def combine_regression_data():
utils.write_data(
input("Output File: "),
utils.join_data(
[input("file 1: "), input("file 2: ")],
expt=[5, 3],
d_filter=[[
utils.sanitize_lower, utils.sanitize_lower, float, float, float
], [utils.sanitize_lower, utils.sanitize_lower, float]]))
def sample_topk(self,
prefix,
end_tag=[u' ', u'.', u',', u'\n', u'\t'],
k=5,
B=100,
N=30):
# B, N: width, depth of beam search
prefix = u" " + prefix
prefix = utils.split_raw(prefix, self.text_modeling)
prefix = np.array(list(map(self.vocab.get, prefix)))
current_state = self.model.initial_rnn_state(1)
for x in prefix[:-1]:
_, current_state, _ = self.model.sample_output(
self.sess, x, current_state)
end_tag = [self.vocab.get(e, -1) for e in end_tag]
# not so efficient beam search. just initial version
candidates = [{
'p': 1.,
'seq': prefix[1:],
'rnn_state': current_state,
'done': False
}]
for depth in range(len(prefix) - 1, N):
new_candidates = []
for d in candidates:
if d['done']:
new_candidates.append(d)
continue
_, next_rnn_state, next_x_prob = self.model.sample_output(
self.sess, d['seq'][-1], d['rnn_state'])
for next_x in range(len(next_x_prob)):
next_p = d['p'] * next_x_prob[next_x]
next_seq = np.concatenate((d['seq'], [next_x]))
done = next_x in end_tag
new_candidates.append({
'p': next_p,
'seq': next_seq,
'rnn_state': next_rnn_state,
'done': done
})
candidates = []
new_candidates.sort(key=lambda x: x['p'], reverse=True)
for idx in range(min(B, len(new_candidates))):
candidates.append(new_candidates[idx])
result = []
for idx in range(min(k, len(candidates))):
d = candidates[idx]
seq = list(map(self.inv_vocab.get, d['seq']))
result.append((utils.join_data(seq, self.text_modeling), d['p']))
return result
def random_sentence(self, prefix=u"", seq_length=100):
prefix = u" " + prefix
prefix = utils.split_raw(prefix, self.text_modeling)
prefix = np.array(list(map(self.vocab.get, prefix)))
current_state = self.model.initial_rnn_state(1)
output = []
for x in prefix[:-1]:
output.append(x)
_, current_state, _ = self.model.sample_output(
self.sess, x, current_state)
x = prefix[-1]
output.append(x)
while len(output) < seq_length:
x, current_state, _ = self.model.sample_output(
self.sess, x, current_state)
output.append(x)
output = list(map(self.inv_vocab.get, output))
output_str = utils.join_data(output, self.text_modeling)
return output_str
def setup_data(npath, ppath):
from ml_data import SequenceDinucLabelsProperties
global data
global limits
global lengths
global shared_input_length
# Join data into a single input vector
data = join_data(SequenceDinucLabelsProperties(npath, ppath))
# Get lenghts of input vectors
lengths = get_input_lengths(data)
# Define limits of data type on single vector
limits = calc_limits(data)
# Compute shared input vector length
shared_input_length = 79
def setup_data(npath, ppath):
from ml_data import SimpleHistData
from ml_data import DinucCrossCovarData
global data
global limits
global lengths
global shared_input_length
# Join data into a single input vector
data = join_data(
SimpleHistData(npath, ppath, k=4, upto=True),
DinucCrossCovarData(npath, ppath, k=3, upto=True),
)
# Get lenghts of input vectors
lengths = get_input_lengths(data)
# Define limits of data type on single vector
limits = calc_limits(data)
# Compute shared input vector length
shared_input_length = limits[-1][-1]
if args.load_dir is not None:
print "Continue from {}".format(args.load_dir)
saver.restore(sess, args.load_dir)
start_time = time.time()
for epoch in range(args.n_epochs):
losses = []
n_batch = train_loader.n_batch(args.batch_size, args.seq_length)
for idx, (batch_x, batch_y) in enumerate(
train_loader.get_batch(args.batch_size, args.seq_length)):
loss = model.run_train_op(sess, train_op, batch_x, batch_y,
model.initial_rnn_state(args.batch_size))
losses.append(loss)
print "Epoch {} ({} / {}), loss: {:.4f}, elapsed time: {:.1f}s".format(
epoch, idx, n_batch, loss,
time.time() - start_time)
writer.add_summary(
sess.run(loss_summary, feed_dict={loss_log: np.mean(losses)}), epoch)
saver.save(sess, args.save_dir)
output, x, current_state = [], train_loader.vocab.get(
unichr(32)), model.initial_rnn_state(1)
for _ in range(100):
x, current_state, _ = model.sample_output(sess, x, current_state)
output.append(x)
output = list(map(train_loader.inv_vocab.get, output))
print output
print utils.join_data(output, args.text_modeling)
