def _train(flag, input_shape=[-1,250], expand_layer=[3,20],
layers=[250,50,1], learning_rate=0.1, momentum=0.8,
n_epoch=100, block_bytes=3*50*8):
"""Train Synonym extraction model
Args:
flag: string, format like _$(dim)_$(window)_$(min_count), identifier
input_shape: list of int, specify the shape of placeholder
expand_layer: list of int, specify the feature expand layer.
For example, if value is [3, 20], that means expand
feature from 3 to 20 with a 3*20 matrix
layers: list of int, specify the number of nerons in the MLP.For
example, if value is [250, 50, 1], that means input layer is
batch*250, hidden layer is 250*50, output layer is 50*1.
"""
with tf.variable_scope('var%s' % flag):
data_dir = "%s/data" % os.path.dirname(os.getcwd())
model_dir = "%s/model" % data_dir
this_dir = "%s/%s" % (model_dir, flag)
if not os.path.exists(model_dir):
os.mkdir(model_dir)
if not os.path.exists(this_dir):
os.mkdir(this_dir)
f_train_x = flag.join(['../data/temp/train_X','.bin'])
f_train_y = flag.join(['../data/temp/train_y','.txt'])
f_test_x = flag.join(['../data/temp/test_X','.bin'])
f_test_y = flag.join(['../data/temp/test_y','.txt'])
placeholder_shape = list(input_shape)
placeholder_shape[0] = None
print '...building graph%s' % flag
sess = tf.Session(config=tf.ConfigProto(
intra_op_parallelism_threads=4))
#Build graph
if len(input_shape) == 2:
model = Model.MLP(momentum=momentum,
input_shape=placeholder_shape, layers=layers)
elif len(input_shape) == 3:
model = Model.MLPwL(momentum=momentum,
input_shape=placeholder_shape,
expand_layer=expand_layer, layers=layers)
else:
raise ValueError
# Build operations
x_placeholder = model.x_placeholder
y_placeholder = model.y_placeholder
pred = model.output(x_placeholder)
loss = model.loss_function(pred, y_placeholder)
train_op = model.train_op(loss)
# Builde saver
saver = tf.train.Saver()
tf.add_to_collection('output', pred)
tf.add_to_collection('loss', loss)
tf.add_to_collection('input', x_placeholder)
# Train model
sess.run(tf.initialize_all_variables())
print '...start training %s' % flag
t0 = time.time()
for epoch in xrange(n_epoch):
gen = Model.data_generator(f_train_x, f_train_y, block_bytes)
batch = list(islice(gen, 50))
while(len(batch) > 0):
batch_x = np.array([l[0] for l in batch])
batch_x = np.reshape(batch_x, input_shape)
batch_y = np.array([l[1] for l in batch])
feed_dict = {x_placeholder:batch_x,
y_placeholder:batch_y}
ls, _ = sess.run([loss, train_op], feed_dict=feed_dict)
batch = list(islice(gen, 50))
t1 = time.time()
message = ('epoch: %d, loss: %.5f, elapsed time: %.5f'
% (epoch + 1, ls, t1 - t0))
_print_bar(epoch + 1, n_epoch, message)
# Print Model Report
gen = Model.data_generator(f_test_x, f_test_y,
block_bytes=block_bytes)
batch = list(islice(gen, 50000))
batch_x = np.array([l[0] for l in batch])
batch_x = np.reshape(batch_x, input_shape)
batch_y = np.array([l[1] for l in batch])
x = sess.run(pred, {x_placeholder:batch_x})
x[x > 0] = 1
x[x < 0] = -1
precision, recall, f1 = Model.metric(x, batch_y)
print '\n'
print 'precision: %f, recall: %f, f1 score: %f' % (precision,
recall, f1)
# Saving graph and architecture and variables
graph_def = sess.graph_def
saver.save(sess, flag.join(['%s/model' % this_dir,'']))
saver_def = saver.saver_def
filename = flag.join(['%s/model' % this_dir,'.meta'])
collection_list = ['output', 'loss', 'input']
tf.train.export_meta_graph(filename=filename,
graph_def=graph_def,
saver_def=saver_def,
collection_list=collection_list)
os.system('rm -rf ../data/temp')
return model_dir
