I'm trying to implement the first part of the google blog entry Inceptionism: Going Deeper into Neural Networks in TensorFlow. So far I have found several resources that either explain it in natural language or focus on other parts or give code snippets for other frameworks. I understand the idea of optimizing a random input image with respect to a class prior and also the maths behind it given in the this paper, section 2, but I'm not able to implement it myself using TensorFlow.
From this SO question and the helpful comment by etarion, I now know that you can give a list of variables to the optimizer, while all other variables are untouched. However, when giving the optimizer a random image as a variable leads to
File "mnist_test.py", line 101, in main
optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(-cost, var_list=[rnd_img])
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 198, in minimize
name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 309, in apply_gradients
(converted_grads_and_vars,))
ValueError: No gradients provided for any variable: ((None,<tensorflow.python.ops.variables.Variable object at 0x7feac1870410>),)
For testing purpose I used a stripped down MNIST example. I tried to keep it as short as possible while still being readable and executable:
def main():
# parameters
learning_rate = 0.001
train_batches = 1000
batch_size = 128
display_step = 50
# net parameters
n_input = 784 #28x28
n_classes = 10
keep_prob = 0.75
weights = {
'wc1': tf.Variable(tf.truncated_normal([5, 5, 1, 32])),
'wc2': tf.Variable(tf.truncated_normal([5, 5, 32, 64])),
'wd1': tf.Variable(tf.truncated_normal([7*7*64, 1024])),
'out': tf.Variable(tf.truncated_normal([1024, n_classes]))
}
biases = {
'bc1': tf.Variable(tf.constant(0.1, shape=[32])),
'bc2': tf.Variable(tf.constant(0.1, shape=[64])),
'bd1': tf.Variable(tf.constant(0.1, shape=[1024])),
'out': tf.Variable(tf.constant(0.1, shape=[n_classes]))
}
# tf inputs
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
dropout = tf.placeholder(tf.float32)
# create net
net = create_net(x, weights, biases, keep_prob)
# define loss
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, y))
# define optimizer
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
# evaluation
pred_correct = tf.equal(tf.argmax(net, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(pred_correct, tf.float32))
print "loading mnist data"
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
for i in xrange(train_batches):
batch_x, batch_y = mnist.train.next_batch(batch_size)
sess.run(optimizer, feed_dict={x: batch_x, y: batch_y, dropout: keep_prob})
if i % display_step == 0:
loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x, y: batch_y, dropout: 1.0})
print "batch: %i, loss: %.5f, accuracy: %.5f" % (i, loss, acc)
acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels, dropout: 1.0})
print "test accuracy: %.5f" % (acc)
# ====== this is where the reconstruction begins =====
rnd_img = tf.Variable(tf.random_normal([1, n_input]))
one_hot = np.zeros(10)
one_hot[4] = 1;
# the next line causes the error
optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(-cost, var_list=[rnd_img])
for i in xrange(1000):
session.run(optimizer2, feed_dict={x: rnd_img, y: one_hot, dropout: 1.0})
sess.close()
if __name__ == "__main__":
main()
The helper functions I used:
def create_net(x, weights, biases, dropout):
x = tf.reshape(x, shape=[-1, 28, 28, 1])
conv1 = conv2d_relu(x, weights['wc1'], biases['bc1'])
conv1 = maxpool2d(conv1, 2)
conv2 = conv2d_relu(conv1, weights['wc2'], biases['bc2'])
conv2 = maxpool2d(conv2, 2)
fc1 = fullyconnected_relu(conv2, weights['wd1'], biases['bd1'])
fc1 = tf.nn.dropout(fc1, dropout)
out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
return out
def conv2d_relu(x, W, b, stride=1):
conv = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding='SAME')
conv = tf.nn.bias_add(conv, b)
return tf.nn.relu(conv)
def maxpool2d(x, k=2, stride=2, padding='VALID'):
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)
def fullyconnected_relu(x, W, b):
fc = tf.reshape(x, [-1, W.get_shape().as_list()[0]])
fc = tf.add(tf.matmul(fc, W), b)
fc = tf.nn.relu(fc)
I've found some sources saying that this error occurs when there is no path within the computation graph between the output and the variables to be optimize, but I don't see why this should be the case here.
My questions are:
- Why isn't the optimizer able to apply any gradients?
- Is this the right way to go in order to implement the visualization of a class?
Thanks in advance.
Edit:
Here is the complete code again, after incorporation of the accepted answer (for anyone who is interested). Anyway, the results are still not as expected, as the script basically produces random images after 100000 rounds of reconstruction. Ideas are welcome.
import tensorflow as tf
import numpy as np
import skimage.io
def conv2d_relu(x, W, b, stride=1):
conv = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding='SAME')
conv = tf.nn.bias_add(conv, b)
return tf.nn.relu(conv)
def maxpool2d(x, k=2, stride=2, padding='VALID'):
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)
def fullyconnected_relu(x, W, b):
fc = tf.reshape(x, [-1, W.get_shape().as_list()[0]])
fc = tf.add(tf.matmul(fc, W), b)
fc = tf.nn.relu(fc)
return fc;
def create_net(x, weights, biases, dropout):
x = tf.reshape(x, shape=[-1, 28, 28, 1])
conv1 = conv2d_relu(x, weights['wc1'], biases['bc1'])
conv1 = maxpool2d(conv1, 2)
conv2 = conv2d_relu(conv1, weights['wc2'], biases['bc2'])
conv2 = maxpool2d(conv2, 2)
fc1 = fullyconnected_relu(conv2, weights['wd1'], biases['bd1'])
fc1 = tf.nn.dropout(fc1, dropout)
out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
return out
def save_image(img_data, name):
img = img_data.reshape(28,28)
mi = np.min(img)
ma = np.max(img)
img = (img-mi)/(ma-mi)
skimage.io.imsave(name, img)
def main():
# parameters
learning_rate = 0.001
train_batches = 1000
batch_size = 100
display_step = 50
# net parameters
n_input = 784 #28x28
n_classes = 10
keep_prob = 0.75
weights = {
'wc1': tf.Variable(tf.truncated_normal([5, 5, 1, 32])),
'wc2': tf.Variable(tf.truncated_normal([5, 5, 32, 64])),
'wd1': tf.Variable(tf.truncated_normal([7*7*64, 1024])),
'out': tf.Variable(tf.truncated_normal([1024, n_classes]))
}
biases = {
'bc1': tf.Variable(tf.constant(0.1, shape=[32])),
'bc2': tf.Variable(tf.constant(0.1, shape=[64])),
'bd1': tf.Variable(tf.constant(0.1, shape=[1024])),
'out': tf.Variable(tf.constant(0.1, shape=[n_classes]))
}
# tf inputs
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
dropout = tf.placeholder(tf.float32)
# create net
net = create_net(x, weights, biases, dropout)
# define loss
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, y))
# define optimizer
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
# evaluation
pred_correct = tf.equal(tf.argmax(net, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(pred_correct, tf.float32))
print "loading mnist data"
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
for i in xrange(train_batches):
batch_x, batch_y = mnist.train.next_batch(batch_size)
sess.run(optimizer, feed_dict={x: batch_x, y: batch_y, dropout: keep_prob})
if i % display_step == 0:
loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x, y: batch_y, dropout: 1.0})
print "batch: %i, loss: %.5f, accuracy: %.5f" % (i, loss, acc)
acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels, dropout: 1.0})
print "test accuracy: %.5f" % (acc)
# reconstruction part
rnd_img = tf.Variable(tf.random_normal([1, n_input]))
one_hot = np.zeros((1, 10))
one_hot[0,1] = 1;
net2 = create_net(rnd_img, weights, biases, dropout)
cost2 = -tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net2, y))
optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(cost2, var_list=[rnd_img])
init_var_list = []
for var in tf.all_variables():
if(not tf.is_variable_initialized(var).eval(session=sess)):
init_var_list.append(var)
sess.run(tf.initialize_variables(init_var_list))
save_image(rnd_img.eval(sess), "bevor.tiff")
for i in xrange(100000):
_, loss = sess.run([optimizer2, cost2], feed_dict={y: one_hot, dropout: 1.0})
if(i%10000 == 0):
cur_img = rnd_img.eval(session=sess)
print "loss:", loss, "mi:", np.min(cur_img), "ma:", np.max(cur_img)
save_image(rnd_img.eval(sess), "after.tiff")
sess.close()
if __name__ == "__main__":
main()
Some explanation: After rebuilding the graph with the new input variable and optimizer, I had to initialize the new variables, i.e. the rnd_img and some helper variables used by the Adam optimizer, hence the loop over all_variables() and checking for initialization status. If somebody knows a more elegant way, let me know. Or maybe that's the reason why I don't get any results?
