def __init__(self,
num_symbols,
num_embed_units,
num_units,
num_layers,
num_labels,
embed,
learning_rate=0.5,
max_gradient_norm=5.0,
model='LSTM'):
#todo: implement placeholders
self.texts = tf.placeholder(dtype=tf.string,
shape=[None, None]) # shape: batch*len
self.texts_length = tf.placeholder(dtype=tf.int32,
shape=None) # shape: batch
self.labels = tf.placeholder(dtype=tf.int64,
shape=None) # shape: batch
self.keep_prob = tf.placeholder(dtype=tf.float32)
self.symbol2index = MutableHashTable(key_dtype=tf.string,
value_dtype=tf.int64,
default_value=UNK_ID,
shared_name="in_table",
name="in_table",
checkpoint=True)
# build the vocab table (string to index)
# initialize the training process
self.learning_rate = tf.Variable(float(learning_rate),
trainable=False,
dtype=tf.float32)
self.global_step = tf.Variable(0, trainable=False)
self.epoch = tf.Variable(0, trainable=False)
self.epoch_add_op = self.epoch.assign(self.epoch + 1)
self.index_input = self.symbol2index.lookup(self.texts) # batch*len
# build the embedding table (index to vector)
if embed is None:
# initialize the embedding randomly
self.embed = tf.get_variable('embed',
[num_symbols, num_embed_units],
tf.float32)
else:
# initialize the embedding by pre-trained word vectors
self.embed = tf.get_variable('embed',
dtype=tf.float32,
initializer=embed)
self.embed_input = tf.nn.embedding_lookup(
self.embed, self.index_input) #batch*len*embed_unit
#todo: implement unfinished networks
if num_layers == 1:
if model == 'LSTM':
cell = BasicLSTMCell(num_units)
elif model == 'RNN':
cell = BasicRNNCell(num_units)
elif model == 'GRU':
cell = GRUCell(num_units)
else:
print("Wrong model!")
return
cell_dr = tf.nn.rnn_cell.DropoutWrapper(
cell, input_keep_prob=1.0, output_keep_prob=self.keep_prob)
outputs, states = dynamic_rnn(cell_dr,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="rnn")
if model == 'LSTM':
h_state = states[0]
else:
h_state = states
else:
if model == 'LSTM':
cell = BasicLSTMCell(num_units)
elif model == 'RNN':
cell = BasicRNNCell(num_units)
elif model == 'GRU':
cell = GRUCell(num_units)
else:
print("Wrong model!")
return
cell_dr = tf.nn.rnn_cell.DropoutWrapper(
cell, input_keep_prob=1.0, output_keep_prob=self.keep_prob)
multi_cell = tf.contrib.rnn.MultiRNNCell([cell_dr] * num_layers,
state_is_tuple=True)
init_state = multi_cell.zero_state(16, tf.float32)
outputs, state = tf.nn.dynamic_rnn(multi_cell,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="rnn",
initial_state=init_state,
time_major=False)
h_state = outputs[:, -1, :]
logits = tf.layers.dense(h_state, num_labels)
self.loss = tf.reduce_sum(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels,
logits=logits),
name='loss')
mean_loss = self.loss / tf.cast(tf.shape(self.labels)[0],
dtype=tf.float32)
predict_labels = tf.argmax(logits, 1, 'predict_labels')
self.accuracy = tf.reduce_sum(tf.cast(
tf.equal(self.labels, predict_labels), tf.int32),
name='accuracy')
self.params = tf.trainable_variables()
# calculate the gradient of parameters
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
gradients = tf.gradients(mean_loss, self.params)
clipped_gradients, self.gradient_norm = tf.clip_by_global_norm(
gradients, max_gradient_norm)
self.update = opt.apply_gradients(zip(clipped_gradients, self.params),
global_step=self.global_step)
tf.summary.scalar('loss/step', self.loss)
for each in tf.trainable_variables():
tf.summary.histogram(each.name, each)
self.merged_summary_op = tf.summary.merge_all()
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=3,
pad_step_number=True,
keep_checkpoint_every_n_hours=1.0)
def __init__(self,
num_symbols, # 18430, vocabulary size.
num_embed_units, # 300, Size of word embedding.
num_units, # 512, Size of each model layer.
num_layers, # 1, Number of layers in the model.
num_labels, # 5, Number of labels.
embed, # (18430, 300), word2vector list.
learning_rate=0.5,
max_gradient_norm=5.0):
# todo: implement placeholders
self.texts = tf.placeholder(dtype=tf.string, shape=[None, None], name='texts') # shape: batch*len
self.texts_length = tf.placeholder(dtype=tf.int64, shape=[None], name='texts_length') # shape: batch
self.labels = tf.placeholder(dtype=tf.int64, shape=[None], name='labels') # shape: batch
self.symbol2index = MutableHashTable(
key_dtype=tf.string,
value_dtype=tf.int64,
default_value=UNK_ID,
shared_name="in_table",
name="in_table",
checkpoint=True)
# build the vocab table (string to index)
# initialize the training process
self.learning_rate = tf.Variable(float(learning_rate), trainable=False, dtype=tf.float32)
self.global_step = tf.Variable(0, trainable=False)
self.epoch = tf.Variable(0, trainable=False)
self.epoch_add_op = self.epoch.assign(self.epoch + 1)
self.index_input = self.symbol2index.lookup(self.texts) # batch*len
# build the embedding table (index to vector)
if embed is None:
# initialize the embedding randomly
self.embed = tf.get_variable('embed', [num_symbols, num_embed_units], tf.float32)
else:
# initialize the embedding by pre-trained word vectors
self.embed = tf.get_variable('embed', dtype=tf.float32, initializer=embed)
self.embed_input = tf.nn.embedding_lookup(self.embed, self.index_input) #batch*len*embed_unit
if num_layers == 1:
# cell = BasicLSTMCell(num_units)
cell = GRUCell(num_units)
# cell = BasicRNNCell(num_units)
keep_prob = 0.95
dropped_input = tf.nn.dropout(self.embed_input, keep_prob=keep_prob)
outputs, states = dynamic_rnn(cell, dropped_input, self.texts_length, dtype=tf.float32, scope="rnn")
# todo: implement unfinished networks
# logits = tf.layers.dense(inputs=states, units=num_labels)
l1 = tf.nn.dropout(states, keep_prob=keep_prob)
inner_layer = tf.layers.dense(inputs=l1, units=256, activation=tf.nn.relu)
l2 = tf.nn.dropout(inner_layer, keep_prob=keep_prob)
logits = tf.layers.dense(inputs=l2, units=num_labels)
self.loss = tf.reduce_sum(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels, logits=logits), name='loss')
mean_loss = self.loss / tf.cast(tf.shape(self.labels)[0], dtype=tf.float32)
predict_labels = tf.argmax(logits, 1, 'predict_labels')
self.accuracy = tf.reduce_sum(tf.cast(tf.equal(self.labels, predict_labels), tf.int32), name='accuracy')
self.params = tf.trainable_variables()
# calculate the gradient of parameters
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
gradients = tf.gradients(mean_loss, self.params)
clipped_gradients, self.gradient_norm = tf.clip_by_global_norm(gradients, max_gradient_norm)
self.update = opt.apply_gradients(zip(clipped_gradients, self.params), global_step=self.global_step)
tf.summary.scalar('loss/step', self.loss)
for each in tf.trainable_variables():
tf.summary.histogram(each.name, each)
self.merged_summary_op = tf.summary.merge_all()
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=3, pad_step_number=True, keep_checkpoint_every_n_hours=1.0)
def __init__(self,
num_symbols,
num_embed_units,
num_units,
num_layers,
num_labels,
embed,
learning_rate=0.5,
max_gradient_norm=5.0,
keep_prob=1.,
weight_decay=1e-10,
RNN_type="BasicRNN"):
#todo: implement placeholders
self.texts = tf.placeholder(dtype = tf.string, shape = [None, None])
self.texts_length = tf.placeholder(dtype = tf.int32, shape = [None])
self.labels = tf.placeholder(dtype = tf.int64, shape = [None])
'''
self.texts = tf.placeholder() # shape: batch*len
self.texts_length = tf.placeholder() # shape: batch
self.labels = tf.placeholder() # shape: batch
'''
self.symbol2index = MutableHashTable(
key_dtype=tf.string,
value_dtype=tf.int64,
default_value=UNK_ID,
shared_name="in_table",
name="in_table",
checkpoint=True)
# build the vocab table (string to index)
# initialize the training process
self.learning_rate = tf.Variable(float(learning_rate), trainable=False, dtype=tf.float32)
self.weight_decay = tf.Variable(float(weight_decay), trainable=False, dtype=tf.float32)
self.keep_prob = tf.Variable(float(keep_prob), trainable=False, dtype=tf.float32)
self.global_step = tf.Variable(0, trainable=False)
self.epoch = tf.Variable(0, trainable=False)
self.epoch_add_op = self.epoch.assign(self.epoch + 1)
self.index_input = self.symbol2index.lookup(self.texts) # batch*len
# build the embedding table (index to vector)
if embed is None:
# initialize the embedding randomly
self.embed = tf.get_variable('embed', [num_symbols, num_embed_units], tf.float32)
else:
# initialize the embedding by pre-trained word vectors
self.embed = tf.get_variable('embed', dtype=tf.float32, initializer=embed)
self.embed_input = tf.nn.embedding_lookup(self.embed, self.index_input) #batch*len*embed_unit
# bi-LSTM
with tf.variable_scope("foward_cell"):
#fw_cell = tf.contrib.rnn.GRUCell(num_units)
if RNN_type == "LSTM":
fw_cell = BasicLSTMCell(num_units)
else:
fw_cell = GRUCell(num_units)
'''
fw_cell = tf.contrib.rnn.GRUCell(num_units)
fw_cell = tf.contrib.rnn.GRUCell(num_units)
'''
with tf.variable_scope("barkward_cell"):
#bw_cell = tf.contrib.rnn.GRUCell(num_units)
if RNN_type == "LSTM":
bw_cell = BasicLSTMCell(num_units)
else:
bw_cell = GRUCell(num_units)
'''
if num_layers == 1:
if RNN_type == "BasicRNN":
cell = BasicRNNCell(num_units)
# cell = tf.contrib.rnn.BasicRNNCell(num_units)
elif RNN_type == "GRU":
cell = GRUCell(num_units)
elif RNN_type == "LSTM":
cell = BasicLSTMCell(num_units)
outputs, states = dynamic_rnn(cell, self.embed_input, self.texts_length, dtype=tf.float32, scope="rnn")
'''
outputs, states = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, self.embed_input, self.texts_length, dtype = tf.float32, scope = "bi_lstm")
#print "***state: ", states
#self.y0 = tf.reduce_max(outputs, axis = 1)
#self.y0 = tf.reduce_max(outputs[0] + outputs[1], axis = 1)
#self.y0 = tf.reduce_sum(states, axis = 0)
self.y0 = states[0][1] + states[1][1]
#print "****** y0:", self.y0
self.y0_dp = tf.nn.dropout(self.y0, keep_prob = self.keep_prob)
self.y1 = tf.layers.dense(inputs = self.y0_dp, units = 128, activation = tf.nn.sigmoid)
self.y2 = tf.layers.dense(inputs = self.y0_dp, units = num_labels)
logits = self.y2
'''
self.W1 = tf.Variable(tf.truncated_normal(stddev = .1, shape = [num_units, 128]))
self.b1 = tf.Variable(tf.constant(.1, shape = [128]))
self.u1 = tf.matmul(self.y0_dp, self.W1) + self.b1
self.y1 = tf.nn.sigmoid(self.u1)
self.W2 = tf.Variable(tf.truncated_normal(stddev = .1, shape = [128, 5]))
self.b2 = tf.Variable(tf.constant(.1, shape = [5]))
self.u2 = tf.matmul(self.y1, self.W2) + self.b2
'''
# logits = tf.layers.dense(inputs = self.y1, units = 5)
# logits = self.u2
#todo: implement unfinished networks
with tf.name_scope("l2_loss"):
vars = tf.trainable_variables()
self.lossL2 = tf.add_n([ tf.nn.l2_loss(v) for v in vars ]) * self.weight_decay
self.loss = tf.reduce_sum(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels, logits=logits), name='loss') + self.lossL2
mean_loss = self.loss / tf.cast(tf.shape(self.labels)[0], dtype=tf.float32)
predict_labels = tf.argmax(logits, 1, 'predict_labels')
self.accuracy = tf.reduce_sum(tf.cast(tf.equal(self.labels, predict_labels), tf.int32), name='accuracy')
self.params = tf.trainable_variables()
# calculate the gradient of parameters
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
#opt = tf.train.AdamOptimizer(self.learning_rate)
gradients = tf.gradients(mean_loss, self.params)
clipped_gradients, self.gradient_norm = tf.clip_by_global_norm(gradients, max_gradient_norm)
self.train_op = opt.apply_gradients(zip(clipped_gradients, self.params), global_step=self.global_step)
#self.train_op = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss, global_step=self.global_step,var_list=self.params)
tf.summary.scalar('loss/step', self.loss)
for each in tf.trainable_variables():
tf.summary.histogram(each.name, each)
self.merged_summary_op = tf.summary.merge_all()
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=3, pad_step_number=True, keep_checkpoint_every_n_hours=1.0)
def __init__(self,
num_symbols,
num_embed_units,
num_units,
num_layers,
num_labels,
embed,
learning_rate=0.5,
max_gradient_norm=5.0):
#todo: implement placeholders
self.texts = tf.placeholder(tf.string, [None, None],
name="texts") # shape: batch*len
self.texts_length = tf.placeholder(tf.int64, [None],
name="texts_length") # shape: batch
self.labels = tf.placeholder(tf.int64, [None],
name="labels") # shape: batch
self.symbol2index = MutableHashTable(key_dtype=tf.string,
value_dtype=tf.int64,
default_value=UNK_ID,
shared_name="in_table",
name="in_table",
checkpoint=True)
# build the vocab table (string to index)
# initialize the training process
self.learning_rate = tf.Variable(float(learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_factor = 0.9
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
self.global_step = tf.Variable(0, trainable=False)
self.epoch = tf.Variable(0, trainable=False)
self.epoch_add_op = self.epoch.assign(self.epoch + 1)
self.index_input = self.symbol2index.lookup(self.texts) # batch*len
# build the embedding table (index to vector)
if embed is None:
# initialize the embedding randomly
self.embed = tf.get_variable('embed',
[num_symbols, num_embed_units],
tf.float32)
else:
# initialize the embedding by pre-trained word vectors
self.embed = tf.get_variable('embed',
dtype=tf.float32,
initializer=embed)
self.embed_input = tf.nn.embedding_lookup(
self.embed, self.index_input) #batch*len*embed_unit
model = 'lstm'
if num_layers == 1:
if (model == 'rnn'):
cell = BasicRNNCell(num_units)
elif (model == 'gru'):
cell = GRUCell(num_units)
elif (model == 'lstm'):
cell = BasicLSTMCell(num_units)
cell_do = tf.nn.rnn_cell.DropoutWrapper(
cell, input_keep_prob=1.0, output_keep_prob=FLAGS.keep_prob)
outputs, states = dynamic_rnn(cell_do,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="rnn")
#todo: implement unfinished networks
outputs_flat = tf.reduce_mean(outputs, 1)
if (model == 'lstm'):
states = states[0]
# W_f = weight_variable([tf.app.flags.FLAGS.units, 5])
# b_f = bias_variable([5])
# logits = tf.matmul(outputs_flat, W_f) + b_f
# fc_layer = tf.layers.dense(inputs = states, units = 32, activation = tf.nn.relu)
logits = tf.layers.dense(inputs=states, units=5, activation=None)
else:
self.reverse_texts = tf.placeholder(
tf.string, [None, None],
name="reverse_texts") # shape: batch*len
self.index_reverse_input = self.symbol2index.lookup(
self.reverse_texts)
self.embed_reverse_input = tf.nn.embedding_lookup(
self.embed, self.index_reverse_input) #batch*len*embed_unit
if (model == 'rnn'):
cell1 = BasicRNNCell(num_units)
cell2 = BasicRNNCell(num_units)
elif (model == 'gru'):
cell1 = GRUCell(num_units)
cell2 = GRUCell(num_units)
elif (model == 'lstm'):
cell1 = BasicLSTMCell(num_units)
cell2 = BasicLSTMCell(num_units)
cell1_do = tf.nn.rnn_cell.DropoutWrapper(
cell1, input_keep_prob=1.0, output_keep_prob=FLAGS.keep_prob)
cell2_do = tf.nn.rnn_cell.DropoutWrapper(
cell2, input_keep_prob=1.0, output_keep_prob=FLAGS.keep_prob)
outputs1, states1 = dynamic_rnn(cell1_do,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="rnn")
outputs2, states2 = dynamic_rnn(cell2_do,
self.embed_reverse_input,
self.texts_length,
dtype=tf.float32,
scope="rnn")
if (model == 'lstm'):
states = states1[0] + states2[0]
else:
states = states1 + states2
# fc_layer = tf.layers.dense(inputs = states, units = 32, activation = tf.nn.relu)
logits = tf.layers.dense(inputs=states, units=5, activation=None)
self.loss = tf.reduce_sum(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels,
logits=logits),
name='loss')
mean_loss = self.loss / tf.cast(tf.shape(self.labels)[0],
dtype=tf.float32)
predict_labels = tf.argmax(logits, 1, 'predict_labels')
self.accuracy = tf.reduce_sum(tf.cast(
tf.equal(self.labels, predict_labels), tf.int32),
name='accuracy')
self.params = tf.trainable_variables()
# calculate the gradient of parameters
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
# opt = tf.train.AdamOptimizer(self.learning_rate)
gradients = tf.gradients(mean_loss, self.params)
clipped_gradients, self.gradient_norm = tf.clip_by_global_norm(
gradients, max_gradient_norm)
self.update = opt.apply_gradients(zip(clipped_gradients, self.params),
global_step=self.global_step)
tf.summary.scalar('loss/step', self.loss)
for each in tf.trainable_variables():
tf.summary.histogram(each.name, each)
self.merged_summary_op = tf.summary.merge_all()
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=3,
pad_step_number=True,
keep_checkpoint_every_n_hours=1.0)
def __init__(self,
num_symbols,
num_embed_units,
num_units,
num_layers,
num_labels,
embed,
learning_rate=0.001,
max_gradient_norm=5.0,
learning_rate_decay_factor=0.1):
#todo: implement placeholders
# PROBLEMS REMAIN
self.texts = tf.placeholder(dtype=tf.string,
shape=[None, None]) # shape: batch*len
self.texts_length = tf.placeholder(dtype=tf.int32,
shape=[None]) # shape: batch
self.labels = tf.placeholder(dtype=tf.int32,
shape=[None]) # shape: batch
self.output_keep_prob = tf.placeholder(dtype=tf.float32, shape=[])
self.symbol2index = MutableHashTable(key_dtype=tf.string,
value_dtype=tf.int64,
default_value=UNK_ID,
shared_name="in_table",
name="in_table",
checkpoint=True)
# build the vocab table (string to index)
# initialize the training process
self.learning_rate = tf.Variable(float(learning_rate),
trainable=False,
dtype=tf.float32)
self.learning_rate_update_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
self.global_step = tf.Variable(0, trainable=False)
self.epoch = tf.Variable(0, trainable=False)
self.epoch_add_op = self.epoch.assign(self.epoch + 1)
self.index_input = self.symbol2index.lookup(self.texts) # batch*len
# build the embedding table (index to vector)
if embed is None:
# initialize the embedding randomly
self.embed = tf.get_variable('embed',
[num_symbols, num_embed_units],
tf.float32)
else:
# initialize the embedding by pre-trained word vectors
self.embed = tf.get_variable('embed',
dtype=tf.float32,
initializer=embed)
self.embed_input = tf.nn.embedding_lookup(
self.embed, self.index_input) #batch*len*embed_unit
if num_layers == 1:
# basic rnn
# cell = BasicRNNCell(num_units)
# outputs, states = dynamic_rnn(cell, self.embed_input, self.texts_length, dtype=tf.float32, scope="rnn")
# gru
# cell = GRUCell(num_units)
# outputs, states = dynamic_rnn(cell, self.embed_input, self.texts_length, dtype=tf.float32, scope="rnn")
# lstm
# cell = BasicLSTMCell(num_units)
# outputs, states = dynamic_rnn(cell, self.embed_input, self.texts_length, dtype=tf.float32, scope="rnn")
# states = states[1]
# final model
cell = tf.nn.rnn_cell.DropoutWrapper(
BasicLSTMCell(num_units),
output_keep_prob=self.output_keep_prob)
cell_bw = tf.nn.rnn_cell.DropoutWrapper(
BasicLSTMCell(num_units),
output_keep_prob=self.output_keep_prob)
outputs, states = bidirectional_dynamic_rnn(cell,
cell_bw,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="rnn")
states = states[0][1] + states[1][1]
else:
cells = []
cells_bw = []
for _ in range(num_layers):
cell = tf.nn.rnn_cell.DropoutWrapper(
GRUCell(num_units), output_keep_prob=output_keep_prob)
cells.append(cell)
cell_bw = tf.nn.rnn_cell.DropoutWrapper(
GRUCell(num_units), output_keep_prob=output_keep_prob)
cells_bw.append(cell_bw)
cell = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True)
cell_bw = tf.contrib.rnn.MultiRNNCell(cells_bw,
state_is_tuple=True)
outputs, states = bidirectional_dynamic_rnn(cell,
cell_bw,
self.embed_input,
self.texts_length,
dtype=tf.float32,
scope="stacked_rnn")
states = states[0][num_layers - 1] + states[1][num_layers - 1]
#todo: implement unfinished networks
self.w1 = tf.Variable(
tf.random_normal(shape=[num_units, num_labels],
stddev=tf.sqrt(2.0 / (num_units + num_labels))))
self.b1 = tf.Variable(tf.constant(0.0, shape=[num_labels]))
logits = tf.matmul(states, self.w1) + self.b1
self.loss = tf.reduce_sum(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels,
logits=logits),
name='loss')
mean_loss = self.loss / tf.cast(tf.shape(self.labels)[0],
dtype=tf.float32)
self.predict_labels = tf.argmax(logits,
1,
'predict_labels',
output_type=tf.int32)
self.accuracy = tf.reduce_sum(tf.cast(
tf.equal(self.labels, self.predict_labels), tf.int32),
name='accuracy')
self.params = tf.trainable_variables()
# calculate the gradient of parameters
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
gradients = tf.gradients(mean_loss, self.params)
clipped_gradients, self.gradient_norm = tf.clip_by_global_norm(
gradients, max_gradient_norm)
self.update = opt.apply_gradients(zip(clipped_gradients, self.params),
global_step=self.global_step)
tf.summary.scalar('loss/step', self.loss)
for each in tf.trainable_variables():
tf.summary.histogram(each.name, each)
self.merged_summary_op = tf.summary.merge_all()
self.saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=3,
pad_step_number=True,
keep_checkpoint_every_n_hours=1.0)