def train_merge():
hyperparams = {
'batch_size': 50,
'learning_rate': 1e-4,
'num_updates': 20000,
'grad_norm_clip': 5
}
model = MnistMergeModel(hyperparams)
loss = tb.Crossentropy(hyperparams)
acc = tb.CatAcc(hyperparams)
evaluator = tb.Evaluator(hyperparams, loss, acc)
optim = tb.AdamOptim(hyperparams)
trainer = tb.Trainer(model, hyperparams, loss, optim, evaluator)
mnist = load_data()
train_xs = {
'image_1': mnist['train']['images'],
'image_2': mnist['train']['images']
}
train_y = mnist['train']['labels']
val_xs = {
'image_1': mnist['test']['images'],
'image_2': mnist['test']['images']
}
val_y = mnist['test']['labels']
trainer.train(train_xs, train_y, val_xs, val_y)
evaluator.eval(model, val_xs, val_y)
def train():
hyperparams = {'batch_size': 50,
'learning_rate': 0.001,
'grad_decay': 0.99,
'grad_epsilon': 0.01,
'num_updates': 20000,
'grad_norm_clip': 5}
model = CoinbaseModel(hyperparams)
loss = tb.Crossentropy(hyperparams)
acc = tb.CatAcc(hyperparams)
evaluator = tb.Evaluator(hyperparams, loss, acc)
optim = tb.RMSPropOptim(hyperparams)
trainer = tb.Trainer(model, hyperparams, loss, optim, evaluator)
split = 10000
data = np.load('data/coinbase_n1.npz')
train_xs = {'first_ticks': data['first_ticks'][:split],
'last_ticks': data['last_ticks'][:split],
'features': data['features'][:split]}
train_y = data['targets'][:split]
val_xs = {'first_ticks': data['first_ticks'][split:],
'last_ticks': data['last_ticks'][split:],
'features': data['features'][split:]}
val_y = data['targets'][split:]
trainer.train(train_xs, train_y,
val_xs, val_y,
val_cmp=True)
evaluator.eval(model, val_xs, val_y)
def train():
hyperparams = {'batch_size': 512,
'learning_rate': 0.0001,
'grad_decay': 0.95,
'grad_epsilon': 0.01,
'num_updates': 100000,
'grad_norm_clip': 5}
with tf.device('/cpu:0'):
model = CoinbaseModel(hyperparams)
loss = tb.Crossentropy(hyperparams)
acc = tb.CatAcc(hyperparams)
evaluator = tb.Evaluator(hyperparams, loss, acc)
optim = tb.RMSPropOptim(hyperparams)
trainer = tb.Trainer(model, hyperparams, loss, optim, evaluator)
split = 90000
data = np.load('data/coinbase-ticks-100000.npz')
print(data['ticks'].shape)
train_xs = {'ticks': data['ticks'][:split]}
train_y = data['targets'][:split]
val_xs = {'ticks': data['ticks'][split:]}
val_y = data['targets'][split:]
with tf.device('/cpu:0'):
trainer.train(train_xs, train_y,
val_xs, val_y,
val_cmp=True)
evaluator.eval(model, val_xs, val_y)
def train_char_rnn():
print("Building network ...")
gen_info = load_data(seqlength=20, text_fnm='data/nietzsche.txt')
train_xs = {'text': gen_info['data']['train']['text']}
train_y = gen_info['data']['train']['targets']
test_xs = {'text': gen_info['data']['test']['text']}
test_y = gen_info['data']['test']['targets']
print(train_xs['text'][:2])
print(train_y[:2])
hyperparams = {
'batch_size': 128,
'learning_rate': 0.001,
'num_updates': 50000,
'grad_norm_clip': 5,
'vocab_size': gen_info['vocab_size'],
'char_to_index': gen_info['char_to_index'],
'index_to_char': gen_info['index_to_char'],
'seqlength': gen_info['seqlength']
}
model = LangModel(hyperparams)
loss = tb.Crossentropy(hyperparams)
acc = tb.Perplexity(hyperparams)
evaluator = tb.SeqGenEvaluator(hyperparams,
loss,
acc,
seed='the quick brown fox jumps')
optim = tb.AdamOptim(hyperparams)
trainer = tb.Trainer(model, hyperparams, loss, optim, evaluator)
trainer.train(train_xs, train_y, test_xs, test_y, display_interval=100)
def train_brnn():
# gen_info = load_data(seqlength=20,
# text_fnm='data/ptb/ptb.train.txt')
# train_xs = {'text': gen_info['data']['train']['text']}
# train_y = gen_info['data']['train']['targets']
# val_xs = {'text': gen_info['data']['test']['text']}
# val_y = gen_info['data']['test']['targets']
# print(train_xs['text'][:2])
# print(train_y[:2])
# hyperparams = {'batch_size': 128,
# 'learning_rate': 0.01,
# 'num_updates': 50000,
# 'grad_norm_clip': 5,
# 'vocab_size': gen_info['vocab_size'],
# 'char_to_index': gen_info['char_to_index'],
# 'index_to_char': gen_info['index_to_char'],
# 'seqlength': gen_info['seqlength']}
seqlength = 20
vocab_size = 10000
data = load_data(seqlength=seqlength)
train_xs = {'text': data['train']['text']}
train_y = data['train']['targets']
val_xs = {'text': data['test']['text']}
val_y = data['test']['targets']
hyperparams = {'batch_size': 128,
'learning_rate': 0.0001,
'num_updates': 50000,
'grad_norm_clip': 5,
'vocab_size': vocab_size,
'seqlength': seqlength}
model = PTBBasic(hyperparams)
loss = tb.Crossentropy(hyperparams)
acc = tb.Perplexity(hyperparams)
evaluator = tb.PerplexityEvaluator(
hyperparams, loss, acc,
seed='the quick brown fox jumps')
optim = tb.AdamOptim(hyperparams)
trainer = tb.Trainer(model, hyperparams, loss, optim, evaluator)
print('Training...')
trainer.train(train_xs, train_y,
val_xs, val_y,
val_cmp=False,
display_interval=100)
evaluator.eval(model, val_xs, val_y)
