right_context = int(os.environ["RIGHT_CONTEXT"])
time_dim = (left_context + right_context + 1)
# Set up training parameters
on_gpu = torch.cuda.is_available()
# Fix random seed for debugging
torch.manual_seed(1)
if on_gpu:
torch.cuda.manual_seed_all(1)
# Set up the model and associated checkpointing directory
model = setup_model(e2e=True)
if on_gpu:
model.cuda()
ckpt_path = best_ckpt_path(e2e=True)
# Load checkpoint
ckpt = torch.load(ckpt_path, map_location=lambda storage,loc: storage)
model.load_state_dict(ckpt["state_dict"])
model.eval()
# Set up data files
scp_dir = os.path.join(os.environ["%s_FEATS" % source_class.upper()], data_dir)
# scp_name = os.path.join(scp_dir, "feats.scp")
scp_file = os.path.join(scp_dir, "feats-norm.scp")
loader_kwargs = {"num_workers": 1, "pin_memory": True} if on_gpu else {}
dataset = KaldiEvalDataset(scp_file,
shuffle_utts=False)
torch.manual_seed(1)
if on_gpu:
torch.cuda.manual_seed_all(1)
# Set up the model and associated checkpointing directory
model = setup_model(denoiser=True)
print(model, flush=True)
# Count number of trainable parameters
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print("Model has %d trainable parameters" % params, flush=True)
if on_gpu:
model.cuda()
ckpt_path = best_ckpt_path(denoiser=True)
# Set up data files
decoder_classes = []
for res_str in os.environ["DECODER_CLASSES_DELIM"].split("_"):
if len(res_str) > 0:
decoder_classes.append(res_str)
training_scps = []
for decoder_class in decoder_classes:
training_scp_dir = os.path.join(
os.environ["%s_FEATS" % decoder_class.upper()], "train")
# training_scp_name = os.path.join(training_scp_dir, "feats.scp")
training_scp_name = os.path.join(training_scp_dir, "feats-norm.scp")
training_scps.append(training_scp_name)
torch.manual_seed(1)
if on_gpu:
torch.cuda.manual_seed_all(1)
# Set up the model and associated checkpointing directory
model = setup_model(phone=True)
print(model, flush=True)
# Count number of trainable parameters
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print("Model has %d trainable parameters" % params, flush=True)
if on_gpu:
model.cuda()
ckpt_path = best_ckpt_path(phone=True)
# Set up data files
decoder_classes = []
for res_str in os.environ["DECODER_CLASSES_DELIM"].split("_"):
if len(res_str) > 0:
decoder_classes.append(res_str)
train_feat_scps = []
for decoder_class in decoder_classes:
train_scp_dir = os.path.join(
os.environ["%s_FEATS" % decoder_class.upper()], "train")
# train_scp_name = os.path.join(train_scp_dir, "feats.scp")
train_scp_name = os.path.join(train_scp_dir, "feats-norm.scp")
train_feat_scps.append(train_scp_name)
torch.manual_seed(1)
if on_gpu:
torch.cuda.manual_seed_all(1)
# Set up the model and associated checkpointing directory
model = setup_model()
print(model, flush=True)
# Count number of trainable parameters
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print("Model has %d trainable parameters" % params, flush=True)
if on_gpu:
model.cuda()
ckpt_path = best_ckpt_path()
# Set up data files
decoder_classes = []
for res_str in os.environ["DECODER_CLASSES_DELIM"].split("_"):
if len(res_str) > 0:
decoder_classes.append(res_str)
training_scps = []
for decoder_class in decoder_classes:
training_scp_dir = os.path.join(
os.environ["%s_FEATS" % decoder_class.upper()], "train")
# training_scp_name = os.path.join(training_scp_dir, "feats.scp")
training_scp_name = os.path.join(training_scp_dir, "feats-norm.scp")
training_scps.append(training_scp_name)