def train():
c = constants()
p = params()
x, y = f.inputs([f.TRAIN_TF_RECORDS_FILE], shuffle=True, name='inputs')
u.shape_log(x)
u.shape_log(y)
logits = model(x, c)
logits = tf.reshape(logits, [-1, c.MAX_TIME, f.OUTPUT_DIM])
labels = tf.reshape(y, [-1, c.MAX_TIME, f.OUTPUT_DIM])
loss = tf.reduce_mean(tf.square(logits - labels))
train_op = tf.train.AdamOptimizer(learning_rate=0.05).minimize(
loss, global_step=p.global_step)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
# initialize global variables
sess.run(init_op)
# enable batch fetchers
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
#setup saver
saver = u.saver_ops(sess=sess,
max_to_keep=3,
global_step=p.global_step)
eng = mo.matlab_connection(c.MATLAB_SESSION)
# run the training operation
for indx in range(20000):
_, log = sess.run([train_op, logits])
# logi.append(log)
u.display_update(indx, 25)
if ((indx + 1) % 2000 == 0):
saver.save()
if ((indx) % 5 == 0):
# CHECK EMERGENCY STOP
if eng.should_stop():
print('stopping')
break
# connect to matlab
# eng = mo.matlab_connection(c.MATLAB_SESSION)
# eng.put_var('logits', logi)
saver.save()
coord.request_stop()
coord.join(threads)
def model(X, c):
cells = [rnn.LSTMCell(num_units=c.NUM_UNITS) for _ in range(c.NUM_LAYERS)]
multicell = rnn.MultiRNNCell(cells, state_is_tuple=True)
reshape_op = tf.reshape(X, [-1, c.MAX_TIME, 1], name='reshape_op')
outputs, H = tf.nn.dynamic_rnn(multicell, reshape_op, dtype=tf.float32)
reshape_op = tf.reshape(outputs, [-1, c.NUM_UNITS])
dense_op = u.distributed_dense(reshape_op, units=1, max_time=100)
logits = tf.reshape(dense_op, [c.MAX_TIME])
u.shape_log(logits)
return logits
def train():
c = constants()
p = params()
x, y = f.inputs([f.TRAIN_TF_RECORDS_FILE], name='inputs')
u.shape_log(x)
u.shape_log(y)
logits = model(x, c)
labels = tf.reshape(y, [c.MAX_TIME])
loss = tf.reduce_mean(tf.square(logits - labels))
train_op = tf.train.AdamOptimizer(learning_rate=0.05).minimize(
loss, global_step=p.global_step)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
# initialize global variables
sess.run(init_op)
# enable batch fetchers
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
#setup saver
saver = u.saver_ops(sess=sess,
max_to_keep=3,
global_step=p.global_step)
logi = []
# run the training operation
for indx in range(10000):
_, log = sess.run([train_op, logits])
logi.append(log)
u.display_update(indx, 50)
# connect to matlab
eng = mo.matlab_connection(c.MATLAB_SESSION)
eng.put_var('logits', logi)
saver.save()
coord.request_stop()
coord.join(threads)
def test():
c = constants()
p = params()
# connect to matlab
eng = mo.matlab_connection(c.MATLAB_SESSION)
x, y = f.inputs([f.TEST_TF_RECORDS_FILE], shuffle=True, name='inputs')
u.shape_log(x)
u.shape_log(y)
logits = model(x, c)
logits = tf.reshape(logits, [-1, c.MAX_TIME, f.OUTPUT_DIM])
labels = tf.reshape(y, [-1, c.MAX_TIME, f.OUTPUT_DIM])
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
# initialize global variables
sess.run(init_op)
# enable batch fetchers
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
#setup saver
saver = u.saver_ops(sess=sess,
max_to_keep=3,
global_step=p.global_step)
logi = []
labs = []
# run the training operation
for indx in range(6):
log, lab = sess.run([logits, labels])
logi.append(log)
labs.append(lab)
u.display_update(indx, 250)
eng.put_var('logits', logi)
eng.put_var('labels', labs)
coord.request_stop()
coord.join(threads)
def model(X, c):
cells = [rnn.LSTMCell(num_units=c.NUM_UNITS) for _ in range(c.NUM_LAYERS)]
multicell = rnn.MultiRNNCell(cells, state_is_tuple=True)
reshape_op = tf.reshape(X, [1, 100, 1], name='reshape_op')
outputs, H = tf.nn.dynamic_rnn(multicell, reshape_op, dtype=tf.float32)
u.shape_log(outputs)
# outputs has shape => [batch_size, max_time, num_units]
# reshape it to => [batch_size*max_time, num_units]
# apply a fully connected layer with shared weight and bias
# that reshapes to => [batch_size*max_time, n_outputs]
reshape_op = tf.reshape(outputs, [-1, c.NUM_UNITS])
unit_kern = tf.Variable(tf.random_normal([1, 1]), dtype=tf.float32)
unit_bias = tf.Variable(tf.random_normal([1, 1]), dtype=tf.float32)
kern = tf.matmul(tf.ones([c.NUM_UNITS, 1], dtype=tf.float32), unit_kern)
bias = tf.matmul(tf.ones([100, 1], dtype=tf.float32), unit_bias)
mult_op = tf.matmul(reshape_op, kern, name='multop')
dense_op = tf.add(mult_op, bias, name='addop')
u.shape_log(dense_op)
logits = tf.reshape(dense_op, [100])
u.shape_log(logits)
return logits
def model(X, c):
cells = [rnn.LSTMCell(num_units=c.NUM_UNITS) for _ in range(c.NUM_LAYERS)]
multicell = rnn.MultiRNNCell(cells, state_is_tuple=True)
reshape_op = tf.reshape(X, [-1, c.MAX_TIME, 2], name='reshape_op')
outputs, H = tf.nn.dynamic_rnn(multicell, reshape_op, dtype=tf.float32)
u.shape_log(outputs)
reshape_op = tf.reshape(outputs, [-1, c.NUM_UNITS])
u.shape_log(reshape_op)
# dense_op = u.distributed_dense(reshape_op, units=2, max_time=100)
dense_op = tf.layers.dense(reshape_op, 2, name='distributed__dense')
u.shape_log(dense_op)
w, b = u.get_layer_vars('distributed__dense')
logits = tf.reshape(dense_op, [c.MAX_TIME * f.OUTPUT_DIM])
u.shape_log(logits)
return logits