src_train_mean = train_stats['src_train_mean'][:]
src_train_std = train_stats['src_train_std'][:]
trg_train_mean = train_stats['trg_train_mean'][:]
trg_train_std = train_stats['trg_train_std'][:]
train_stats.close()
# Load test data
print('Loading test data...', end='')
with h5py.File('data/test_datatable.h5', 'r') as test_datatable:
test_data = test_datatable['test_data'][:, :]
test_datatable.close()
src_test_data = test_data[:, 0:40] # Source data
src_test_data = utils.reshape_lstm(src_test_data, tsteps, data_dim)
src_test_data = (src_test_data - src_train_mean) / src_train_std
trg_test_data = test_data[:, 43:83] # Target data
print('done')
################
# Predict data #
################
print('Predicting')
prediction_test = model.predict(src_test_data, batch_size=batch_size)
prediction_test = prediction_test.reshape(-1, data_dim)
# De-normalize predicted output
prediction_test = (prediction_test * trg_train_std) + trg_train_mean
# Remove means and normalize
src_train_mean = np.mean(src_train_data, axis=0)
src_train_std = np.std(src_train_data, axis=0)
src_train_data = (src_train_data - src_train_mean) / src_train_std
src_valid_data = (src_valid_data - src_train_mean) / src_train_std
src_test_data = (src_test_data - src_train_mean) / src_train_std
trg_train_mean = np.mean(trg_train_data, axis=0)
trg_train_std = np.std(trg_train_data, axis=0)
trg_train_data = (trg_train_data - trg_train_mean) / trg_train_std
trg_valid_data = (trg_valid_data - trg_train_mean) / trg_train_std
# Zero-pad and reshape data
src_train_data = utils.reshape_lstm(src_train_data, tsteps, data_dim)
src_valid_data = utils.reshape_lstm(src_valid_data, tsteps, data_dim)
src_test_data = utils.reshape_lstm(src_test_data, tsteps, data_dim)
trg_train_data = utils.reshape_lstm(trg_train_data, tsteps, data_dim)
trg_valid_data = utils.reshape_lstm(trg_valid_data, tsteps, data_dim)
# Save training statistics
with h5py.File('models/mcp_train_stats.h5', 'w') as f:
h5_src_train_mean = f.create_dataset("src_train_mean", data=src_train_mean)
h5_src_train_std = f.create_dataset("src_train_std", data=src_train_std)
h5_trg_train_mean = f.create_dataset("trg_train_mean", data=trg_train_mean)
h5_trg_train_std = f.create_dataset("trg_train_std", data=trg_train_std)
f.close()
# Remove means and normalize
src_train_mean = np.mean(src_train_data, axis=0)
src_train_std = np.std(src_train_data, axis=0)
src_train_data = (src_train_data - src_train_mean) / src_train_std
src_valid_data = (src_valid_data - src_train_mean) / src_train_std
src_test_data = (src_test_data - src_train_mean) / src_train_std
trg_train_mean = np.mean(trg_train_data, axis=0)
trg_train_std = np.std(trg_train_data, axis=0)
trg_train_data = (trg_train_data - trg_train_mean) / trg_train_std
trg_valid_data = (trg_valid_data - trg_train_mean) / trg_train_std
# Zero-pad and reshape data
src_train_data = utils.reshape_lstm(src_train_data, tsteps, data_dim)
src_valid_data = utils.reshape_lstm(src_valid_data, tsteps, data_dim)
src_test_data = utils.reshape_lstm(src_test_data, tsteps, data_dim)
trg_train_data = utils.reshape_lstm(trg_train_data, tsteps, data_dim)
trg_valid_data = utils.reshape_lstm(trg_valid_data, tsteps, data_dim)
# Save training statistics
with h5py.File('models/mcp_train_stats.h5', 'w') as f:
h5_src_train_mean = f.create_dataset("src_train_mean", data=src_train_mean)
h5_src_train_std = f.create_dataset("src_train_std", data=src_train_std)
h5_trg_train_mean = f.create_dataset("trg_train_mean", data=trg_train_mean)
h5_trg_train_std = f.create_dataset("trg_train_std", data=trg_train_std)
f.close()
mcp_params = parse_file(40,
'data/test/vocoded/SF1/' + basename + '.mcp.dat')
lf0_params = parse_file(1,
'data/test/vocoded/SF1/' + basename + '.lf0.i.dat')
mvf_params = parse_file(1,
'data/test/vocoded/SF1/' + basename + '.vf.i.dat')
# Compute U/V flags
assert mvf_params.shape == lf0_params.shape
uv_flags = np.empty(mvf_params.shape)
for index, vf in enumerate(uv_flags):
uv_flags[index] = 1 - utils.kronecker_delta(mvf_params[index])
# Prepare data for prediction
mcp_params = (mcp_params - src_mcp_mean) / src_mcp_std
mcp_params = utils.reshape_lstm(mcp_params, mcp_tsteps, mcp_data_dim)
lf0_params = (lf0_params - src_lf0_mean) / src_lf0_std
lf0_params = utils.reshape_lstm(np.column_stack((lf0_params, uv_flags)),
lf0_tsteps, lf0_data_dim)
mvf_params = (mvf_params - src_mvf_mean) / src_mvf_std
mvf_params = utils.apply_context(mvf_params, context_size) # Apply context
######################
# Predict parameters #
######################
mvf_prediction = mvf_model.predict(np.column_stack((mvf_params, uv_flags)))
mvf_prediction[:, 0] = (mvf_prediction[:, 0] * trg_mvf_std) + trg_mvf_mean
lf0_prediction = lf0_model.predict(lf0_params, batch_size=lf0_batch_size)
trg_train_mean = train_stats['trg_train_mean'].value
trg_train_std = train_stats['trg_train_std'].value
train_stats.close()
# Load test data
print('Loading test data...', end='')
with h5py.File('data/test_datatable.h5', 'r') as test_datatable:
test_data = test_datatable['test_data'][:, :]
test_datatable.close()
src_test_frames = np.column_stack((test_data[:, 40], test_data[:, 42]))
src_test_frames[:,
0] = (src_test_frames[:, 0] - src_train_mean) / src_train_std
src_test_frames = utils.reshape_lstm(src_test_frames, tsteps, data_dim)
# Zero-pad and reshape target test data
trg_test_frames = utils.reshape_lstm(
np.column_stack((test_data[:, 83], test_data[:, 85])), tsteps,
data_dim).reshape(-1, 2)
print('done')
################
# Predict data #
################
print('Predicting')
prediction_test = model.predict(src_test_frames, batch_size=batch_size)
prediction_test = prediction_test.reshape(-1, 2)
src_train_mean = train_stats['src_train_mean'][:]
src_train_std = train_stats['src_train_std'][:]
trg_train_mean = train_stats['trg_train_mean'][:]
trg_train_std = train_stats['trg_train_std'][:]
train_stats.close()
# Load test data
print('Loading test data...', end='')
with h5py.File('data/test_datatable.h5', 'r') as test_datatable:
test_data = test_datatable['test_data'][:, :]
test_datatable.close()
src_test_data = test_data[:, 0:40] # Source data
src_test_data = utils.reshape_lstm(src_test_data, tsteps, data_dim)
src_test_data = (src_test_data - src_train_mean) / src_train_std
trg_test_data = test_data[:, 43:83] # Target data
print('done')
################
# Predict data #
################
print('Predicting')
prediction_test = model.predict(src_test_data, batch_size=batch_size)
prediction_test = prediction_test.reshape(-1, data_dim)
# De-normalize predicted output
prediction_test = (prediction_test * trg_train_std) + trg_train_mean