def __init__(self, options, inp_dim):
super(MOCKINGJAY, self).__init__()
all_states = torch.load(options["ckpt_file"], map_location='cpu')
self.config = all_states['Settings']['Config']
self.no_grad = bool(strtobool(options["no_grad"]))
# increase dropout
if str(options['dropout']) != 'default':
self.config['mockingjay']['hidden_dropout_prob'] = float(
options['dropout'])
self.config['mockingjay']['attention_probs_dropout_prob'] = float(
options['dropout'])
# Model Config
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
# Build model
self.device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
self.model = MockingjayModel(self.model_config,
inp_dim).to(self.device)
# Load from a PyTorch state_dict
load = bool(strtobool(options["load_pretrain"]))
if load:
self.load_model(all_states['Mockingjay'])
print('[Mockingjay] - Number of parameters: ' + str(
sum(p.numel()
for p in self.model.parameters() if p.requires_grad)))
self.out_dim = 768 # This attribute is for pytorch-kaldi
def set_model(self, inference=False, with_head=False, from_path=None, output_attention=False):
self.verbose('Initializing Mockingjay model.')
# uild the Mockingjay model with speech prediction head
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
self.output_attention = output_attention
if not inference or with_head:
self.model = MockingjayForMaskedAcousticModel(self.model_config, self.input_dim, self.output_dim, self.output_attention).to(self.device)
self.verbose('Number of parameters: ' + str(sum(p.numel() for p in self.model.parameters() if p.requires_grad)))
self.mockingjay = self.model.Mockingjay
if inference and not with_head:
self.mockingjay = MockingjayModel(self.model_config, self.input_dim, self.output_attention).to(self.device)
self.verbose('Number of parameters: ' + str(sum(p.numel() for p in self.mockingjay.parameters() if p.requires_grad)))
self.mockingjay.eval()
elif inference and with_head:
self.model.eval()
elif not inference:
self.model.train()
# Setup optimizer
param_optimizer = list(self.model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
num_train_optimization_steps = self.total_steps // self.gradient_accumulation_steps
if self.apex:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=self.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if self.config['optimizer']['loss_scale'] == 0:
self.optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
self.optimizer = FP16_Optimizer(optimizer, static_loss_scale=self.config['optimizer']['loss_scale'])
self.warmup_linear = WarmupLinearSchedule(warmup=self.warmup_proportion,
t_total=num_train_optimization_steps)
else:
self.optimizer = BertAdam(optimizer_grouped_parameters,
lr=self.learning_rate,
warmup=self.warmup_proportion,
t_total=num_train_optimization_steps)
else:
raise NotImplementedError('Invalid Arguments!')
if self.load: # This will be set to True by default when Tester is running set_model()
self.load_model(inference=inference, with_head=with_head, from_path=from_path)
def __init__(self, options, inp_dim):
super(MOCKINGJAY, self).__init__()
all_states = torch.load(options["ckpt_file"], map_location='cpu')
self.config = all_states['Settings']['Config']
self.no_grad = bool(strtobool(options['no_grad']))
self.spec_aug = bool(strtobool(options['spec_aug']))
self.spec_aug_prev = bool(strtobool(options['spec_aug_prev']))
self.weighted_sum = bool(strtobool(options['weighted_sum']))
self.select_layer = int(options['select_layer'])
if (not self.no_grad) and (not self.spec_aug_prev): raise RuntimeError('Only one of them can be set False!')
# increase dropout
if str(options['dropout']) != 'default':
self.config['mockingjay']['hidden_dropout_prob'] = float(options['dropout'])
self.config['mockingjay']['attention_probs_dropout_prob'] = float(options['dropout'])
# Model Config
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
self.num_layers = self.model_config.num_hidden_layers
if not (self.select_layer in list(range(-1, self.num_layers))): raise RuntimeError('Out of range int for \'select_layer\'!')
# use weighted sum from all layers
if self.weighted_sum:
self.weight = nn.Parameter(torch.ones(self.num_layers) / self.num_layers)
# Build model
self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
self.model = MockingjayModel(self.model_config, inp_dim).to(self.device)
self.model.eval() if self.no_grad else self.model.train()
# Load from a PyTorch state_dict
load = bool(strtobool(options["load_pretrain"]))
if load:
self.load_model(all_states['Mockingjay'])
print('[Mockingjay] - Number of parameters: ' + str(sum(p.numel() for p in self.model.parameters() if p.requires_grad)))
self.out_dim = self.hidden_size # 768, This attribute is for pytorch-kaldi
class Solver():
''' Super class Solver for all kinds of tasks'''
def __init__(self, config, paras):
# General Settings
self.config = config
self.paras = paras
self.device = torch.device('cuda') if (
self.paras.gpu
and torch.cuda.is_available()) else torch.device('cpu')
if torch.cuda.is_available(): self.verbose('CUDA is available!')
# path and directories
self.exp_name = paras.name
if self.exp_name is None:
self.exp_name = '_'.join([
paras.config.split('/')[-1].replace('.yaml', ''),
'sd' + str(paras.seed)
])
self.ckpdir = os.path.join(paras.ckpdir, self.exp_name)
self.load = paras.load
# only for test
self.ckpt = os.path.join(paras.ckpdir, paras.ckpt)
# model
self.load_model_list = config['solver']['load_model_list']
self.duo_feature = config['solver']['duo_feature']
self.output_dim = num_freq if self.duo_feature else None # output dim is the same as input dim if not using duo features
if 'input_dim' in config['mockingjay']:
self.input_dim = config['mockingjay']['input_dim']
self.verbose('Using `input_dim` setting from config.')
else:
input_dim = fmllr_dim if 'fmllr' in self.config['dataloader'][
'data_path'] else mfcc_dim if 'mfcc' in self.config[
'dataloader']['data_path'] else mel_dim
def verbose(self, msg, end='\n'):
''' Verbose function for print information to stdout'''
if self.paras.verbose:
print('[SOLVER] - ', msg, end=end)
def load_data(self, split='train', load_mel_only=False):
''' Load data for training / testing'''
if split == 'train':
self.verbose('Loading source data ' +
str(self.config['dataloader']['train_set']) +
' from ' + self.config['dataloader']['data_path'])
if self.duo_feature:
self.verbose('Loading target data ' +
str(self.config['dataloader']['train_set']) +
' from ' +
self.config['dataloader']['target_path'])
elif split == 'test':
self.verbose('Loading testing data ' +
str(self.config['dataloader']['test_set']) +
' from ' + self.config['dataloader']['data_path'])
else:
raise NotImplementedError('Invalid `split` argument!')
if self.duo_feature and not load_mel_only:
setattr(self, 'dataloader', get_Dataloader(split, load='duo', use_gpu=self.paras.gpu, \
mock_config=self.config['mockingjay'], **self.config['dataloader'])) # Currently the duo feature dataloader only supports mockingjay training, no need to specify `run_mockingjay`
else:
setattr(self, 'dataloader', get_Dataloader(split, load='acoustic', use_gpu=self.paras.gpu, \
run_mockingjay=True if not load_mel_only else False, mock_config=self.config['mockingjay'], \
**self.config['dataloader'])) # specify `run_mockingjay` so dataloader will process mockingjay MAM data
def set_model(self,
inference=False,
with_head=False,
from_path=None,
output_attention=False):
self.verbose('Initializing Mockingjay model.')
# uild the Mockingjay model with speech prediction head
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
self.with_head = with_head
self.output_attention = output_attention
if not inference or with_head:
self.model = MockingjayForMaskedAcousticModel(
self.model_config, self.input_dim, self.output_dim,
self.output_attention).to(self.device)
self.mockingjay = self.model.Mockingjay
if self.paras.multi_gpu:
self.model = torch.nn.DataParallel(self.model)
self.mockingjay = torch.nn.DataParallel(self.mockingjay)
self.verbose('Multi-GPU training Enabled: ' +
str(torch.cuda.device_count()))
self.verbose('Number of parameters: ' + str(
sum(p.numel()
for p in self.model.parameters() if p.requires_grad)))
if inference and not with_head:
self.mockingjay = MockingjayModel(
self.model_config, self.input_dim,
self.output_attention).to(self.device)
if self.paras.multi_gpu:
self.mockingjay = torch.nn.DataParallel(self.mockingjay)
self.verbose('Multi-GPU training Enabled: ' +
str(torch.cuda.device_count()))
self.verbose('Number of parameters: ' + str(
sum(p.numel()
for p in self.mockingjay.parameters() if p.requires_grad)))
self.mockingjay.eval()
elif inference and with_head:
self.model.eval()
elif not inference:
self.model.train()
# Setup optimizer
param_optimizer = list(self.model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params': [
p for n, p in param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
if self.apex:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=self.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if self.config['optimizer']['loss_scale'] == 0:
self.optimizer = FP16_Optimizer(optimizer,
dynamic_loss_scale=True)
else:
self.optimizer = FP16_Optimizer(
optimizer,
static_loss_scale=self.config['optimizer']
['loss_scale'])
self.warmup_linear = WarmupLinearSchedule(
warmup=self.warmup_proportion, t_total=self.total_steps)
else:
self.optimizer = BertAdam(optimizer_grouped_parameters,
lr=self.learning_rate,
warmup=self.warmup_proportion,
t_total=self.total_steps)
else:
raise NotImplementedError('Invalid Arguments!')
if self.load: # This will be set to True by default when Tester is running set_model()
self.load_model(inference=inference,
with_head=with_head,
from_path=from_path)
def save_model(self, name, model_all=True):
if model_all:
all_states = {
'SpecHead':
self.model.SpecHead.state_dict() if not self.paras.multi_gpu
else self.model.module.SpecHead.state_dict(),
'Mockingjay':
self.mockingjay.state_dict() if not self.paras.multi_gpu else
self.mockingjay.module.state_dict(),
'Optimizer':
self.optimizer.state_dict(),
'Global_step':
self.global_step,
'Settings': {
'Config': self.config,
'Paras': self.paras,
},
}
else:
all_states = {
'Mockingjay':
self.mockingjay.state_dict() if not self.paras.multi_gpu else
self.mockingjay.module.state_dict(),
'Settings': {
'Config': self.config,
'Paras': self.paras,
},
}
new_model_path = '{}/{}-{}.ckpt'.format(self.ckpdir, name,
self.global_step)
torch.save(all_states, new_model_path)
self.model_kept.append(new_model_path)
if len(self.model_kept) >= self.max_keep:
os.remove(self.model_kept[0])
self.model_kept.pop(0)
def load_model(self, inference=False, with_head=False, from_path=None):
if from_path is not None:
self.verbose('Load model from {}'.format(from_path))
all_states = torch.load(from_path, map_location='cpu')
self.load_model_list = ['Mockingjay']
else:
self.verbose('Load model from {}'.format(self.ckpt))
all_states = torch.load(self.ckpt, map_location='cpu')
if 'SpecHead' in self.load_model_list:
if not inference or with_head:
try:
if not self.paras.multi_gpu:
self.model.SpecHead.load_state_dict(
all_states['SpecHead'])
else:
self.model.module.SpecHead.load_state_dict(
all_states['SpecHead'])
self.verbose('[SpecHead] - Loaded')
except:
self.verbose('[SpecHead - X]')
if 'Mockingjay' in self.load_model_list:
try:
state_dict = all_states['Mockingjay']
# Load from a PyTorch state_dict
old_keys = []
new_keys = []
for key in state_dict.keys():
new_key = None
if 'gamma' in key:
new_key = key.replace('gamma', 'weight')
if 'beta' in key:
new_key = key.replace('beta', 'bias')
if new_key:
old_keys.append(key)
new_keys.append(new_key)
for old_key, new_key in zip(old_keys, new_keys):
state_dict[new_key] = state_dict.pop(old_key)
missing_keys = []
unexpected_keys = []
error_msgs = []
# copy state_dict so _load_from_state_dict can modify it
metadata = getattr(state_dict, '_metadata', None)
state_dict = state_dict.copy()
if metadata is not None:
state_dict._metadata = metadata
def load(module, prefix=''):
local_metadata = {} if metadata is None else metadata.get(
prefix[:-1], {})
module._load_from_state_dict(state_dict, prefix,
local_metadata, True,
missing_keys, unexpected_keys,
error_msgs)
for name, child in module._modules.items():
if child is not None:
load(child, prefix + name + '.')
# perform load
if not self.paras.multi_gpu:
load(self.mockingjay)
else:
load(self.mockingjay.module)
if len(missing_keys) > 0:
self.verbose(
"Weights of {} not initialized from pretrained model: {}"
.format(self.mockingjay.__class__.__name__,
missing_keys))
if len(unexpected_keys) > 0:
self.verbose(
"Weights from pretrained model not used in {}: {}".
format(self.mockingjay.__class__.__name__,
unexpected_keys))
if len(error_msgs) > 0:
raise RuntimeError(
'Error(s) in loading state_dict for {}:\n\t{}'.format(
self.mockingjay.__class__.__name__,
"\n\t".join(error_msgs)))
self.verbose('[Mockingjay] - Loaded')
except:
self.verbose('[Mockingjay - X]')
if 'Optimizer' in self.load_model_list and not inference:
try:
self.optimizer.load_state_dict(all_states['Optimizer'])
for state in self.optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
self.verbose('[Optimizer] - Loaded')
except:
self.verbose('[Optimizer - X]')
if 'Global_step' in self.load_model_list and not inference:
try:
self.global_step = all_states['Global_step']
self.verbose('[Global_step] - Loaded')
except:
self.verbose('[Global_step - X]')
self.verbose('Model loading complete!')
def up_sample_frames(self, spec, return_first=False):
if len(spec.shape) != 3:
spec = spec.unsqueeze(0)
assert (len(spec.shape) == 3
), 'Input should have acoustic feature of shape BxTxD'
# spec shape: [batch_size, sequence_length // downsample_rate, output_dim * downsample_rate]
spec_flatten = spec.view(spec.shape[0], spec.shape[1] * self.dr,
spec.shape[2] // self.dr)
if return_first: return spec_flatten[0]
return spec_flatten # spec_flatten shape: [batch_size, sequence_length * downsample_rate, output_dim // downsample_rate]
def down_sample_frames(self, spec):
left_over = spec.shape[1] % self.dr
if left_over != 0: spec = spec[:, :-left_over, :]
spec_stacked = spec.view(spec.shape[0], spec.shape[1] // self.dr,
spec.shape[2] * self.dr)
return spec_stacked
def position_encoding(self, seq_len, batch_size=None, padding_idx=None):
''' Sinusoid position encoding table '''
def cal_angle(position, hid_idx):
return position / np.power(10000, 2 *
(hid_idx // 2) / self.hidden_size)
def get_posi_angle_vec(position):
return [
cal_angle(position, hid_j) for hid_j in range(self.hidden_size)
]
sinusoid_table = np.array(
[get_posi_angle_vec(pos_i) for pos_i in range(seq_len)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
if padding_idx is not None:
sinusoid_table[
padding_idx:] = 0. # zero vector for padding dimension
if batch_size is not None:
batch_sinusoid_table = np.repeat(sinusoid_table[np.newaxis, ...],
batch_size,
axis=0)
return batch_sinusoid_table # (batch_size, seq_len, hidden_size)
else:
return sinusoid_table # (seq_len, hidden_size)
class MOCKINGJAY(nn.Module):
def __init__(self, options, inp_dim):
super(MOCKINGJAY, self).__init__()
all_states = torch.load(options["ckpt_file"], map_location='cpu')
self.config = all_states['Settings']['Config']
self.no_grad = bool(strtobool(options['no_grad']))
self.spec_aug = bool(strtobool(options['spec_aug']))
self.spec_aug_prev = bool(strtobool(options['spec_aug_prev']))
self.weighted_sum = bool(strtobool(options['weighted_sum']))
self.select_layer = int(options['select_layer'])
if (not self.no_grad) and (not self.spec_aug_prev): raise RuntimeError('Only one of them can be set False!')
# increase dropout
if str(options['dropout']) != 'default':
self.config['mockingjay']['hidden_dropout_prob'] = float(options['dropout'])
self.config['mockingjay']['attention_probs_dropout_prob'] = float(options['dropout'])
# Model Config
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
self.num_layers = self.model_config.num_hidden_layers
if not (self.select_layer in list(range(-1, self.num_layers))): raise RuntimeError('Out of range int for \'select_layer\'!')
# use weighted sum from all layers
if self.weighted_sum:
self.weight = nn.Parameter(torch.ones(self.num_layers) / self.num_layers)
# Build model
self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
self.model = MockingjayModel(self.model_config, inp_dim).to(self.device)
self.model.eval() if self.no_grad else self.model.train()
# Load from a PyTorch state_dict
load = bool(strtobool(options["load_pretrain"]))
if load:
self.load_model(all_states['Mockingjay'])
print('[Mockingjay] - Number of parameters: ' + str(sum(p.numel() for p in self.model.parameters() if p.requires_grad)))
self.out_dim = self.hidden_size # 768, This attribute is for pytorch-kaldi
def load_model(self, state_dict):
try:
old_keys = []
new_keys = []
for key in state_dict.keys():
new_key = None
if 'gamma' in key:
new_key = key.replace('gamma', 'weight')
if 'beta' in key:
new_key = key.replace('beta', 'bias')
if new_key:
old_keys.append(key)
new_keys.append(new_key)
for old_key, new_key in zip(old_keys, new_keys):
state_dict[new_key] = state_dict.pop(old_key)
missing_keys = []
unexpected_keys = []
error_msgs = []
# copy state_dict so _load_from_state_dict can modify it
metadata = getattr(state_dict, '_metadata', None)
state_dict = state_dict.copy()
if metadata is not None:
state_dict._metadata = metadata
def load(module, prefix=''):
local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
module._load_from_state_dict(
state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)
for name, child in module._modules.items():
if child is not None:
load(child, prefix + name + '.')
load(self.model)
if len(missing_keys) > 0:
print('Weights of {} not initialized from pretrained model: {}'.format(
self.model.__class__.__name__, missing_keys))
if len(unexpected_keys) > 0:
print('Weights from pretrained model not used in {}: {}'.format(
self.model.__class__.__name__, unexpected_keys))
if len(error_msgs) > 0:
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
self.model.__class__.__name__, '\n\t'.join(error_msgs)))
print('[Mockingjay] - Pre-trained weights loaded!')
except: print('[Mockingjay] - Pre-trained weights NOT loaded!')
def down_sample_frames(self, spec):
spec = spec.contiguous()
left_over = spec.shape[1] % self.dr
if left_over != 0: spec = spec[:, :-left_over, :]
spec_stacked = spec.view(spec.shape[0], spec.shape[1]//self.dr, spec.shape[2]*self.dr)
return spec_stacked
def process_input_data(self, spec):
"""Process input data for the model"""
# add arbitary batch axis B if input `spec` has shape of TxD
if len(spec.shape) == 2:
spec = spec.unsqueeze(0)
# input `spec` should have shape BxTxD
elif len(spec.shape) != 3:
raise ValueError('Input argument `spec` has invalid shape: {}'.format(spec.shape))
# Down sample
spec_stacked = self.down_sample_frames(spec) # (batch_size, seq_len, feature_dim * dr)
# Record length for each uttr
spec_len = np.sum(np.sum(spec_stacked.cpu().data.numpy(), axis=-1) != 0, axis=-1)
spec_len = [int(sl) for sl in spec_len]
batch_size = spec_stacked.shape[0]
seq_len = spec_stacked.shape[1]
pos_enc = position_encoding(seq_len, self.hidden_size) # (seq_len, hidden_size)
attn_mask = np.ones((batch_size, seq_len)) # (batch_size, seq_len)
# zero vectors for padding dimension
for idx in range(len(spec_stacked)):
attn_mask[idx][spec_len[idx]:] = 0
if self.spec_aug and self.spec_aug_prev and self.model.training:
spec_stacked = spec_augment(spec_stacked, mask_T=70, mask_F=4, num_T=2, num_F=2, p=1.0) # (batch_size, seq_len, feature_dim * dr)
spec_stacked = spec_stacked.to(device=self.device, dtype=torch.float32) # (batch_size, seq_len, feature_dim * dr)
pos_enc = torch.FloatTensor(pos_enc).to(device=self.device, dtype=torch.float32).expand(spec_stacked.size(0), *pos_enc.size()) # (batch_size, seq_len, hidden_size)
attn_mask = torch.FloatTensor(attn_mask).to(device=self.device, dtype=torch.float32) # (batch_size, seq_len)
return spec_stacked, pos_enc, attn_mask # (x, pos_enc, attention_mask)
def tile_representations(self, reps):
"""
Tile up the mockingjay representations to match the amount of input frames.
Input - encoded_layers shape: (batch_size, sequence_length, hidden_size)
Output - tiled_encoded_layers shape: (batch_size, sequence_length * downsample_rate, hidden_size)
"""
if len(reps.shape) != 3:
raise ValueError('Input argument `reps` has invalid shape: {}'.format(reps.shape))
tiled_reps = reps.repeat(1, 1, self.dr)
tiled_reps = tiled_reps.reshape(reps.size(0), reps.size(1)*self.dr, reps.size(2))
return tiled_reps # (batch_size, sequence_length * downsample_rate, hidden_size)
def _forward(self, x):
# Compute padding to compromise the downsample loss
input_len = x.shape[0]
left_over = input_len % self.dr
if left_over % 2 == 0:
left_pad = left_over // 2
right_pad = left_pad
else:
left_pad = left_over // 2
right_pad = left_over // 2 + 1
# Model forwarding
x = x.permute(1, 0, 2).contiguous() # (T, B, D) -> (B, T, D)
spec_stacked, pos_enc, attn_mask = self.process_input_data(x)
x = self.model(spec_stacked, pos_enc, attn_mask, output_all_encoded_layers=self.weighted_sum or self.select_layer != -1) # (B, T, D) or # (N, B, T, D)
# Apply weighted sum
if self.weighted_sum:
if type(x) is list: x = torch.stack(x)
softmax_weight = nn.functional.softmax(self.weight, dim=-1)
B, T, D = x.shape[1], x.shape[2], x.shape[3]
x = x.reshape(self.num_layers, -1)
x = torch.matmul(softmax_weight, x).reshape(B, T, D)
# Select a specific layer
elif self.select_layer != -1:
x = x[self.select_layer]
if self.spec_aug and not self.spec_aug_prev and self.model.training:
x = spec_augment(x, mask_T=70, mask_F=86, num_T=2, num_F=2, p=1.0) # (B, T, D)
# If using a downsampling model, apply tile and padding
if x.shape[1] != input_len:
x = self.tile_representations(x)
# padding
x = x.permute(0, 2, 1).contiguous() # (B, T, D) -> (B, D, T)
padding = nn.ReplicationPad1d((left_pad, right_pad))
x = padding(x)
x = x.permute(2, 0, 1).contiguous() # (B, D, T) -> (T, B, D)
# If not using a downsampling model, permute to output
else:
x = x.permute(1, 0, 2).contiguous() # (B, T, D) -> (T, B, D)
return x
def forward(self, x):
if self.no_grad:
with torch.no_grad():
self.model.eval()
x = self._forward(x)
else:
x = self._forward(x)
return x
class MOCKINGJAY(nn.Module):
def __init__(self, options, inp_dim):
super(MOCKINGJAY, self).__init__()
all_states = torch.load(options["ckpt_file"], map_location='cpu')
self.config = all_states['Settings']['Config']
self.no_grad = bool(strtobool(options["no_grad"]))
# increase dropout
if str(options['dropout']) != 'default':
self.config['mockingjay']['hidden_dropout_prob'] = float(
options['dropout'])
self.config['mockingjay']['attention_probs_dropout_prob'] = float(
options['dropout'])
# Model Config
self.model_config = MockingjayConfig(self.config)
self.dr = self.model_config.downsample_rate
self.hidden_size = self.model_config.hidden_size
# Build model
self.device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
self.model = MockingjayModel(self.model_config,
inp_dim).to(self.device)
# Load from a PyTorch state_dict
load = bool(strtobool(options["load_pretrain"]))
if load:
self.load_model(all_states['Mockingjay'])
print('[Mockingjay] - Number of parameters: ' + str(
sum(p.numel()
for p in self.model.parameters() if p.requires_grad)))
self.out_dim = 768 # This attribute is for pytorch-kaldi
def load_model(self, state_dict):
try:
old_keys = []
new_keys = []
for key in state_dict.keys():
new_key = None
if 'gamma' in key:
new_key = key.replace('gamma', 'weight')
if 'beta' in key:
new_key = key.replace('beta', 'bias')
if new_key:
old_keys.append(key)
new_keys.append(new_key)
for old_key, new_key in zip(old_keys, new_keys):
state_dict[new_key] = state_dict.pop(old_key)
missing_keys = []
unexpected_keys = []
error_msgs = []
# copy state_dict so _load_from_state_dict can modify it
metadata = getattr(state_dict, '_metadata', None)
state_dict = state_dict.copy()
if metadata is not None:
state_dict._metadata = metadata
def load(module, prefix=''):
local_metadata = {} if metadata is None else metadata.get(
prefix[:-1], {})
module._load_from_state_dict(state_dict, prefix,
local_metadata, True,
missing_keys, unexpected_keys,
error_msgs)
for name, child in module._modules.items():
if child is not None:
load(child, prefix + name + '.')
load(self.model)
if len(missing_keys) > 0:
print(
"Weights of {} not initialized from pretrained model: {}".
format(self.model.__class__.__name__, missing_keys))
if len(unexpected_keys) > 0:
print(
"Weights from pretrained model not used in {}: {}".format(
self.model.__class__.__name__, unexpected_keys))
if len(error_msgs) > 0:
raise RuntimeError(
'Error(s) in loading state_dict for {}:\n\t{}'.format(
self.model.__class__.__name__,
"\n\t".join(error_msgs)))
print('[Mockingjay] - Pre-trained weights loaded!')
except:
print('[Mockingjay] - Pre-trained weights NOT loaded!')
def down_sample_frames(self, spec):
spec = spec.contiguous()
left_over = spec.shape[1] % self.dr
if left_over != 0: spec = spec[:, :-left_over, :]
spec_stacked = spec.view(spec.shape[0], spec.shape[1] // self.dr,
spec.shape[2] * self.dr)
return spec_stacked
def position_encoding(self, seq_len, batch_size=None, padding_idx=None):
''' Sinusoid position encoding table '''
def cal_angle(position, hid_idx):
return position / np.power(10000, 2 *
(hid_idx // 2) / self.hidden_size)
def get_posi_angle_vec(position):
return [
cal_angle(position, hid_j) for hid_j in range(self.hidden_size)
]
sinusoid_table = np.array(
[get_posi_angle_vec(pos_i) for pos_i in range(seq_len)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
if padding_idx is not None:
sinusoid_table[
padding_idx:] = 0. # zero vector for padding dimension
if batch_size is not None:
batch_sinusoid_table = np.repeat(sinusoid_table[np.newaxis, ...],
batch_size,
axis=0)
return batch_sinusoid_table # (batch_size, seq_len, hidden_size)
else:
return sinusoid_table # (seq_len, hidden_size)
def process_MAM_data(self, spec):
"""Process testing data for the masked acoustic model"""
# add arbitary batch axis B if input `spec` has shape of TxD
if len(spec.shape) == 2:
spec = spec.unsqueeze(0)
# input `spec` should have shape BxTxD
elif len(spec.shape) != 3:
raise ValueError(
'Input argument `spec` has invalid shape: {}'.format(
spec.shape))
# Down sample
spec_stacked = self.down_sample_frames(
spec) # (batch_size, seq_len, mel_dim * dr)
# Record length for each uttr
spec_len = np.sum(
np.sum(spec_stacked.cpu().data.numpy(), axis=-1) != 0, axis=-1)
spec_len = [int(sl) for sl in spec_len]
batch_size = spec_stacked.shape[0]
seq_len = spec_stacked.shape[1]
pos_enc = self.position_encoding(
seq_len, batch_size) # (batch_size, seq_len, hidden_size)
attn_mask = np.ones((batch_size, seq_len)) # (batch_size, seq_len)
# zero vectors for padding dimension
for idx in range(len(spec_stacked)):
pos_enc[idx][spec_len[idx]:] = 0
attn_mask[idx][spec_len[idx]:] = 0
spec_stacked = spec_stacked.to(device=self.device, dtype=torch.float32)
pos_enc = torch.FloatTensor(pos_enc).to(device=self.device,
dtype=torch.float32)
attn_mask = torch.FloatTensor(attn_mask).to(device=self.device,
dtype=torch.float32)
return spec_stacked, pos_enc, attn_mask # (x, pos_enc, attention_mask)
def tile_representations(self, reps):
"""
Tile up the mockingjay representations to match the amount of input frames.
Input - encoded_layers shape: (batch_size, sequence_length, hidden_size)
Output - tiled_encoded_layers shape: (batch_size, sequence_length * downsample_rate, hidden_size)
"""
if len(reps.shape) != 3:
raise ValueError(
'Input argument `reps` has invalid shape: {}'.format(
reps.shape))
tiled_reps = reps.repeat(1, 1, self.dr)
tiled_reps = tiled_reps.reshape(reps.size(0),
reps.size(1) * self.dr, reps.size(2))
return tiled_reps # (batch_size, sequence_length * downsample_rate, hidden_size)
def _forward(self, x):
# Compute padding to compromise the downsample loss
input_len = x.shape[0]
left_over = input_len % self.dr
if left_over % 2 == 0:
left_pad = left_over // 2
right_pad = left_pad
else:
left_pad = left_over // 2
right_pad = left_over // 2 + 1
# Model forwarding
x = x.permute(1, 0, 2).contiguous() # (T, B, D) -> (B, T, D)
spec_stacked, pos_enc, attn_mask = self.process_MAM_data(x)
x = self.model(spec_stacked,
pos_enc,
attn_mask,
output_all_encoded_layers=False) # (B, T, D)
# If using a downsampling model, apply tile and padding
if x.shape[1] != input_len:
x = self.tile_representations(x)
# padding
x = x.permute(0, 2, 1).contiguous() # (B, T, D) -> (B, D, T)
padding = nn.ReplicationPad1d((left_pad, right_pad))
x = padding(x)
x = x.permute(2, 0, 1).contiguous() # (B, D, T) -> (T, B, D)
# If not using a downsampling model, permute to output
else:
x = x.permute(1, 0, 2).contiguous() # (B, T, D) -> (T, B, D)
return x
def forward(self, x):
if self.no_grad:
with torch.no_grad():
self.model.eval()
x = self._forward(x)
else:
x = self._forward(x)
torch.cuda.empty_cache()
return x