After the Keras Example, let's build a tensorflow-based model as a comparision.
This example uses the same feature extraction techniques as the Keras Example.
In summary, the data prep follows these steps...
In [1]:
from pandas import read_csv
srooms_df = read_csv('../data/agaricus-lepiota.data.csv')
from sklearn_pandas import DataFrameMapper
import sklearn
import numpy as np
mappings = ([
('edibility', sklearn.preprocessing.LabelEncoder()),
('odor', sklearn.preprocessing.LabelBinarizer()),
('habitat', sklearn.preprocessing.LabelBinarizer()),
('spore-print-color', sklearn.preprocessing.LabelBinarizer())
])
mapper = DataFrameMapper(mappings)
srooms_np = mapper.fit_transform(srooms_df.copy()).astype(np.float32)
from sklearn.model_selection import train_test_split
train, test = train_test_split(srooms_np, test_size = 0.2, random_state=7)
train_labels = train[:,0:1]
train_data = train[:,1:]
test_labels = test[:,0:1]
test_data = test[:,1:]
Tensorflow requies a bit more work than Keras to define the network because we need to define the model's parameters (i.e. the weights and biases). Here is a Keras code snippnet for comparison:
from keras.models import Sequential
from keras.layers import Dense, Dropout
model = Sequential()
model.add(Dense(20, activation='relu', input_dim=25))
model.add(Dropout(0.5))
model.add(Dense(1, activation='sigmoid'))Here are the key differences:
In [2]:
import tensorflow as tf
import math
def inference(samples, input_dim, dense1_units, dense2_units):
with tf.name_scope('dense_1'):
weights = tf.Variable(
tf.truncated_normal([input_dim, dense1_units],
stddev=1.0 / math.sqrt(float(input_dim))),
name='weights')
biases = tf.Variable(tf.zeros([dense1_units]),
name='biases')
dense1 = tf.nn.relu(tf.nn.xw_plus_b(samples, weights, biases))
with tf.name_scope('dropout'):
dropout = tf.nn.dropout(dense1, 0.5)
with tf.name_scope('dense_2'):
weights = tf.Variable(
tf.truncated_normal([dense1_units, dense2_units],
stddev=1.0 / math.sqrt(float(dense2_units))),
name='weights')
biases = tf.Variable(tf.zeros([dense2_units]),
name='biases')
output = tf.sigmoid(tf.nn.xw_plus_b(dropout, weights, biases))
return output
Unlike Keras, TensorFlow doesn't provide pre-canned functions for training. The model needs the following functions defined.
log(0). Again, Keras hides these details by providing pre-canned loss and accuracy functions. The same definition in keras is a one liner.
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
In [3]:
def loss(output, labels, from_logits=False):
if not from_logits:
epsilon = 10e-8
output = tf.clip_by_value(output, epsilon, 1 - epsilon)
output = tf.log(output / (1 - output))
xentropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=output)
return tf.reduce_mean(xentropy)
def training(loss):
tf.summary.scalar('loss', loss)
optimizer = tf.train.AdamOptimizer()
global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss, global_step=global_step)
return train_op
def predict(output):
return tf.round(output)
def accuracy(output, labels):
return tf.reduce_mean(tf.to_float(tf.equal(predict(output),labels)))
This entire code block represents a single line in Keras...
model.fit(train_data, train_labels, epochs=10, batch_size=32, callbacks=[tensor_board])So, what's going on here?
In [4]:
import time
log_dir = './logs/tensor_srooms'
num_epochs=10
batch_size=64
with tf.Graph().as_default():
with tf.name_scope('input'):
features_initializer = tf.placeholder(dtype=tf.float32, shape=train_data.shape)
labels_initializer = tf.placeholder(dtype=tf.float32, shape=train_labels.shape)
input_features = tf.Variable(features_initializer, trainable=False, collections=[])
input_labels = tf.Variable(labels_initializer, trainable=False, collections=[])
# Shuffle the training data between epochs and train in batchs
feature, label = tf.train.slice_input_producer([input_features, input_labels], num_epochs=num_epochs)
features, labels = tf.train.batch([feature, label], batch_size=batch_size)
# Define layers dimensions
output = inference(features, 25, 20, 1)
loss_op = loss(output, labels)
train_op = training(loss_op)
# Define the metrics op
acc_op = accuracy(predict(output), labels)
# Initialize all variables op
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
summary_op = tf.summary.merge_all()
# Saver for the weights
saver = tf.train.Saver()
print('create saver')
# Start Session
sess = tf.Session()
sess.run(init_op)
print('session started')
# Load up the data.
sess.run(input_features.initializer, feed_dict={features_initializer: train_data})
sess.run(input_labels.initializer, feed_dict={labels_initializer: train_labels})
print('loaded data')
# Write the summary for tensorboard
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
# coordinate reading threads
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while not coord.should_stop():
start_time = time.time()
# Run one step of the model.
_, loss_value, acc_value = sess.run([train_op, loss_op, acc_op])
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f, acc = %.3f (%.3f sec)' % (step, loss_value, acc_value, duration))
# Update the events file.
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
step += 1
except tf.errors.OutOfRangeError:
print('Saving')
saver.save(sess, log_dir, global_step=step)
print('Done training for %d epochs, %d steps.' % (num_epochs, step))
finally:
# When done, ask the threads to stop.
coord.request_stop()
# Wait for threads to finish.
coord.join(threads)
sess.close()
In [ ]:
In [ ]: