def __init__(self, data_loader, num_epochs, chckpnt_dir, weights_path,
summary_write_freq, model_save_freq):
self.data_loader = data_loader
self.num_epochs = num_epochs
self.chckpnt_dir = chckpnt_dir
self.weights_path = weights_path
self.summary_write_freq = summary_write_freq
self.model_save_freq = model_save_freq
self.num_train = data_loader.get_num_train()
self.input, self.label, self.loss, self.outputs = model.build_graph()
self.train_op = tf.train.AdamOptimizer(1e-3).minimize(self.loss)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # This is needed otherwise, tensorflow assigns the entire GPU memory for one process
self.saver = tf.train.Saver()
self.sess = tf.InteractiveSession(config=config)
self.summary_writer = tf.summary.FileWriter(self.chckpnt_dir,
self.sess.graph)
self.sess.run(tf.global_variables_initializer())
self.load_weights(self.weights_path) # Load weights from VGG model
ckpt = tf.train.get_checkpoint_state(
self.chckpnt_dir) # Continue from previous checkpoint
if (ckpt and ckpt.model_checkpoint_path):
saver.restore(sess.ckpt.model_checkpoint_path)
def train():
with tf.Session() as sess:
# initialization
graph = build_graph()
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
# multi thread
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# log writer
train_writer = tf.summary.FileWriter(FLAGS.log_dir + '/train', sess.graph)
test_writer = tf.summary.FileWriter(FLAGS.log_dir + '/val')
# restore model
if FLAGS.restore:
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
if ckpt:
saver.restore(sess, ckpt)
print("restore from the checkpoint {}".format(ckpt))
# training begins
try:
while not coord.should_stop():
for step, (x_batch, y_batch) in enumerate(train_data_iterator()):
start_time = time.time()
feed_dict = {graph['images']: x_batch,
graph['labels']: y_batch,
graph['keep_prob']: 0.8,
graph['is_training']: True}
train_opts = [graph['train_op'], graph['loss'], graph['merged_summary_op']]
_, loss_val, train_summary = sess.run(train_opts, feed_dict=feed_dict)
train_writer.add_summary(train_summary, step)
end_time = time.time()
print("the step {0} takes {1} loss {2}".format(step, end_time - start_time, loss_val))
# eval stage
if step % FLAGS.eval_steps == 0:
x_batch_test, y_batch_test = test_data_helper(128)
feed_dict = {graph['images']: x_batch_test,
graph['labels']: y_batch_test,
graph['keep_prob']: 1.0,
graph['is_training']: False}
test_opts = [graph['accuracy'], graph['merged_summary_op']]
accuracy_test, test_summary = sess.run(test_opts, feed_dict=feed_dict)
test_writer.add_summary(test_summary, step)
print('===============Eval a batch=======================')
print('the step {0} test accuracy: {1}'.format(step, accuracy_test))
print('===============Eval a batch=======================')
# save stage
if step % FLAGS.save_steps == 0 and step > FLAGS.min_save_steps:
saver.save(sess, os.path.join(FLAGS.checkpoint_dir, FLAGS.model_name), global_step=step)
except tf.errors.OutOfRangeError:
print('==================Train Finished================')
saver.save(sess, os.path.join(FLAGS.checkpoint_dir, FLAGS.model_name), global_step=step)
finally:
coord.request_stop()
coord.join(threads)
def main():
images, labels = dataset.load_test_images()
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
z, _x = sess.run([t.z_r, t.x_r], feed_dict={t.x: _images})
plot.scatter_labeled_z(z, label_id, dir=config.ckpt_dir)
plot.tile_images(_x[:100], dir=config.ckpt_dir)
plot.plot_loss_tendency(config.ckpt_dir)
hist_value, hist_head = plot.load_pickle_to_data(config.ckpt_dir)
for loss_name in ['reconstruction']:
plot.plot_loss_trace(hist_value[loss_name], loss_name, config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator'],
hist_value['generator'], 'z', config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_z'],
hist_value['generator_z'], 'z',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_img'],
hist_value['generator_img'], 'z',
config.ckpt_dir)
def main():
# load MNIST images
images, labels = dataset.load_test_images()
# Settings
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
representation, x_reconstruction = sess.run([t.yz, t.x_r],
feed_dict={t.x: _images})
plot.scatter_labeled_z(representation, label_id, dir=config.ckpt_dir)
plot.tile_images(x_reconstruction[:100], dir=config.ckpt_dir)
# z distributed plot
num_segments = 20
limit = (-5, 5)
x_values = np.linspace(limit[0], limit[1], num_segments)
y_values = np.linspace(limit[0], limit[1], num_segments)
vacant = np.zeros((28 * num_segments, 28 * num_segments))
for i, x_element in enumerate(x_values):
for j, y_element in enumerate(y_values):
x_reconstruction = sess.run(
t.x_r,
feed_dict={
t.yz: np.reshape([x_element, y_element], [1, 2])
})
vacant[(num_segments - 1 - i) * 28:(num_segments - i) * 28,
j * 28:(j + 1) * 28] = x_reconstruction.reshape(28, 28)
vacant = (vacant + 1) / 2
pylab.figure(figsize=(10, 10), dpi=400, facecolor='white')
pylab.imshow(vacant, cmap='gray', origin='upper')
pylab.tight_layout()
pylab.axis('off')
pylab.savefig("{}/clusters.png".format(config.ckpt_dir))
# loss part
hist_value, hist_head = plot.load_pickle_to_data(config.ckpt_dir)
for loss_name in ['reconstruction', 'supervised']:
plot.plot_loss_trace(hist_value[loss_name], loss_name,
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_y'],
hist_value['generator_y'], 'y',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_z'],
hist_value['generator_z'], 'z',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['validation_accuracy'],
hist_value['transform'],
'validation_accuracy', config.ckpt_dir)
def build_train_model(hparams, scope="train"):
"""Builds a training Seq2Seq model
Args:
hparams: a HParams object
scope: scope of train model
Returns:
model: a NTModel tuple, representing a handle to the model
"""
src_lang = hparams.src_lang
src_vocab_file_name = hparams.src_vocab_file_name
tgt_lang = hparams.tgt_lang
tgt_vocab_file_name = hparams.tgt_vocab_file_name
tf.reset_default_graph()
train_graph = tf.Graph()
with train_graph.as_default() as g:
with tf.container(scope):
src_vocab, tgt_vocab = load_vocabs(src_lang, src_vocab_file_name,
tgt_lang, tgt_vocab_file_name)
src_dataset_file_name = tf.placeholder(
tf.string, name="src_dataset_file_name")
tgt_dataset_file_name = tf.placeholder(
tf.string, name="tgt_dataset_file_name")
src_dataset = tf.data.TextLineDataset(src_dataset_file_name)
tgt_dataset = tf.data.TextLineDataset(tgt_dataset_file_name)
batch_size = tf.placeholder(tf.int64, name="batch_size")
# maximum sequence length for training example
max_len = tf.placeholder(tf.int64, name="max_len")
iterator = Iterator(src_dataset,
src_vocab,
tgt_dataset,
tgt_vocab,
batch_size=batch_size,
max_len=max_len)
# actual TensorFlow Dataset Iterator
iterator_tf = iterator.create_iterator()
model_class = _get_model_from_str_type(hparams.model_name)
model = model_class(hparams, src_vocab, tgt_vocab)
model_graph = model.build_graph(iterator_tf,
tf.contrib.learn.ModeKeys.TRAIN,
batch_size, g)
return NTModel(src_vocab=src_vocab,
tgt_vocab=tgt_vocab,
iterator_tf=iterator_tf,
model_graph=model_graph,
model=model,
hparams=hparams,
mode=tf.contrib.learn.ModeKeys.TRAIN)
def test():
""" Testing a pretrained network """
# First create our Tensorflow graph.
# Start by setting some of our hyperparamters.
num_dimensions = 300
# Load in data structures
wordsList = data.wordsList # encode words as UTF-8 (necessary?)
wordVectors = data.wordVectors
model.build_graph()
saver = tf.train().Saver()
with tf.InteractiveSession as sess:
saver.restore(sess, tf.train().latest_checkpoint('models'))
max_length = util.MAX_SEQ_LENGTH
print('Write a movie review to analyse. Enter to exit. Max length is ', max_length)
while True:
line = _get_user_input()
if len(line) > 0 and line[-1] == '\n':
continue
if line == '':
break
"""
Before we input our own text, let's first define a couple of functions.
The first is a function to make sure the sentence is in the proper format,
and the second is a function that obtains the word vectors for each of the words
in a given sentence.
"""
# Get token-ids for the input sentence
input_matrix = get_sentence_matrix(line)
if len(input_matrix) > max_length:
print('Max length I can handle is:', max_length)
line = _get_user_input()
continue
predicted_sentiment = sess.run(model.prediction, {model.input_data: input_matrix})
if predicted_sentiment[0] > predicted_sentiment[1]:
print("Positive statement")
else:
print("Negative statement")
def main():
# load MNIST images
images, labels = dataset.load_test_images()
# Settings
num_anologies = 10
pylab.gray()
# generate style vector z
x = dataset.sample_unlabeled_data(images, num_anologies)
x = (x + 1) / 2
with tf.device(config.device):
x_input, img_y, img_z, reconstruction = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
z = sess.run(img_z, feed_dict={x_input: x})
for m in range(num_anologies):
pylab.subplot(num_anologies, config.ndim_y + 2, m * 12 + 1)
pylab.imshow(x[m].reshape((28, 28)), interpolation='none')
pylab.axis('off')
all_y = np.identity(config.ndim_y, dtype=np.float32)
for m in range(num_anologies):
fixed_z = np.repeat(z[m].reshape(1, -1), config.ndim_y, axis=0)
gen_x = sess.run(reconstruction,
feed_dict={
img_z: fixed_z,
img_y: all_y
})
gen_x = (gen_x + 1) / 2
for n in range(config.ndim_y):
pylab.subplot(num_anologies, config.ndim_y + 2, m * 12 + 3 + n)
pylab.imshow(gen_x[n].reshape((28, 28)), interpolation='none')
pylab.axis('off')
fig = pylab.gcf()
fig.set_size_inches(num_anologies, config.ndim_y)
pylab.savefig('{}/analogy.png'.format(config.ckpt_dir))
hist_value, hist_head = plot.load_pickle_to_data(config.ckpt_dir)
for loss_name in [
'reconstruction', 'validation_accuracy', 'supervised'
]:
plot.plot_loss_trace(hist_value[loss_name], loss_name,
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_y'],
hist_value['generator_y'], 'y',
config.ckpt_dir)
plot.plot_adversarial_trace(hist_value['discriminator_z'],
hist_value['generator_z'], 'z',
config.ckpt_dir)
def train():
trainset = sys.argv[1]
inputs = melt.read_sparse.inputs
X, y = inputs(trainset,
decode=decode,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
num_preprocess_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
train_with_validation = len(
sys.argv) > 2 and not sys.argv[2].startswith('-')
print('train_with_validation', train_with_validation)
if train_with_validation:
validset = sys.argv[2]
eval_X, eval_y = inputs(
validset,
decode=decode,
batch_size=FLAGS.batch_size * 10,
num_preprocess_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
loss, accuracy = model.build_graph(X, y)
train_op = melt.gen_train_op(loss, FLAGS.learning_rate)
if train_with_validation:
tf.get_variable_scope().reuse_variables()
eval_loss, eval_accuracy = model.build_graph(eval_X, eval_y)
tf.scalar_summary('loss_eval', eval_loss)
eval_ops = [eval_loss, eval_accuracy]
else:
eval_ops = None
train_flow(
[train_op, loss, accuracy],
deal_results=melt.show_precision_at_k,
#deal_results=None,
eval_ops=eval_ops,
deal_eval_results=lambda x: melt.print_results(x,
names=['precision@1']),
print_avg_loss=True,
eval_interval_steps=FLAGS.eval_interval_steps)
def train():
assert FLAGS.num_classes > 0 and FLAGS.num_features > 0, 'you must pass num_classes and num_features according to your data'
print('num_features:', FLAGS.num_features, 'num_classes:', FLAGS.num_classes)
model.set_input_info(num_features=FLAGS.num_features, num_classes=FLAGS.num_classes)
trainset = sys.argv[1]
inputs = melt.shuffle_then_decode.inputs
X, y = inputs(
trainset,
decode=decode,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
num_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
train_with_validation = len(sys.argv) > 2
if train_with_validation:
validset = sys.argv[2]
eval_X, eval_y = inputs(
validset,
decode=decode,
batch_size=FLAGS.batch_size * 10,
num_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
with tf.variable_scope('main') as scope:
loss, accuracy = model.build_graph(X, y)
scope.reuse_variables()
if train_with_validation:
eval_loss, eval_accuracy = model.build_graph(eval_X, eval_y)
eval_ops = [eval_loss, eval_accuracy]
else:
eval_ops = None
melt.apps.train_flow(
[loss, accuracy],
deal_results=melt.show_precision_at_k,
eval_ops=eval_ops,
deal_eval_results= lambda x: melt.print_results(x, names=['precision@1']),
model_dir=FLAGS.model_dir
)
def main():
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
logger.info(config.ckpt_path)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_path))
logger.info("Loading model completely")
z_latent = sampler_switch(config)
d_q = sess.run(t.p_o,
feed_dict={
t.z_e: dp.test.e,
t.x_c: dp.test.c,
t.z_l: z_latent,
t.p_in: dp.test.rd,
})
r_p = sess.run(t.p_i,
feed_dict={
t.x_c: dp.test.c,
t.z_l: z_latent,
t.z_e: dp.test.e,
t.p_in: dp.test.rd
})
# inverse the scaled output
qm, qr, rdm, rdr = dp.out.qm, dp.out.qr, dp.out.rdm, dp.out.rdr
actual_Q = anti_norm(dp.test.q, qm, qr)
result_Q = anti_norm(d_q, qm, qr)
actual_r = anti_norm(dp.test.rd, rdm, rdr)
result_r = anti_norm(r_p, rdm, rdr)
# save the result
ensemble = {
'actual_Q': actual_Q,
'result_Q': result_Q,
'actual_r': actual_r,
'result_r': result_r
}
path = os.path.join(config.logs_path, config.description + '-test.pkl')
pickle_save(ensemble, 'test_result', path)
copy_file(path, config.history_test_path)
# visualize the process
vis.cplot(actual_Q[:, 0], result_Q[:, 0], ['Q1', 'origin', 'modify'],
config.t_p)
vis.cplot(actual_Q[:, 1], result_Q[:, 1], ['Q2', 'origin', 'modify'],
config.t_p)
for num in range(6):
vis.cplot(actual_r[:, num], result_r[:, num],
['R{}'.format(num + 1), 'origin', 'modify'], config.t_p)
def build_eval_model(hparams, scope="eval"):
src_lang = hparams.src_lang
src_vocab_file_name = hparams.src_vocab_file_name
tgt_lang = hparams.tgt_lang
tgt_vocab_file_name = hparams.tgt_vocab_file_name
tf.reset_default_graph()
eval_graph = tf.Graph()
with eval_graph.as_default() as g:
with tf.container(scope):
src_vocab, tgt_vocab = load_vocabs(src_lang, src_vocab_file_name,
tgt_lang, tgt_vocab_file_name)
src_dataset_file_name = tf.placeholder(
tf.string, name="src_dataset_file_name")
tgt_dataset_file_name = tf.placeholder(
tf.string, name="tgt_dataset_file_name")
src_dataset = tf.data.TextLineDataset(src_dataset_file_name)
tgt_dataset = tf.data.TextLineDataset(tgt_dataset_file_name)
batch_size = tf.placeholder(tf.int64, name="batch_size")
# maximum sequence length for training example
max_len = tf.placeholder(tf.int64, name="max_len")
iterator = Iterator(src_dataset,
src_vocab,
tgt_dataset,
tgt_vocab,
batch_size=batch_size,
max_len=max_len)
# actual TensorFlow Dataset Iterator
iterator_tf = iterator.create_iterator(shuffle=False)
model_class = _get_model_from_str_type(hparams.model_name)
model = model_class(hparams, src_vocab, tgt_vocab)
model_graph = model.build_graph(iterator_tf,
tf.contrib.learn.ModeKeys.EVAL,
batch_size, g)
return NTModel(src_vocab=src_vocab,
tgt_vocab=tgt_vocab,
iterator_tf=iterator_tf,
model_graph=model_graph,
model=model,
hparams=hparams,
mode=tf.contrib.learn.ModeKeys.EVAL)
def train():
trainset = sys.argv[1]
inputs = melt.shuffle_then_decode.inputs
X, y = inputs(trainset,
decode=decode,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
num_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
train_with_validation = len(sys.argv) > 2
if train_with_validation:
validset = sys.argv[2]
eval_X, eval_y = inputs(validset,
decode=decode,
batch_size=FLAGS.batch_size * 10,
num_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
loss, accuracy = model.build_graph(X, y)
if train_with_validation:
tf.get_variable_scope().reuse_variables()
eval_loss, eval_accuracy = model.build_graph(eval_X, eval_y)
eval_ops = [eval_loss, eval_accuracy]
else:
eval_ops = None
melt.apps.train_flow(
[loss, accuracy],
deal_results=melt.show_precision_at_k,
eval_ops=eval_ops,
deal_eval_results=lambda x: melt.print_results(x,
names=['precision@1']),
)
def main():
hparams = hyperparameters.hparams
print("************************************")
print("*********** Begin Train ************")
print("************************************")
print("Model: %s" % hparams.model_name)
print("Optimizer: %s" % hparams.optimizer_name)
print("Data directory: %s" % hparams.data_dir)
print("Log directory: %s" % hparams.log_dir)
inputs, loss, train_op = model.build_graph(hparams)
train.run_train(hparams, inputs, loss, train_op)
def main():
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_path))
q_errors = []
r_adjs = []
z_adjs = []
z_true = sess.run(t.z_img, feed_dict={t.x: dp.test.rd})
for time in range(test_times):
z_latent = sampler_switch(config)
q_error = sess.run(t.dq,
feed_dict={
t.z_e: dp.test.e,
t.x_c: dp.test.c,
t.z_l: z_latent,
t.p_in: dp.test.rd,
t.p_t: dp.test.q,
})
r_adj = sess.run(t.x_lat,
feed_dict={
t.x_c: dp.test.c,
t.z_l: z_latent,
t.z_e: dp.test.e,
})
z_adj = sess.run(t.z_img, feed_dict={t.x: r_adj})
q_errors.append(q_error)
r_adjs.append(r_adj)
z_adjs.append(z_adj)
q_errors = (np.array(q_errors) - np.expand_dims(dp.test.e, axis=0))**2
r_adjs = np.array(r_adjs).reshape(-1, config.ndim_x)
z_adjs = np.array(z_adjs).reshape(-1, config.ndim_z)
pickle_save([q_errors, r_adjs, z_adjs, z_true],
["productions", "adjustment", "latent_variables"],
'{}/{}-metric_plus.pkl'.format(config.logs_path,
config.description))
def main():
images, labels = dataset.load_test_images()
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
x, z_respresentation, x_construction = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
z, _x = sess.run([z_respresentation, x_construction],
feed_dict={x: _images})
scatter_labeled_z(z, label_id, dir=config.ckpt_dir)
tile_images(_x[:100], dir=config.ckpt_dir)
plot_loss_tendency(config.ckpt_dir)
def target():
"""
Label the data using a particular model and save the softmax values.
Generates one softmax values per file
"""
flags = parse_flags()
hparams = parse_hparams(flags.hparams)
num_classes = 41
df = pd.read_csv(flags.infer_csv_file)
file_names = df.iloc[:, 0].values
count = 0
sr = 32000
df = pd.read_csv(flags.infer_csv_file)
x, _ = utils_tf._load_dataset(cfg.to_dataset('training'))
generator = utils.fit_scaler(x)
file_names = df.iloc[:, 0].values
with tf.Graph().as_default() as graph:
mel_filt = librosa.filters.mel(sr=32000, n_fft=1024, n_mels=64).T
mel_filt = tf.convert_to_tensor(mel_filt, dtype=tf.float32)
pcm = tf.placeholder(tf.float32, shape=[None], name='input_audio')
model = CleverHansModel(flags.save_model_dir + '.meta', sr, generator,
mel_filt)
saver = model.build_graph(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, flags.save_model_dir)
print(len(file_names))
for i in range(100):
data, _ = utils_tf._preprocess_data(flags.infer_audio_dir,
file_names[i])
l = sess.run([model.get_probs()], feed_dict={pcm: data})
l = np.squeeze(l)
if (l.ndim != 1):
l = np.mean(l, axis=0)
lab = utils_tf._convert_label_name_to_label(df.iloc[i, 1])
if (lab == np.argmax(l)):
count += 1
print(lab, np.argmax(l))
print(count / 100)
def test():
X, y = inputs(sys.argv[1],
decode=decode,
batch_size=FLAGS.batch_size,
num_epochs=1,
num_preprocess_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
loss, accuracy = model.build_graph(X, y)
eval_names = ['loss', 'precision@1']
print('eval_names:', eval_names)
test_flow([loss, accuracy],
names=eval_names,
model_dir=FLAGS.model_path,
num_steps=FLAGS.num_steps,
eval_times=FLAGS.eval_times)
def test():
X, y = inputs(
sys.argv[1],
decode=decode,
batch_size=FLAGS.batch_size,
num_epochs=
0, #for test we will set num_epochs as 1! but now has epoch variable model loading problem, so set 0..
num_preprocess_threads=FLAGS.num_preprocess_threads,
batch_join=FLAGS.batch_join,
shuffle=FLAGS.shuffle)
loss, accuracy = model.build_graph(X, y)
eval_names = ['loss', 'precision@1']
print('eval_names:', eval_names)
test_flow([loss, accuracy],
names=eval_names,
model_dir=FLAGS.model_path,
num_steps=FLAGS.num_steps,
eval_times=FLAGS.eval_times)
def main():
# load MNIST images
images, labels = dataset.load_test_images()
# Settings
num_scatter = len(images)
_images, _, label_id = dataset.sample_labeled_data(images, labels,
num_scatter)
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
# z distributed plot
num_segments = 20
limit = (-5, 5)
x_values = np.linspace(limit[0], limit[1], num_segments)
y_values = np.linspace(limit[0], limit[1], num_segments)
vacant = np.zeros((28 * num_segments, 28 * num_segments))
for i, x_element in enumerate(x_values):
for j, y_element in enumerate(y_values):
x_reconstruction = sess.run(
t.x_r,
feed_dict={
t.yz: np.reshape([x_element, y_element], [1, 2])
})
vacant[(num_segments - 1 - i) * 28:(num_segments - i) * 28,
j * 28:(j + 1) * 28] = x_reconstruction.reshape(28, 28)
vacant = (vacant + 1) / 2
pylab.figure(figsize=(10, 10), dpi=400, facecolor='white')
pylab.imshow(vacant, cmap='gray', origin='upper')
pylab.tight_layout()
pylab.axis('off')
pylab.savefig("{}/clusters.png".format(config.ckpt_dir))
def train():
BATCH_SIZE = 128
y1, y2, y3, ai, bi, ci, train_op, loss = build_graph()
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# saver.restore(sess, "./tmp/model.ckpt")
for i in range(5000):
_, s1, s2, s3, s4, s5, s6, s7 = sess.run(
[train_op, y1, y2, y3, ai, bi, ci, loss])
if i % 100 == 0:
accuracy = get_accuracy(s1, s2, s3, s4, s5, s6, BATCH_SIZE)
print('Step %s: Accuracy of this batch is %s%%. Loss: %s ' %
(i, accuracy * 100, s7))
print('The first number: ', np.round(s1[:10, 0]))
print('The second number:', np.round(s2[:10, 0]))
print('The third number: ', np.round(s3[:10, 0]))
# if i > 1000:
# accuracy = get_accuracy(s1, s2, s3, s4, s5, s6, BATCH_SIZE)
# if(accuracy == 1 and s7<0.06):
# break
checkpoint_path = os.path.join('./tmp', 'model.ckpt')
saver.save(sess, checkpoint_path)
if True:
partial_trees,node_type_size,level_width,method_indexes,parse_tree_depth = load_data()
print 'num of tress',len(partial_trees)
print 'node type size:',node_type_size
print 'max width of every level:',level_width
# print num_sub_trees
# '''
previous_accuracy = 0
not_increase = 0
g = model.build_graph(batch_size=batch_size,state_size=state_size,learning_rate=learning_rate,node_type_size = node_type_size,vector_size=vector_size,level_max_width = level_width,parse_tree_depth=parse_tree_depth,num_classes=num_classes)
sess = tf.Session()
# initialize
sess.run(tf.global_variables_initializer())
# g['saver'].restore(sess,'weight_average_tmp/GCD/weight')
count = 0
while True:
count += 1
print 'EPOCH',count
if not os.path.exists('weight_average/GCD'):
os.makedirs('weight_average/GCD')
if not os.path.exists('weight_average_tmp/GCD'):
os.makedirs('weight_average_tmp/GCD')
import tensorflow as tf
from model import build_graph
from config import FLAGS
from preprocess import get_X, int2text
def predict(img):
size = (FLAGS.image_size, FLAGS.image_size)
image = get_X(img, size)
predict_opts = [graph['predicted_val_top_k'], graph['predicted_index_top_k']]
feed_dict = {graph['images']: [image],
graph['keep_prob']: 1.0,
graph['is_training']: False}
predict_val, predict_index = sess.run(predict_opts, feed_dict=feed_dict)
predict_index = predict_index.flatten()
text = int2text(predict_index[0], FLAGS.wordset)
return text
# when import from other modules, load the model
sess = tf.Session()
graph = build_graph(top_k=1)
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
saver.restore(sess, ckpt)
import os, sys
import tensorflow as tf
import numpy as np
import pylab
sys.path.append(os.path.abspath(os.path.join(os.getcwd(), "../../")))
from model import build_graph, config
with tf.device(config.device):
t = build_graph(is_test=True)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
# z distributed plot
num_segments = 20
xlimit = (-6, 2)
ylimit = (-6, 2)
x_values = np.linspace(xlimit[0], xlimit[1], num_segments)
y_values = np.linspace(ylimit[0], ylimit[1], num_segments)
vacant = np.zeros((28 * num_segments, 28 * num_segments))
z_batch = []
for x_element in x_values:
for y_element in y_values:
z_batch.append([x_element, y_element])
z_batch = np.array(z_batch).reshape(-1, 2)
x_reconstruction = sess.run(t.x_r, feed_dict={t.z_r: z_batch})
x_reconstruction = (x_reconstruction + 1) / 2
# Validation
val_imgs_path = './cac_numpy_data/val_imgs.npy'
val_labs_path = './cac_numpy_data/val_labs.npy'
train_imgs , train_labs, val_imgs , val_labs , test_imgs , test_labs = \
map(np.load , [train_imgs_path , train_labs_path , val_imgs_path , val_labs_path , test_imgs_path , test_labs_path])
test_imgs = test_imgs / 255.
val_imgs = val_imgs / 255.
x_, y_, lr_, is_training, global_step = define_inputs(
shape=[None, 540, 540, 3], n_classes=3) # 1~9 / 10~100 / 100~inf
logits = build_graph(x_=x_,
y_=y_,
is_training=is_training,
aug_flag=True,
actmap_flag=True,
model='vgg_11',
random_crop_resize=540,
bn=True)
sess, saver, summary_writer = sess_start(logs_path='./logs/vgg_11')
pred, pred_cls, cost_op, cost_mean, train_op, correct_pred, accuracy_op = algorithm(
logits, y_, lr_, 'sgd', 'mse')
for i in range(100000):
batch_xs, batch_ys = batch_selector(train_imgs, train_labs, 5, 5, 5, 5)
#print np.shape(batch_xs) , np.shape(batch_ys)
batch_xs = batch_xs / 255.
batch_ys = batch_ys.reshape([-1, 1])
train_fetches = [train_op, accuracy_op, cost_mean, logits]
train_feedDict = {x_: batch_xs, y_: batch_ys, lr_: lr, is_training: True}
_, train_acc, train_loss, train_preds = sess.run(fetches=train_fetches,
def begin_eval(para, repo, train_index, val_index):
methods = np.load('./data/comment_%s_data/comment_method_vector.npy' %
repo)
print 'shape of method vectors', methods.shape
comment_array = np.load('./data/comment_%s_data/comment_array.npy' % repo)
print 'shape of comment array', comment_array.shape
vector_size = len(methods[0])
print 'vector size', vector_size
bodies_array = np.load('./data/body/body_%s/bodies_array.npy' % repo)
print 'shape of bodies array', bodies_array.shape
body_length = len(bodies_array[0])
bodyWordToIndex, bodyIndexToWord = reader._read_comments_word(
'./data/body_words/bodyWordMap.txt')
body_vocab_size = len(bodyWordToIndex)
wordToIndex, indexToWord = reader._read_comments_word(
'./data/comment_words/commentWordMap.txt')
print 'length', len(wordToIndex), len(indexToWord)
num_classes = len(wordToIndex)
#random_index = np.load('./comment_%s_data/random_index_array.npy'%repo)
if para == '3':
state_size = 1
else:
state_size = 512
g = model.build_graph(batch_size, state_size, learning_rate, num_steps,
vector_size, num_classes, body_length,
body_vocab_size)
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
intra_op_parallelism_threads=8))
if para == 1:
g['saver'].restore(sess, 'CR_temp/weight')
else:
#g['saver'].restore(sess,'CR/weight')
g['saver'].restore(sess, 'CR_%s_tab/weight' % repo)
system_strings, model_strings = eval(sess, val_index - train_index,
batch_size, num_steps, num_classes,
methods[train_index:val_index],
comment_array[train_index:val_index],
bodies_array[train_index:val_index],
g, indexToWord, wordToIndex, para)
sess.close()
r.system_filename_pattern = 'a.(\d+).txt'
r.model_filename_pattern = 'a.[A-Z].#ID#.txt'
output = r.convert_and_evaluate()
# print(output)
index = output.find('ROUGE-2 Average_F: ')
value = float(output[index + 19:index + 26])
#output_dict = r.output_to_dict(output)
return value, output
def train(batch_size, lr, epochs, keep_prob, chkp_dir, output_pb):
click.echo(
click.style(
"lr: {}, keep_prob: {}, output pbfile: {}".format(
lr, keep_prob, output_pb),
fg="cyan",
bold=True,
))
cifar10_train = cifar.CIFAR10("./cifar10_data", download=True, train=True)
cifar10_test = cifar.CIFAR10("./cifar10_data", download=True, train=False)
mean = ((cifar10_train.train_data.astype("float32") /
255.0).mean(axis=(0, 1, 2)).tolist())
std = ((cifar10_train.train_data.astype("float32") /
255.0).std(axis=(1, 2)).mean(axis=0).tolist())
cifar10_train.transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomAffine(degrees=0, translate=(0.1, 0.1)),
transforms.ToTensor(),
transforms.Normalize(mean, std),
])
cifar10_test.transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize(mean, std)])
train_loader = torch.utils.data.DataLoader(cifar10_train,
batch_size=batch_size,
shuffle=True,
num_workers=2)
eval_loader = torch.utils.data.DataLoader(cifar10_test,
batch_size=len(cifar10_test),
shuffle=False,
num_workers=2)
graph = tf.Graph()
with graph.as_default():
tf_image_batch = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
tf_labels = tf.placeholder(tf.float32, shape=[None, 10])
tf_keep_prob = tf.placeholder(tf.float32, name="keep_prob")
tf_pred, train_op, tf_total_loss, saver = build_graph(tf_image_batch,
tf_labels,
tf_keep_prob,
lr=lr)
best_acc = 0.0
chkp_cnt = 0
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
for epoch in range(1, epochs + 1):
for i, (img_batch, label_batch) in enumerate(train_loader, 1):
np_img_batch = img_batch.numpy().transpose((0, 2, 3, 1))
np_label_batch = label_batch.numpy()
_ = sess.run(
train_op,
feed_dict={
tf_image_batch: np_img_batch,
tf_labels: one_hot(np_label_batch),
tf_keep_prob: keep_prob,
},
)
if (i % 100) == 0:
img_batch, label_batch = next(iter(eval_loader))
np_img_batch = img_batch.numpy().transpose((0, 2, 3, 1))
np_label_batch = label_batch.numpy()
pred_label = sess.run(
tf_pred,
feed_dict={
tf_image_batch: np_img_batch,
tf_keep_prob: 1.0
},
)
acc = (pred_label
== np_label_batch).sum() / np_label_batch.shape[0]
click.echo(
click.style(
"[epoch {}: {}], accuracy {:0.2f}%".format(
epoch, i, acc * 100),
fg="yellow",
bold=True,
))
if acc >= best_acc:
best_acc = acc
chkp_cnt += 1
click.echo(
click.style(
"[epoch {}: {}] saving checkpoint, {} with acc {:0.2f}%"
.format(epoch, i, chkp_cnt, best_acc * 100),
fg="white",
bold=True,
))
best_chkp = saver.save(sess,
chkp_dir,
global_step=chkp_cnt)
best_graph_def = prepare_meta_graph("{}.meta".format(best_chkp),
output_nodes=[tf_pred.op.name])
with open(output_pb, "wb") as fid:
fid.write(best_graph_def.SerializeToString())
click.echo(
click.style("{} saved".format(output_pb), fg="blue", bold=True))
def main(run_load_from_file=False):
# load MNIST images
images, labels = dataset.load_test_images()
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
max_epoch = 510
num_trains_per_epoch = 500
batch_size_u = 100
# training
with tf.device(config.device):
h = build_graph()
sess_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
saver = tf.train.Saver(max_to_keep=2)
with tf.Session(config=sess_config) as sess:
'''
Load from checkpoint or start a new session
'''
if run_load_from_file:
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
training_epoch_loss, _ = pickle.load(
open(config.ckpt_dir + '/pickle.pkl', 'rb'))
else:
sess.run(tf.global_variables_initializer())
training_epoch_loss = []
# Recording loss per epoch
process = Process()
for epoch in range(max_epoch):
process.start_epoch(epoch, max_epoch)
'''
Learning rate generator
'''
learning_rate = 0.0001
# Recording loss per iteration
sum_loss_reconstruction = 0
sum_loss_discrminator_z = 0
sum_loss_discrminator_img = 0
sum_loss_generator_z = 0
sum_loss_generator_img = 0
process_iteration = Process()
for i in range(num_trains_per_epoch):
process_iteration.start_epoch(i, num_trains_per_epoch)
# Inputs
'''
_l -> labeled
_u -> unlabeled
'''
images_u = dataset.sample_unlabeled_data(images, batch_size_u)
if config.distribution_sampler == 'swiss_roll':
z_true_u = sampler.swiss_roll(batch_size_u, config.ndim_z,
config.num_types_of_label)
elif config.distribution_sampler == 'gaussian_mixture':
z_true_u = sampler.gaussian_mixture(
batch_size_u, config.ndim_z, config.num_types_of_label)
elif config.distribution_sampler == 'uniform_desk':
z_true_u = sampler.uniform_desk(batch_size_u,
config.ndim_z,
radius=2)
elif config.distribution_sampler == 'gaussian':
z_true_u = sampler.gaussian(batch_size_u,
config.ndim_z,
var=1)
elif config.distribution_sampler == 'uniform':
z_true_u = sampler.uniform(batch_size_u,
config.ndim_z,
minv=-1,
maxv=1)
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
# adversarial phase for discriminator_z
images_u_s = dataset.sample_unlabeled_data(
images, batch_size_u)
_, loss_discriminator_z = sess.run([h.opt_dz, h.loss_dz],
feed_dict={
h.x: images_u,
h.z: z_true_u,
h.lr: learning_rate
})
_, loss_discriminator_img = sess.run([h.opt_dimg, h.loss_dimg],
feed_dict={
h.x: images_u,
h.x_s: images_u_s,
h.lr: learning_rate
})
# adversarial phase for generator
_, loss_generator_z = sess.run([h.opt_e, h.loss_e],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
_, loss_generator_img = sess.run([h.opt_d, h.loss_d],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
sum_loss_reconstruction += loss_reconstruction
sum_loss_discrminator_z += loss_discriminator_z
sum_loss_discrminator_img += loss_discriminator_img
sum_loss_generator_z += loss_generator_z
sum_loss_generator_img += loss_generator_img
if i % 1000 == 0:
process_iteration.show_table_2d(
i, num_trains_per_epoch, {
'reconstruction':
sum_loss_reconstruction / (i + 1),
'discriminator_z':
sum_loss_discrminator_z / (i + 1),
'discriminator_img':
sum_loss_discrminator_img / (i + 1),
'generator_z':
sum_loss_generator_z / (i + 1),
'generator_img':
sum_loss_generator_img / (i + 1),
})
average_loss_per_epoch = [
sum_loss_reconstruction / num_trains_per_epoch,
sum_loss_discrminator_z / num_trains_per_epoch,
sum_loss_discrminator_img / num_trains_per_epoch,
sum_loss_generator_z / num_trains_per_epoch,
sum_loss_generator_img / num_trains_per_epoch,
(sum_loss_discrminator_z + sum_loss_discrminator_img) /
num_trains_per_epoch,
(sum_loss_generator_z + sum_loss_generator_img) /
num_trains_per_epoch
]
training_epoch_loss.append(average_loss_per_epoch)
training_loss_name = [
'reconstruction', 'discriminator_z', 'discriminator_img',
'generator_z', 'generator_img', 'discriminator', 'generator'
]
if epoch % 1 == 0:
process.show_bar(
epoch, max_epoch, {
'loss_r': average_loss_per_epoch[0],
'loss_d': average_loss_per_epoch[5],
'loss_g': average_loss_per_epoch[6]
})
plt.scatter_labeled_z(
sess.run(h.z_r, feed_dict={h.x: images[:1000]}),
[int(var) for var in labels[:1000]],
dir=config.log_dir,
filename='z_representation-{}'.format(epoch))
if epoch % 10 == 0:
saver.save(sess,
os.path.join(config.ckpt_dir, 'model_ckptpoint'),
global_step=epoch)
pickle.dump((training_epoch_loss, training_loss_name),
open(config.ckpt_dir + '/pickle.pkl', 'wb'))
def main(run_load_from_file=False):
# load MNIST images
images, labels = dataset.load_train_images()
# config
opt = Operation()
opt.check_dir(config.ckpt_dir, is_restart=False)
opt.check_dir(config.log_dir, is_restart=True)
# setting
max_epoch = 510
num_trains_per_epoch = 500
batch_size_u = 100
# training
with tf.device(config.device):
h = build_graph()
sess_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
saver = tf.train.Saver(max_to_keep=2)
with tf.Session(config=sess_config) as sess:
'''
Load from checkpoint or start a new session
'''
if run_load_from_file:
saver.restore(sess, tf.train.latest_checkpoint(config.ckpt_dir))
training_epoch_loss, _ = pickle.load(
open(config.ckpt_dir + '/pickle.pkl', 'rb'))
else:
sess.run(tf.global_variables_initializer())
training_epoch_loss = []
# Recording loss per epoch
process = Process()
for epoch in range(max_epoch):
process.start_epoch(epoch, max_epoch)
'''
Learning rate generator
'''
learning_rate = opt.ladder_learning_rate(epoch +
len(training_epoch_loss))
# Recording loss per iteration
training_loss_set = []
sum_loss_reconstruction = 0
sum_loss_supervised = 0
sum_loss_discrminator = 0
sum_loss_generator = 0
process_iteration = Process()
for i in range(num_trains_per_epoch):
process_iteration.start_epoch(i, num_trains_per_epoch)
# sample from data distribution
images_u = dataset.sample_unlabeled_data(images, batch_size_u)
# reconstruction_phase
_, loss_reconstruction = sess.run([h.opt_r, h.loss_r],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
z_true_u = sampler.gaussian(batch_size_u,
config.ndim_z,
mean=0,
var=1)
y_true_u = sampler.onehot_categorical(batch_size_u,
config.ndim_y)
# adversarial phase for discriminator
_, loss_discriminator_y = sess.run([h.opt_dy, h.loss_dy],
feed_dict={
h.x: images_u,
h.y: y_true_u,
h.lr: learning_rate
})
_, loss_discriminator_z = sess.run([h.opt_dz, h.loss_dz],
feed_dict={
h.x: images_u,
h.z: z_true_u,
h.lr: learning_rate
})
loss_discriminator = loss_discriminator_y + loss_discriminator_z
# adversarial phase for generator
_, loss_generator_y, loss_generator_z = sess.run(
[h.opt_e, h.loss_gy, h.loss_gz],
feed_dict={
h.x: images_u,
h.lr: learning_rate
})
loss_generator = loss_generator_y + loss_generator_z
training_loss_set.append([
loss_reconstruction,
loss_discriminator,
loss_discriminator_y,
loss_discriminator_z,
loss_generator,
loss_generator_z,
loss_generator_y,
])
sum_loss_reconstruction += loss_reconstruction
sum_loss_discrminator += loss_discriminator
sum_loss_generator += loss_generator
if i % 1000 == 0:
process_iteration.show_table_2d(
i, num_trains_per_epoch, {
'reconstruction': sum_loss_reconstruction /
(i + 1),
'discriminator': sum_loss_discrminator / (i + 1),
'generator': sum_loss_generator / (i + 1),
})
# In end of epoch, summary the loss
average_training_loss_per_epoch = np.mean(
np.array(training_loss_set), axis=0)
# append validation accuracy to the training loss
training_epoch_loss.append(average_training_loss_per_epoch)
loss_name_per_epoch = [
'reconstruction',
'discriminator',
'discriminator_y',
'discriminator_z',
'generator',
'generator_z',
'generator_y',
]
if epoch % 1 == 0:
process.show_bar(
epoch, max_epoch, {
'loss_r': average_training_loss_per_epoch[0],
'loss_d': average_training_loss_per_epoch[1],
'loss_g': average_training_loss_per_epoch[4],
})
plt.tile_images(sess.run(h.x_, feed_dict={h.x: images_u}),
dir=config.log_dir,
filename='x_rec_epoch_{}'.format(
str(epoch).zfill(3)))
if epoch % 10 == 0:
saver.save(sess,
os.path.join(config.ckpt_dir, 'model_ckptpoint'),
global_step=epoch)
pickle.dump((training_epoch_loss, loss_name_per_epoch),
open(config.ckpt_dir + '/pickle.pkl', 'wb'))
eval_graph = tf.Graph()
with train_graph.as_default():
pitch = tf.placeholder(tf.int32, (None, None))
velocity = tf.placeholder(tf.int32, (None, None))
dt = tf.placeholder(tf.int32, (None, None))
duration = tf.placeholder(tf.int32, (None, None))
noteseq = {
"pitch": pitch,
"velocity": velocity,
"dt": dt,
"duration": duration
}
rnn = build_graph(noteseq, config, True, False)
global_step = tf.train.create_global_step()
lr = tf.train.exponential_decay(config.learning_rate, global_step, 8000,
0.989)
opt = tf.train.AdamOptimizer(learning_rate=lr)
grads, var = zip(*opt.compute_gradients(rnn["loss"]))
grads, _ = tf.clip_by_global_norm(grads, 5.0)
train_op = opt.apply_gradients(zip(grads, var), global_step=global_step)
init = tf.global_variables_initializer()
train_saver = tf.train.Saver()
with tf.variable_scope("feature_loss"):
summaries = [
if not os.path.exists(result_path): os.makedirs(result_path)
log_path = os.path.join(result_path, "log.txt")
paths['log_path'] = log_path
utils.get_logger(log_path).info(str(args))
## training model
if args.mode == 'train':
# 引入第二步建立的模型
model = model.BiLSTM_CRF(args,
embeddings,
data.tag2label,
word2id,
paths,
config=config)
# 创建节点，无返回值
model.build_graph()
## hyperparameters-tuning, split train/dev
# dev_data = train_data[:5000]; dev_size = len(dev_data)
# train_data = train_data[5000:]; train_size = len(train_data)
# print("train data: {0}\ndev data: {1}".format(train_size, dev_size))
# model.train(train=train_data, dev=dev_data)
## train model on the whole training data
print("train data: {}".format(len(train_data)))
# 训练
model.train(train=train_data, dev=test_data
) # use test_data as the dev_data to see overfitting phenomena
## testing model
elif args.mode == 'test':