def main():
# Settings
use_cuda = True
log_dir = 'identification_model_ckpt'
cp_num = 50 # Which checkpoint to use?
test_frames = 200
# Model params
resnet_version = 'resnet18'
embedding_size = 128
n_classes = 200
# Load model from checkpoint
model = load_model(use_cuda, log_dir, cp_num, embedding_size, n_classes,
resnet_version)
# Get the dataframe for enroll DB
featrues_path = '/cas/DeepLearn/elperu/tmp/speech_datasets/LibriSpeech/train_test_split/test'
DB_all = read_feats_structure(featrues_path)
# Where to save embeddings
embedding_dir = '/cas/DeepLearn/elperu/tmp/speech_datasets/LibriSpeech/embd_identification'
# Perform the enrollment and save the results
enroll_per_spk(use_cuda, test_frames, model, DB_all, embedding_dir)
def load_dataset(M):
train_DB = read_feats_structure(c.TRAIN_FEAT_DIR)
print(f'\nTraining set {len(train_DB)}')
file_loader = read_MFB # numpy array:(n_frames, n_dims)
transform = transforms.Compose([
TruncatedInputfromMFB(), # numpy array:(1, n_frames, n_dims)
ToTensorInput() # torch tensor:(1, n_dims, n_frames)
])
transform_T = ToTensorDevInput()
speaker_list = sorted(set(
train_DB['speaker_id'])) # len(speaker_list) == n_speakers
spk_to_idx = {spk: i for i, spk in enumerate(speaker_list)}
train_dataset = contrastiveDataset(DB=train_DB,
loader=file_loader,
transform=transform,
spk_to_idx=spk_to_idx,
spk_list=speaker_list,
M=M)
n_classes = len(speaker_list) # How many speakers? 240
print('\nNumber of classes (speakers):\n{}\n'.format(n_classes))
return train_dataset, n_classes
def split_enroll_and_test(dataroot_dir):
DB_all = read_feats_structure(dataroot_dir)
enroll_DB = pd.DataFrame()
test_DB = pd.DataFrame()
enroll_DB = DB_all[DB_all['filename'].str.contains('enroll.p')]
test_DB = DB_all[DB_all['filename'].str.contains('test.p')]
# Reset the index
enroll_DB = enroll_DB.reset_index(drop=True)
test_DB = test_DB.reset_index(drop=True)
return enroll_DB, test_DB
def split_train_dev(train_feat_dir, valid_ratio):
train_valid_DB = read_feats_structure(train_feat_dir)
total_len = len(train_valid_DB) # 148642
valid_len = int(total_len * valid_ratio/100.)
train_len = total_len - valid_len
shuffled_train_valid_DB = train_valid_DB.sample(frac=1).reset_index(drop=True)
# Split the DB into train and valid set
train_DB = shuffled_train_valid_DB.iloc[:train_len]
valid_DB = shuffled_train_valid_DB.iloc[train_len:]
# Reset the index
train_DB = train_DB.reset_index(drop=True)
valid_DB = valid_DB.reset_index(drop=True)
print('\nTraining set %d utts (%0.1f%%)' %(train_len, (train_len/total_len)*100))
print('Validation set %d utts (%0.1f%%)' %(valid_len, (valid_len/total_len)*100))
print('Total %d utts' %(total_len))
return train_DB, valid_DB