tensorflow-cnn-visualization

mnist visualization example

python mnist_test.py --summary_path=yourpath
# after training
cd yourpath
tensorboard --logdir=./ --host=0.0.0.0

conv layers

pooling layers

filters

Usage

Visualize feature maps(activations) on TensorBoard

summary feature maps function

import visualizer
visualizer.summary_feature_maps(validation_sample_data, inputs, end_points, sess)

example code

import visualizer
...
inputs = tf.placeholder(tf.float16, [batch_size, image_size, image_size, image_channel])
conv1_weights = tf.Variable(tf.truncated_normal([5, 5, image_channel, 32], stddev=0.1, dtype=tf.float16))
conv1_biases = tf.Variable(tf.zeros([32], dtype=tf.float16))
conv = tf.nn.conv2d(inputs, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
end_points["conv1"] = conv
pool = tf.nn.max_pool(relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
end_points["pool1"] = pool
with tf.Session() as sess:
 tf.global_variables_initializer().run()
 ...
 visualizer.summary_feature_maps(validation_sample_data, inputs, end_points, sess)
 ...
 merged = tf.summary.merge_all()
 writer = tf.summary.FileWriter(FLAGS.summary_path)
 summary = sess.run(merged, feed_dict={inputs: validation_sample_data})
 writer.add_summary(summary, 0)
 writer.close()

Visualize filters(weights, kernels) on TensorBoard

summary filter function

summary filters function

import visualizer
visualizer.summary_filters([conv1_weights, conv2_weights])

example code

import visualizer
...
conv1_weights = tf.Variable(tf.truncated_normal([5, 5, image_channel, 32], stddev=0.1, dtype=tf.float16))
conv1_biases = tf.Variable(tf.zeros([32], dtype=tf.float16))
conv = tf.nn.conv2d(inputs, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
pool = tf.nn.max_pool(relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
visualizer.summary_filters([conv1_weights, conv2_weights])
with tf.Session() as sess:
 tf.global_variables_initializer().run()
 ...
 merged = tf.summary.merge_all()
 writer = tf.summary.FileWriter(FLAGS.summary_path)
 summary = sess.run(merged, feed_dict={inputs: validation_sample_data})
 writer.add_summary(summary, 0)
 writer.close()