def test_model(self, inputs, outputs, keep_prob=1):
batch_test_inputs, batch_test_seq_len = pad_sequences(
inputs, maxlen=self.max_len)
batch_test_targets = sparse_tuple_from(outputs)
feed = {
self.model_input: batch_test_inputs,
self.model_targets: batch_test_targets,
self.model_seqlen: batch_test_seq_len,
self.keep_prob: keep_prob
}
d = self.sess.run(self.decoded[0], feed_dict=feed)
dense_decoded = tf.sparse_tensor_to_dense(
d, default_value=-1).eval(session=self.sess)
error = self.sess.run(self.ler, feed_dict=feed)
return error
def write_summary_for(self, epoch, inputs_path, outputs_path):
with open(inputs_path, "rb") as fp: # Unpickling
inputs = pickle.load(fp)
with open(outputs_path, "rb") as fp: # Unpickling
outputs = pickle.load(fp)
batch_test_inputs, batch_test_seq_len = pad_sequences(
inputs, maxlen=self.max_len)
batch_test_targets = sparse_tuple_from(outputs)
feed = {
self.model_input: batch_test_inputs,
self.model_targets: batch_test_targets,
self.model_seqlen: batch_test_seq_len,
self.keep_prob: 1
}
_, summ = self.sess.run([self.ler, self.summeries], feed_dict=feed)
self.writer.add_summary(summ, global_step=epoch + 1)
_, summ = self.sess.run([self.cost, self.summeries], feed_dict=feed)
self.writer.add_summary(summ, global_step=epoch + 1)
def print_outputs(self, inputs, outputs, keep_prob=1):
batch_test_inputs, batch_test_seq_len = pad_sequences(
inputs, maxlen=self.max_len)
batch_test_targets = sparse_tuple_from(outputs)
feed = {
self.model_input: batch_test_inputs,
self.model_targets: batch_test_targets,
self.model_seqlen: batch_test_seq_len,
self.keep_prob: keep_prob
}
d = self.sess.run(self.decoded[0], feed_dict=feed)
dense_decoded = tf.sparse_tensor_to_dense(
d, default_value=-1).eval(session=self.sess)
all_originals = []
all_decodeds = []
for i, seq in enumerate(dense_decoded):
seq = [s for s in seq if s != -1]
all_originals.append(self.number_to_lex(outputs[i]))
all_decodeds.append(self.number_to_lex(seq))
return [self.no_delimiters(array) for array in all_originals
], [self.no_delimiters(array) for array in all_decodeds]
def train_model(self,
num_epochs=20,
num_batches=10,
num_batches_test=1,
keep_prob=1,
data_usage=1,
valid_proportion=0.1,
valid_patience=5,
continue_model=False):
last_epoch = 0
if continue_model:
with open(self.save_dir + '/last_epoch', "rb") as fp: # Unpickling
last_epoch = pickle.load(fp) + 1
self.restore_model(epoch=last_epoch - 1)
else:
self.sess.run(tf.global_variables_initializer())
shutil.rmtree(self.graphs_dir)
self.writer = tf.summary.FileWriter(self.graphs_dir)
for curr_epoch in range(last_epoch, num_epochs):
train_cost = train_ler = 0
all_batches_lengths = 0
batches = list(range(0, num_batches))
random.shuffle(batches)
for batch in batches[1:]:
sys.stdout.write("\repoch %d " % int(curr_epoch + 1))
sys.stdout.write("%d%%" %
int(100 * batches.index(batch) / num_batches))
with open(self.inputs_train_path + self.batch_append(batch),
"rb") as fp: # Unpickling
batch_train_inputs = pickle.load(fp)
with open(self.outputs_train_path + self.batch_append(batch),
"rb") as fp: # Unpickling
batch_train_targets = sparse_tuple_from(pickle.load(fp))
batch_train_inputs, batch_train_seq_len = pad_sequences(
batch_train_inputs, maxlen=self.max_len)
feed = {
self.model_input: batch_train_inputs,
self.model_targets: batch_train_targets,
self.model_seqlen: batch_train_seq_len,
self.keep_prob: keep_prob
}
batch_cost, _ = self.sess.run([self.cost, self.optimizer],
feed)
train_cost += batch_cost * len(batch_train_seq_len)
train_ler += self.sess.run(
self.ler, feed_dict=feed) * len(batch_train_seq_len)
all_batches_lengths += len(batch_train_seq_len)
train_cost /= all_batches_lengths
train_ler /= all_batches_lengths
test_ler = self.test_ler(num_batches_test)
log = "Epoch {}/{}, train_cost = {:.3f}, train_ler = {:.3f}, test_ler = {:.3f}"
sys.stdout.write("\r")
print(
log.format(curr_epoch + 1, num_epochs, train_cost, train_ler,
test_ler))
self.write_summary_for(
curr_epoch,
self.inputs_train_path + self.batch_append(batches[0]),
self.outputs_train_path + self.batch_append(batches[0]))
self.save_model(curr_epoch)