def test_save_load(self):
# Data
x_batch = np.random.randint(0, VOCABULARY_SIZE,
[BATCH_SIZE, SEQUENCE_LENGTH])
y_batch = np.eye(2)[np.ones(BATCH_SIZE, dtype=np.int32)]
_, checkpoint_path = tempfile.mkstemp()
with self.graph.as_default(), self.sess.as_default():
rnn = self._build_classifier()
self.sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
feed_dict = {rnn.input_x: x_batch, rnn.input_y: y_batch}
loss1 = self.sess.run(rnn.total_loss, feed_dict)
saver.save(self.sess, checkpoint_path)
# Save model hyperparameters
f, path = tempfile.mkstemp()
rnn.save_to_file(path)
with tf.Graph().as_default(), tf.Session() as sess:
rnn2 = RNNClassifier.load_from_file(path)
x = tf.placeholder(tf.int32, [BATCH_SIZE, SEQUENCE_LENGTH])
y = tf.placeholder(tf.float32, [BATCH_SIZE, 2])
rnn2.build_graph(x, y)
saver = tf.train.Saver()
saver.restore(sess, checkpoint_path)
feed_dict = {rnn2.input_x: x_batch, rnn2.input_y: y_batch}
loss2 = sess.run(rnn2.total_loss, feed_dict)
self.assertEqual(loss1, loss2)
def test_save_load(self):
# Data
x_batch = np.random.randint(0, VOCABULARY_SIZE, [BATCH_SIZE, SEQUENCE_LENGTH])
y_batch = np.eye(2)[np.ones(BATCH_SIZE, dtype=np.int32)]
_, checkpoint_path = tempfile.mkstemp()
with self.graph.as_default(), self.sess.as_default():
rnn = self._build_classifier()
self.sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
feed_dict = {rnn.input_x: x_batch, rnn.input_y: y_batch}
loss1 = self.sess.run(rnn.total_loss, feed_dict)
saver.save(self.sess, checkpoint_path)
# Save model hyperparameters
f, path = tempfile.mkstemp()
rnn.save_to_file(path)
with tf.Graph().as_default(), tf.Session() as sess:
rnn2 = RNNClassifier.load_from_file(path)
x = tf.placeholder(tf.int32, [BATCH_SIZE, SEQUENCE_LENGTH])
y = tf.placeholder(tf.float32, [BATCH_SIZE, 2])
rnn2.build_graph(x, y)
saver = tf.train.Saver()
saver.restore(sess, checkpoint_path)
feed_dict = {rnn2.input_x: x_batch, rnn2.input_y: y_batch}
loss2 = sess.run(rnn2.total_loss, feed_dict)
self.assertEqual(loss1, loss2)
def _build_classifier(self):
with self.graph.as_default(), self.sess.as_default():
tf.set_random_seed(42)
x = tf.placeholder(tf.int32, [BATCH_SIZE, SEQUENCE_LENGTH])
y = tf.placeholder(tf.float32, [BATCH_SIZE, 2])
rnn = RNNClassifier(sequence_length=SEQUENCE_LENGTH,
vocabulary_size=VOCABULARY_SIZE,
num_classes=2,
batch_size=BATCH_SIZE,
backprop_truncate_after=SEQUENCE_LENGTH,
embedding_dim=128,
hidden_dim=128,
affine_dim=128,
cell_class="LSTM",
num_layers=2,
dropout_keep_prob_embedding=1.0,
dropout_keep_prob_affine=1.0,
dropout_keep_prob_cell_input=1.0,
dropout_keep_prob_cell_output=1.0)
rnn.build_graph(x, y)
return rnn
def _build_classifier(self):
with self.graph.as_default(), self.sess.as_default():
tf.set_random_seed(42)
x = tf.placeholder(tf.int32, [BATCH_SIZE, SEQUENCE_LENGTH])
y = tf.placeholder(tf.float32, [BATCH_SIZE, 2])
rnn = RNNClassifier(
sequence_length=SEQUENCE_LENGTH,
vocabulary_size=VOCABULARY_SIZE,
num_classes=2,
batch_size=BATCH_SIZE,
backprop_truncate_after=SEQUENCE_LENGTH,
embedding_dim=128,
hidden_dim=128,
affine_dim=128,
cell_class="LSTM",
num_layers=2,
dropout_keep_prob_embedding=1.0,
dropout_keep_prob_affine=1.0,
dropout_keep_prob_cell_input=1.0,
dropout_keep_prob_cell_output=1.0,
)
rnn.build_graph(x, y)
return rnn
from tfmodels.data.mr import MRData
from tfmodels.data.imdb import IMDBData
from tfmodels.data.ymrjp import YMRJPData
from datetime import datetime
from tfmodels.models.rnn.rnn_classifier import RNNClassifier, RNNClassifierTrainer, RNNClassifierEvaluator
pp = pprint.PrettyPrinter(indent=2)
# Pick training data
tf.flags.DEFINE_boolean("data_mr", False, "Dataset: MR")
tf.flags.DEFINE_boolean("data_sst", False, "Dataset: SST")
tf.flags.DEFINE_boolean("data_ymrjp", False,
"Dataset: Yahoo Movie Reviews (Japanese)")
# Classifier parameters
RNNClassifier.add_flags()
# Training parameters
tf.flags.DEFINE_integer("random_state", 42,
"Random state initialization for reproducibility")
tf.flags.DEFINE_integer("max_sequence_length", 512,
"Examples will be padded/truncated to this length")
tf.flags.DEFINE_integer("num_epochs", 20, "Number of training epochs")
tf.flags.DEFINE_integer(
"evaluate_every", 25,
"Evaluate model on dev set after this number of steps")
tf.flags.DEFINE_integer(
"checkpoint_every", 100,
"Evaluate model on dev set after this number of steps")
# Session Parameters
import tfmodels.data.utils
from tfmodels.data.mr import MRData
from tfmodels.data.imdb import IMDBData
from tfmodels.data.ymrjp import YMRJPData
from datetime import datetime
from tfmodels.models.rnn.rnn_classifier import RNNClassifier, RNNClassifierTrainer, RNNClassifierEvaluator
pp = pprint.PrettyPrinter(indent=2)
# Pick training data
tf.flags.DEFINE_boolean("data_mr", False, "Dataset: MR")
tf.flags.DEFINE_boolean("data_sst", False, "Dataset: SST")
tf.flags.DEFINE_boolean("data_ymrjp", False, "Dataset: Yahoo Movie Reviews (Japanese)")
# Classifier parameters
RNNClassifier.add_flags()
# Training parameters
tf.flags.DEFINE_integer("random_state", 42, "Random state initialization for reproducibility")
tf.flags.DEFINE_integer("max_sequence_length", 512, "Examples will be padded/truncated to this length")
tf.flags.DEFINE_integer("num_epochs", 20, "Number of training epochs")
tf.flags.DEFINE_integer("evaluate_every", 25, "Evaluate model on dev set after this number of steps")
tf.flags.DEFINE_integer("checkpoint_every", 100, "Evaluate model on dev set after this number of steps")
# Session Parameters
tf.flags.DEFINE_boolean("allow_soft_placement", False, "Allow soft device placement (e.g. no GPU)")
tf.flags.DEFINE_boolean("log_device_placement", False, "Log placement of ops on devices")
# Print parameters
FLAGS = tf.flags.FLAGS
FLAGS.batch_size
