k += 1
return p_means, p_covariances, q_means, q_covariances, num_phones_mask
spk = 0
phone = 2
# Get all the p/q vects
pkl_file = '/home/alta/BLTSpeaking/exp-vr313/data/mfcc13/GKTS4-D3/grader/BLXXXgrd02/BLXXXgrd02.pkl'
pkl = pickle.load(open(pkl_file, "rb"))
print("loaded pkl")
# get the phones
phones = get_phones()
max_len_frames = 4000
p_vects, q_vects, p_mask, q_mask, mask = get_vects(pkl, phones, max_len_frames)
# Convert to pytorch tensors
p_vects = torch.from_numpy(p_vects).float()
q_vects = torch.from_numpy(q_vects).float()
p_mask = torch.from_numpy(p_mask).float()
q_mask = torch.from_numpy(q_mask).float()
mask = torch.from_numpy(mask).float()
# Apply torch operations
# Get p/q_lengths
p_lengths = torch.sum(p_mask[:, :, :, 0].squeeze(),
dim=2).unsqueeze(dim=2).repeat(1, 1, 13)
q_lengths = torch.sum(q_mask[:, :, :, 0].squeeze(),
dim=2).unsqueeze(dim=2).repeat(1, 1, 13)
checkpoint = args.checkpoint
# Save the command run
if not os.path.isdir('CMDs'):
os.mkdir('CMDs')
with open('CMDs/training_spectral_attack_mse.cmd', 'a') as f:
f.write(' '.join(sys.argv) + '\n')
pkl = pickle.load(open(pkl_file, "rb"))
print("Loaded pkl")
# Get the phones
phones = get_phones()
# Get the batched tensors
X1, X2, M1, M2 = get_vects(pkl, phones, N, F)
# Get the output labels
y = (pkl['score'])
# Convert to tensors
X1 = torch.from_numpy(X1).float()
X2 = torch.from_numpy(X2).float()
M1 = torch.from_numpy(M1).float()
M2 = torch.from_numpy(M2).float()
y = torch.FloatTensor(y)
# Split into training and validation sets
validation_size = 50
X1_train = X1[validation_size:]
X1_val = X1[:validation_size]