return np.exp(loss / label.size)
if __name__ == '__main__':
batch_size = 128
buckets = [10, 20, 30, 40, 50, 60]
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 2
learning_rate = 0.01
momentum = 0.0
contexts = [mx.context.gpu(i) for i in range(1)]
vocab = default_build_vocab(os.path.join(data_dir, 'ptb.train.txt'))
init_h = [
mx.io.DataDesc('LSTM_state', (num_lstm_layer, batch_size, num_hidden),
layout='TNC')
]
init_c = [
mx.io.DataDesc('LSTM_state_cell',
(num_lstm_layer, batch_size, num_hidden),
layout='TNC')
]
init_states = init_c + init_h
data_train = BucketSentenceIter(os.path.join(data_dir, 'ptb.train.txt'),
vocab,
buckets,
return data
batch_size = 20
seq_len = 35
num_hidden = 400
num_embed = 200
num_lstm_layer = 8
num_round = 25
learning_rate = 0.1
wd = 0.
momentum = 0.0
max_grad_norm = 5.0
update_period = 1
dic = default_build_vocab("./data/ptb.train.txt")
vocab = len(dic)
# static buckets
buckets = [8, 16, 24, 32, 60]
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_states = init_c + init_h
X_train_batch = BucketSentenceIter("./data/ptb.train.txt",
dic,
buckets,
batch_size,
#buckets = [32]
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
#num_epoch = 25
num_epoch = 2
learning_rate = 0.01
momentum = 0.0
# dummy data is used to test speed without IO
dummy_data = False
contexts = [mx.context.gpu(i) for i in range(1)]
vocab = default_build_vocab(os.path.join(data_dir, "ptb.train.txt"))
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
data_train = BucketSentenceIter(os.path.join(data_dir, "ptb.train.txt"), vocab,
buckets, batch_size, init_states)
data_val = BucketSentenceIter(os.path.join(data_dir, "ptb.valid.txt"), vocab,
buckets, batch_size, init_states)
if dummy_data:
data_train = DummyIter(data_train)
data_val = DummyIter(data_val)
state_names = [x[0] for x in init_states]
buckets = []
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 25
learning_rate = 0.01
momentum = 0.0
# dummy data is used to test speed without IO
dummy_data = False
#contexts = [mx.context.gpu(i) for i in range(1)]
contexts = mx.context.cpu()
vocab = default_build_vocab("./data/ptb.train.txt")
def sym_gen(seq_len):
return gru_unroll(num_lstm_layer, seq_len, len(vocab),
num_hidden=num_hidden, num_embed=num_embed,
num_label=len(vocab))
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
data_train = BucketSentenceIter("./data/ptb.train.txt", vocab,
buckets, batch_size, init_h)
data_val = BucketSentenceIter("./data/ptb.valid.txt", vocab,
buckets, batch_size, init_h)
if dummy_data:
data_train = DummyIter(data_train)
stream = p.open(format=p.get_format_from_width(2),
channels=1,
rate=44100,
output=True)
if __name__ == '__main__':
batch_size = 1
buckets = 30
num_hidden = 2500
num_label = 1500
num_lstm_layer = 2
num_epoch = 326
print(batch_size, buckets, num_hidden, num_lstm_layer, num_epoch)
img_data, wave_data = default_build_vocab("./data/data/2.mp4",
"./data/data/2.mp3")
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_states = init_c + init_h
data_train = BucketSentenceIter(img_data, wave_data, buckets, batch_size,
init_states, num_label)
model = mx.model.FeedForward.load('model/lip',
num_epoch,
ctx=mx.context.gpu(1),
num_epoch=500,
learning_rate=0.5)
import logging
head = '%(asctime)-15s %(message)s'
buckets = [32]
#buckets = []
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 25
learning_rate = 0.01
momentum = 0.0
# dummy data is used to test speed without IO
dummy_data = False
contexts = [mx.context.gpu(i) for i in range(N)]
vocab = default_build_vocab("./data/sherlockholmes.train.txt")
def sym_gen(seq_len):
return lstm_unroll(num_lstm_layer, seq_len, len(vocab),
num_hidden=num_hidden, num_embed=num_embed,
num_label=len(vocab))
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
data_train = BucketSentenceIter("./data/sherlockholmes.train.txt", vocab,
buckets, batch_size, init_states)
data_val = BucketSentenceIter("./data/sherlockholmes.valid.txt", vocab,
buckets, batch_size, init_states)
return np.exp(loss / label.size)
if __name__ == '__main__':
batch_size = 128
buckets = [10, 20, 30, 40, 50, 60]
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 2
learning_rate = 0.01
momentum = 0.0
contexts = [mx.context.gpu(i) for i in range(4)]
vocab = default_build_vocab(
os.path.join(data_dir, 'sherlockholmes.train.txt'))
init_h = [('LSTM_init_h', (batch_size, num_lstm_layer, num_hidden))]
init_c = [('LSTM_init_c', (batch_size, num_lstm_layer, num_hidden))]
init_states = init_c + init_h
data_train = BucketSentenceIter(
os.path.join(data_dir, 'sherlockholmes.train.txt'), vocab, buckets,
batch_size, init_states)
data_val = BucketSentenceIter(
os.path.join(data_dir, 'sherlockholmes.valid.txt'), vocab, buckets,
batch_size, init_states)
def sym_gen(seq_len):
data = mx.sym.Variable('data')
label = mx.sym.Variable('softmax_label')
return loss / pred.shape[0]
if __name__ == '__main__':
batch_size = 1
buckets = 15
num_hidden = 1500
num_label = 1764
num_lstm_layer = 2
num_epoch = 500
learning_rate = 1
momentum = 0.0
print(batch_size, buckets, num_hidden, num_lstm_layer, num_epoch,
learning_rate)
contexts = [mx.context.gpu(0)]
img_data, wave_data = default_build_vocab("./data/train_x_2.mp4",
"./data/train_y_2.mp3")
img_data1, wave_data1 = default_build_vocab("./data/train_x_2.mp4",
"./data/train_y_2.mp3")
def sym_gen(seq_len):
return bi_lstm_unroll(num_lstm_layer,
seq_len,
len(img_data),
num_hidden=num_hidden,
num_label=num_label)
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_states = init_c + init_h
buckets = []
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 25
learning_rate = 0.01
momentum = 0.0
# dummy data is used to test speed without IO
dummy_data = False
#contexts = [mx.context.gpu(i) for i in range(1)]
contexts = mx.context.cpu()
vocab = default_build_vocab("./data/sherlockholmes.train.txt")
def sym_gen(seq_len):
return gru_unroll(num_lstm_layer, seq_len, len(vocab),
num_hidden=num_hidden, num_embed=num_embed,
num_label=len(vocab))
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
data_train = BucketSentenceIter("./data/sherlockholmes.train.txt", vocab,
buckets, batch_size, init_h)
data_val = BucketSentenceIter("./data/sherlockholmes.valid.txt", vocab,
buckets, batch_size, init_h)
if dummy_data:
data_train = DummyIter(data_train)
return np.exp(loss / label.size)
if __name__ == '__main__':
batch_size = 128
buckets = [10, 20, 30, 40, 50, 60]
num_hidden = 200
num_embed = 200
num_lstm_layer = 2
num_epoch = 2
learning_rate = 0.01
momentum = 0.0
contexts = [mx.context.gpu(i) for i in range(4)]
vocab = default_build_vocab(os.path.join(data_dir, 'sherlockholmes.train.txt'))
init_h = [('LSTM_init_h', (batch_size, num_lstm_layer, num_hidden))]
init_c = [('LSTM_init_c', (batch_size, num_lstm_layer, num_hidden))]
init_states = init_c + init_h
data_train = BucketSentenceIter(os.path.join(data_dir, 'sherlockholmes.train.txt'),
vocab, buckets, batch_size, init_states)
data_val = BucketSentenceIter(os.path.join(data_dir, 'sherlockholmes.valid.txt'),
vocab, buckets, batch_size, init_states)
def sym_gen(seq_len):
data = mx.sym.Variable('data')
label = mx.sym.Variable('softmax_label')
embed = mx.sym.Embedding(data=data, input_dim=len(vocab),
output_dim=num_embed, name='embed')
return loss / pred.shape[0]
if __name__ == '__main__':
batch_size = 1
buckets = 30
num_hidden = 2500
num_label = 1445
num_lstm_layer = 2
num_epoch = 500
learning_rate = 1
momentum = 0.0
print(batch_size, buckets, num_hidden, num_lstm_layer, num_epoch,
learning_rate)
contexts = [mx.context.gpu(0)]
img_data, wave_data = default_build_vocab("./data/4.mkv", "./data/4.mp3")
def sym_gen(seq_len):
return lstm_unroll(num_lstm_layer,
seq_len,
len(img_data),
num_hidden=num_hidden,
num_label=num_label)
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_states = init_c + init_h
data_train = BucketSentenceIter(img_data, wave_data, buckets, batch_size,
batch_size = args.batch_size
buckets = [args.seq_len]
num_hidden = args.num_hidden
num_embed = args.num_embed
num_lstm_layer = args.lstm_layer
num_epoch = 1
learning_rate = 0.01
momentum = 0.0
# dummy data is used to test speed without IO
dummy_data = False
contexts = [mx.context.gpu(i) for i in range(1)]
vocab = default_build_vocab(args.data_path+"/ptb.train.txt")
def sym_gen(seq_len):
return lstm_unroll(num_lstm_layer, seq_len, 10000,
num_hidden=num_hidden, num_embed=num_embed,
num_label=10000)
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
data_train = BucketSentenceIter(args.data_path+"/ptb.train.txt", vocab,
buckets, batch_size, args.num_batch, init_states)
if dummy_data:
data_train = DummyIter(data_train)