gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
sess = tf.InteractiveSession(config=tf.ConfigProto(
gpu_options=gpu_options))
logger.info('Creating model')
vocab_size = embeddings.shape[0]
embedding_size = embeddings.shape[1]
# test
print("embeddings size: ", vocab_size, " , ", embedding_size)
if args.model == 'mlp':
model = MultiFeedForwardClassifier(args.num_units,
2,
vocab_size,
embedding_size,
use_intra_attention=args.use_intra,
training=True,
project_input=args.no_project,
optimizer=args.optim)
else:
model = LSTMClassifier(args.num_units,
2,
vocab_size,
embedding_size,
training=True,
project_input=args.no_project,
optimizer=args.optim)
model.initialize(sess, embeddings)
# this assertion is just for type hinting for the IDE
"""
Plot weights in a RTE FF model
"""
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('model',
help='Directory containing the '
'trained model')
parser.add_argument('-o',
help='Output directory for images '
'(default is same as the model)',
dest='output')
args = parser.parse_args()
output = args.output or args.model
sess = tf.InteractiveSession()
model = MultiFeedForwardClassifier.load(args.model, sess)
for var in tf.trainable_variables():
if 'bias' in var.name:
continue
# create valid filenames
name = var.name.replace(':0', '').replace('/', '_')
values = var.eval()
pl.matshow(values)
pl.colorbar()
path = os.path.join(output, name)
pl.savefig(path)
train_data.sentences2.shape))
logger.debug(msg.format('Validation',
valid_data.sentences1.shape,
valid_data.sentences2.shape))
logger.debug(msg.format('Test',
test_data.sentences1.shape,
test_data.sentences2.shape))
sess = tf.InteractiveSession()
logger.info('Creating model')
vocab_size = embeddings.shape[0]
embedding_size = embeddings.shape[1]
model = MultiFeedForwardClassifier(args.num_units,
3, vocab_size, embedding_size, max_len,
use_intra_attention=args.use_intra,
training=True, learning_rate=args.rate,
clip_value=args.clip_norm, l2_constant=args.l2,
project_input=args.no_project, mode=args.mode)
model.initialize(sess, embeddings)
if args.load:
vars = tf.contrib.framework.list_variables(args.load)
# with tf.Graph().as_default(), tf.Session().as_default() as sess:
# new_vars = []
#for name, shape in vars:
# print(name)
# v = tf.contrib.framework.load_variable(args.load, name)
# new_vars.append(tf.Variable(v, name = name.replace('biases', 'bias')))
# saver = tf.train.Saver(new_vars)
# sess.run(tf.global_variables_initializer())
# saver.save(sess, args.load + '/new-model')
embedding_size = embeddings.shape[1]
if is_really_cont:
model_class = utils.get_model_class(params)
model, saver = model_class.load(warmup_model,
sess,
training=True,
embeddings=embeddings)
else:
if args.model == 'mlp':
model = MultiFeedForwardClassifier(
args.num_units,
3,
vocab_size,
embedding_size,
use_intra_attention=args.use_intra,
training=True,
project_input=args.no_project,
optimizer=args.optim)
else:
model = LSTMClassifier(args.num_units,
3,
vocab_size,
embedding_size,
training=True,
project_input=args.no_project,
optimizer=args.optim)
model.initialize(sess, embeddings)