{model_name}.csv'
if os.path.isfile(csv_path):
print('Results already available')
df = pd.read_csv(csv_path)
else:
def main(conf):
model_path = os.path.join(conf["exp_dir"], "best_model.pth")
model = ConvTasNet.from_pretrained(model_path)
model = LambdaOverlapAdd(
nnet=model, # function to apply to each segment.
n_src=2, # number of sources in the output of nnet
window_size=64000, # Size of segmenting window
hop_size=None, # segmentation hop size
window="hanning", # Type of the window (see scipy.signal.get_window
reorder_chunks=False, # Whether to reorder each consecutive segment.
enable_grad=
False, # Set gradient calculation on of off (see torch.set_grad_enabled)
)
# Handle device placement
if conf["use_gpu"]:
model.cuda()
model_device = next(model.parameters()).device
# Evaluation is mode using 'remix' mixture
dataset_kwargs = {
"root_path": Path(conf["train_conf"]["data"]["root_path"]),
"task": conf["train_conf"]["data"]["task"],
"sample_rate": conf["train_conf"]["data"]["sample_rate"],
"num_workers": conf["train_conf"]["training"]["num_workers"],
"mixture": "remix",
}
test_set = DAMPVSEPDataset(split="test", **dataset_kwargs)
# Randomly choose the indexes of sentences to save.
eval_save_dir = os.path.join(conf["exp_dir"], conf["out_dir"])
ex_save_dir = os.path.join(eval_save_dir, "examples/")
if conf["n_save_ex"] == -1:
conf["n_save_ex"] = len(test_set)
save_idx = random.sample(range(len(test_set)), conf["n_save_ex"])
series_list = []
torch.no_grad().__enter__()
for idx in tqdm(range(len(test_set))):
# Forward the network on the mixture.
mix, sources = test_set[idx]
mix = mix.to(model_device)
est_sources = model.forward(mix.unsqueeze(0).unsqueeze(1))
mix_np = mix.squeeze(0).cpu().data.numpy()
sources_np = sources.cpu().data.numpy()
est_sources_np = est_sources.squeeze(0).cpu().data.numpy()
# For each utterance, we get a dictionary with the mixture path,
# the input and output metrics
utt_metrics = get_metrics(
mix_np,
sources_np,
est_sources_np,
sample_rate=conf["sample_rate"],
metrics_list=compute_metrics,
average=False,
)
utt_metrics = split_metric_dict(utt_metrics)
utt_metrics["mix_path"] = test_set.mixture_path
series_list.append(pd.Series(utt_metrics))
# Save some examples in a folder. Wav files and metrics as text.
if idx in save_idx:
local_save_dir = os.path.join(ex_save_dir, "ex_{}/".format(idx))
os.makedirs(local_save_dir, exist_ok=True)
sf.write(local_save_dir + "mixture.wav", mix_np / max(abs(mix_np)),
conf["sample_rate"])
# Loop over the sources and estimates
for src_idx, src in enumerate(sources_np):
sf.write(local_save_dir + "s{}.wav".format(src_idx), src,
conf["sample_rate"])
for src_idx, est_src in enumerate(est_sources_np):
est_src *= np.max(np.abs(mix_np)) / np.max(np.abs(est_src))
sf.write(
local_save_dir + "s{}_estimate.wav".format(src_idx),
est_src,
conf["sample_rate"],
)
# Write local metrics to the example folder.
with open(local_save_dir + "metrics.json", "w") as f:
json.dump(utt_metrics, f, indent=0)
# Save all metrics to the experiment folder.
all_metrics_df = pd.DataFrame(series_list)
all_metrics_df.to_csv(os.path.join(eval_save_dir, "all_metrics.csv"))
# Print and save summary metrics
final_results = {}
for metric_name in compute_metrics:
for s in ["", "_s0", "_s1"]:
input_metric_name = "input_" + f"{metric_name}{s}"
ldf = all_metrics_df[f"{metric_name}{s}"] - all_metrics_df[
input_metric_name]
final_results[f"{metric_name}{s}"] = all_metrics_df[
f"{metric_name}{s}"].mean()
final_results[f"{metric_name}{s}" + "_imp"] = ldf.mean()
print("Overall metrics :")
pprint(final_results)
with open(os.path.join(eval_save_dir, "final_metrics.json"), "w") as f:
json.dump(final_results, f, indent=0)
model_dict = torch.load(model_path, map_location="cpu")
os.makedirs(os.path.join(conf["exp_dir"], "publish_dir"), exist_ok=True)
publishable = save_publishable(
os.path.join(conf["exp_dir"], "publish_dir"),
model_dict,
metrics=final_results,
train_conf=train_conf,
)
def main(conf):
compute_metrics = update_compute_metrics(conf["compute_wer"], COMPUTE_METRICS)
anno_df = pd.read_csv(Path(conf["test_dir"]).parent.parent.parent / "test_annotations.csv")
wer_tracker = (
MockWERTracker() if not conf["compute_wer"] else WERTracker(ASR_MODEL_PATH, anno_df)
)
model_path = os.path.join(conf["exp_dir"], "best_model.pth")
model = ConvTasNet.from_pretrained(model_path)
# Handle device placement
if conf["use_gpu"]:
model.cuda()
model_device = next(model.parameters()).device
test_set = LibriMix(
csv_dir=conf["test_dir"],
task=conf["task"],
sample_rate=conf["sample_rate"],
n_src=conf["train_conf"]["data"]["n_src"],
segment=None,
return_id=True,
) # Uses all segment length
# Used to reorder sources only
loss_func = PITLossWrapper(pairwise_neg_sisdr, pit_from="pw_mtx")
# Randomly choose the indexes of sentences to save.
eval_save_dir = os.path.join(conf["exp_dir"], conf["out_dir"])
ex_save_dir = os.path.join(eval_save_dir, "examples/")
if conf["n_save_ex"] == -1:
conf["n_save_ex"] = len(test_set)
save_idx = random.sample(range(len(test_set)), conf["n_save_ex"])
series_list = []
torch.no_grad().__enter__()
for idx in tqdm(range(len(test_set))):
# Forward the network on the mixture.
mix, sources, ids = test_set[idx]
mix, sources = tensors_to_device([mix, sources], device=model_device)
est_sources = model(mix.unsqueeze(0))
loss, reordered_sources = loss_func(est_sources, sources[None], return_est=True)
mix_np = mix.cpu().data.numpy()
sources_np = sources.cpu().data.numpy()
est_sources_np = reordered_sources.squeeze(0).cpu().data.numpy()
# For each utterance, we get a dictionary with the mixture path,
# the input and output metrics
utt_metrics = get_metrics(
mix_np,
sources_np,
est_sources_np,
sample_rate=conf["sample_rate"],
metrics_list=COMPUTE_METRICS,
)
utt_metrics["mix_path"] = test_set.mixture_path
utt_metrics.update(
**wer_tracker(
mix=mix_np,
clean=sources_np,
estimate=est_sources_np,
wav_id=ids,
sample_rate=conf["sample_rate"],
)
)
series_list.append(pd.Series(utt_metrics))
# Save some examples in a folder. Wav files and metrics as text.
if idx in save_idx:
local_save_dir = os.path.join(ex_save_dir, "ex_{}/".format(idx))
os.makedirs(local_save_dir, exist_ok=True)
sf.write(local_save_dir + "mixture.wav", mix_np, conf["sample_rate"])
# Loop over the sources and estimates
for src_idx, src in enumerate(sources_np):
sf.write(local_save_dir + "s{}.wav".format(src_idx), src, conf["sample_rate"])
for src_idx, est_src in enumerate(est_sources_np):
est_src *= np.max(np.abs(mix_np)) / np.max(np.abs(est_src))
sf.write(
local_save_dir + "s{}_estimate.wav".format(src_idx),
est_src,
conf["sample_rate"],
)
# Write local metrics to the example folder.
with open(local_save_dir + "metrics.json", "w") as f:
json.dump(utt_metrics, f, indent=0)
# Save all metrics to the experiment folder.
all_metrics_df = pd.DataFrame(series_list)
all_metrics_df.to_csv(os.path.join(eval_save_dir, "all_metrics.csv"))
# Print and save summary metrics
final_results = {}
for metric_name in compute_metrics:
input_metric_name = "input_" + metric_name
ldf = all_metrics_df[metric_name] - all_metrics_df[input_metric_name]
final_results[metric_name] = all_metrics_df[metric_name].mean()
final_results[metric_name + "_imp"] = ldf.mean()
print("Overall metrics :")
pprint(final_results)
if conf["compute_wer"]:
print("\nWER report")
wer_card = wer_tracker.final_report_as_markdown()
print(wer_card)
# Save the report
with open(os.path.join(eval_save_dir, "final_wer.md"), "w") as f:
f.write(wer_card)
with open(os.path.join(eval_save_dir, "final_metrics.json"), "w") as f:
json.dump(final_results, f, indent=0)
model_dict = torch.load(model_path, map_location="cpu")
os.makedirs(os.path.join(conf["exp_dir"], "publish_dir"), exist_ok=True)
publishable = save_publishable(
os.path.join(conf["exp_dir"], "publish_dir"),
model_dict,
metrics=final_results,
train_conf=train_conf,
)