def __init__(self, input_dim, num_class, CONFIG):
super(Model, self).__init__()
self.bidirectional = True
self._hid_dim = 256
self._rnn2_input_size = self._hid_dim * 2 if self.bidirectional else self._hid_dim
self._ffn_input_size = self._hid_dim * 2 if self.bidirectional else self._hid_dim
# Define RNN model
self.rnn1 = nn.LSTM(input_size=input_dim * 2,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn2 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn3 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn4 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn5 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
# Define FFN after RNN
self.fc = nn.Linear(self._ffn_input_size, num_class + 1)
self.dropout = nn.Dropout(CONFIG['dropout'])
self.use_cuda = CONFIG['cuda']
self.ctc_loss = ctc.CTCLoss()
self.blank = num_class
self.clip = CONFIG['optimizer']['clip']
self.reduce_factor = 2
self.ngram_lm = n_gram_lm(lm_path='WSJ_Character_7.lm',
n_gram=3,
character_set='Character_Set.txt')
if CONFIG['xavier']:
self.init_weights()
def loss(self, batch):
x, y, x_lens, y_lens = self.collate(*batch)
out = self.forward_impl(x)
loss_fn = ctc.CTCLoss()
loss = loss_fn(out, y, x_lens, y_lens)
return loss
def awni_loss(out, labels, input_lens, label_lens, blank_idx):
"""Calculates the loss using awni hannun's warpctc bindings.
Only works with pytorch 0.4.
"""
import functions.ctc as ctc #awni hannun's ctc bindings
loss_fn = ctc.CTCLoss(blank_label=blank_idx)
loss = loss_fn(out, labels, input_lens, label_lens)
return loss
def __init__(self, input_dim, num_class, CONFIG):
super(Model, self).__init__()
self.bidirectional = True
self._hid_dim = 256
self._rnn2_input_size = self._hid_dim * 2 if self.bidirectional else self._hid_dim
self._ffn_input_size = self._hid_dim * 2 if self.bidirectional else self._hid_dim
# Define RNN model
self.rnn1 = nn.LSTM(input_size=input_dim * 2,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn2 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn3 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn4 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
self.rnn5 = nn.LSTM(input_size=self._rnn2_input_size,
hidden_size=256,
num_layers=1,
batch_first=True,
dropout=0,
bidirectional=True)
# Define FFN after RNN
self.fc = nn.Linear(self._ffn_input_size, num_class + 1)
self.dropout = nn.Dropout(0.1)
self.use_cuda = CONFIG['cuda']
self.ctc_loss = ctc.CTCLoss()
self.blank = num_class
self.clip = CONFIG['optimizer']['clip']
self.reduce_factor = 2
def test_model(test_label):
"""
this checks if there are nans in a variety of test scenarios
Arguments:
test_label (str): the label passed into alter_logspec
"""
with open(CONFIG_PATH, 'r') as fid:
config = json.load(fid)
opt_cfg = config["optimizer"]
model_cfg = config["model"]
model, preproc = load(MODEL_PATH)
optimizer = torch.optim.SGD(model.parameters(),
lr=opt_cfg["learning_rate"],
momentum=opt_cfg["momentum"])
log_spec, samp_rate = load_audio(AUDIO_PATH, preproc)
inputs = alter_logspec(log_spec, alter=test_label)
labels = [preproc.encode(TEXT)]
batch = ((inputs, labels))
x, y, x_lens, y_lens = model.collate(*batch)
out, rnn_args = model.forward_impl(x)
loss_fn = ctc.CTCLoss()
loss = loss_fn(out, y, x_lens, y_lens)
loss.backward()
loss = loss.item()
grad_norm = nn.utils.clip_grad_norm_(model.parameters(), 200)
print(f"loss: {loss}, grad_norm: {grad_norm}")
optimizer.step()
print(f"Are there nan's? : {check_nan(model)}")
def loss(self, out, batch):
x, y, x_lens, y_lens = self.collate(*batch)
batch_size, _, out_dim = out.size()
loss_fn = ctc.CTCLoss()
loss = loss_fn(out, y, x_lens, y_lens)
return loss