def train(train_set, test_set):
input_dim = 2
n_classes = 2
# inference model
with tf.name_scope('model'):
i = tf.placeholder("float", name='input')
o = tf.placeholder("int32", name='true_output')
l1 = linear(
input=i,
input_size=input_dim,
output_size=3,
name='linear_1'
)
a1 = tf.nn.tanh(l1, name='tanh_activation')
l2 = linear(
input=a1,
input_size=l1.input_size,
output_size=n_classes,
name='linear_2'
)
p_o_i = tf.nn.softmax(l2, name="softmax_output")
with tf.name_scope('loss'):
one_hot_labels = dense_to_one_hot(o, n_classes)
loss = tf.reduce_mean(-one_hot_labels * tf.log(p_o_i), name='loss')
tf.scalar_summary('loss', loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(one_hot_labels, 1), tf.argmax(p_o_i, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
tf.scalar_summary('accuracy', accuracy)
with tf.Session() as sess:
# Merge all the summaries and write them out to ./log
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter('./log', sess.graph_def)
saver = tf.train.Saver()
# training
train_op = tf.train.RMSPropOptimizer(FLAGS.learning_rate, FLAGS.decay, name='trainer').minimize(loss)
tf.initialize_all_variables().run()
for epoch in range(FLAGS.max_epochs):
sess.run(train_op, feed_dict={i: train_set['features'], o: train_set['targets']})
if epoch % max(int(FLAGS.max_epochs / 100), 1) == 0:
summary, lss, acc = sess.run([merged, loss, accuracy],
feed_dict={i: test_set['features'], o: test_set['targets']})
writer.add_summary(summary, epoch)
print()
print('Epoch: {epoch}'.format(epoch=epoch))
print(' - accuracy = {acc}'.format(acc=acc))
print(' - loss = {lss}'.format(lss=lss))
save_path = saver.save(sess, "model.ckpt")
print()
print("Model saved in file: %s" % save_path)
print()
print('Test features')
print(test_set['features'])
print('Test targets')
print(test_set['targets'])
print('Predictions')
p_o_i = sess.run(p_o_i, feed_dict={i: test_set['features'], o: test_set['targets']})
print(p_o_i)
print('Argmax predictions')
print(np.argmax(p_o_i, 1).reshape((-1, 1)))
def train(train_set, test_set, idx2word, word2idx):
embedding_size = 5
vocabulary_length = len(idx2word)
sequence_size = train_set['features'].shape[1]
lstm_size = 5
# inference model
with tf.name_scope('model'):
i = tf.placeholder("int32", name='input')
o = tf.placeholder("int32", name='true_output')
with tf.variable_scope("batch_size"):
batch_size = tf.shape(i)[0]
e = embedding(
input=i,
length=vocabulary_length,
size=embedding_size,
name='embedding'
)
with tf.name_scope("RNNCell"):
cell = LSTMCell(lstm_size, input_size=embedding_size)
state = cell.zero_state(batch_size, tf.float32)
outputs, states = rnn(
cell=cell,
inputs=[e[:, j, :] for j in range(sequence_size)],
initial_state=state,
name='RNN'
)
final_state = states[-1]
l = linear(
input=final_state,
input_size=cell.state_size,
output_size=vocabulary_length,
name='linear'
)
p_o_i = tf.nn.softmax(l, name="softmax_output")
with tf.name_scope('loss'):
one_hot_labels = dense_to_one_hot(o, vocabulary_length)
loss = tf.reduce_mean(-one_hot_labels * tf.log(p_o_i), name='loss')
tf.scalar_summary('loss', loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(one_hot_labels, 1), tf.argmax(p_o_i, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
tf.scalar_summary('accuracy', accuracy)
with tf.Session() as sess:
# Merge all the summaries and write them out to ./log
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter('./log', sess.graph_def)
saver = tf.train.Saver()
# training
train_op = tf.train.AdamOptimizer(FLAGS.learning_rate, name='trainer').minimize(loss)
tf.initialize_all_variables().run()
for epoch in range(FLAGS.max_epochs):
sess.run(train_op, feed_dict={i: train_set['features'], o: train_set['targets']})
if epoch % max(int(FLAGS.max_epochs / 100), 1) == 0:
summary, lss, acc = sess.run([merged, loss, accuracy],
feed_dict={i: test_set['features'], o: test_set['targets']})
writer.add_summary(summary, epoch)
print()
print('Epoch: {epoch}'.format(epoch=epoch))
print(' - accuracy = {acc}'.format(acc=acc))
print(' - loss = {lss}'.format(lss=lss))
save_path = saver.save(sess, "model.ckpt")
print()
print("Model saved in file: %s" % save_path)
print()
print('Test features')
print(test_set['features'])
print('Test targets')
print(test_set['targets'])
# print('Predictions')
p_o_i = sess.run(p_o_i, feed_dict={i: test_set['features'], o: test_set['targets']})
# print(p_o_i)
print('Argmax predictions')
print(np.argmax(p_o_i, 1).reshape((-1, 1)))