def main(_):
dataset = MNIST(batch_size=BATCH_SIZE, is_train=False)
images, labels = dataset.dummy_inputs()
network = Network()
with tf.device("/gpu:0"):
logits = network.inference(images)
#loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
# labels=labels, logits=logits))
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=logits)
batch = tf.Variable(0, dtype=tf.int32, trainable=False)
variables = tf.trainable_variables()
var_grads = []
grads = []
for cur_var in variables:
print cur_var.shape
cur_grad = jacobian(losses, cur_var)
var_grads.append((cur_var, cur_grad))
grads.append(cur_grad)
for cur_var, grad in var_grads:
print "name: {} grad: {}".format(cur_var.op.name, grad)
if cur_var.op.name == "layer0/conv0/W":
first_conv_grad = grad
with tf.Session() as sess:
tf.global_variables_initializer().run()
tf.summary.FileWriter(logdir="./log", graph=sess.graph)
print('Initialized!')
# value check
"""
cur_images, cur_labels = dataset.next_batch()
value_grad = sess.run(first_conv_grad, feed_dict={images: cur_images, labels: cur_labels})
print value_grad[3]
print value_grad[0].shape
sys.exit(0)
"""
# warmup
for step in xrange(1):
cur_images, cur_labels = dataset.next_batch()
_ = sess.run(grads, feed_dict={images:cur_images, labels:cur_labels})
num_iteration = TOTAL_DATA / BATCH_SIZE
print ("loop {} data batch_size {} iteration {}".format(TOTAL_DATA, BATCH_SIZE, num_iteration))
start_time = time.time()
for step in xrange(num_iteration):
cur_images, cur_labels = dataset.next_batch()
_ = sess.run(grads, feed_dict={images: cur_images, labels: cur_labels})
duration = time.time() - start_time
num_data = num_iteration * BATCH_SIZE
print ("duraction {} time/data {}msec".format(duration, duration*1000/num_data))
def main(_):
dataset = MNIST(batch_size=BATCH_SIZE, is_train=False)
images, labels = dataset.dummy_inputs()
network = Network()
list_images = tf.split(images, BATCH_SIZE)
with tf.device("/gpu:0"):
list_logits = [
network.inference(cur_image) for cur_image in list_images
]
logits = tf.concat(list_logits, axis=0)
print logits
loss = tf.reduce_sum(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits))
batch = tf.Variable(0, dtype=tf.int32, trainable=False)
# variables = tf.trainable_variables()
# key:original value:copy
var_dict = network.get_all_variables()
all_vars = []
for cur_original, cur_copies in var_dict.items():
all_vars.extend(cur_copies)
grads = tf.gradients(loss, all_vars)
first_conv_grad = []
for cur_var, grad in zip(all_vars, grads):
if "layer0/conv0/W" in cur_var.op.name:
first_conv_grad.append(grad)
with tf.Session() as sess:
tf.global_variables_initializer().run()
tf.summary.FileWriter(logdir="./log", graph=sess.graph)
print('Initialized!')
# value check
#cur_images, cur_labels = dataset.next_batch()
#value_grad = sess.run(first_conv_grad, feed_dict={images: cur_images, labels: cur_labels})
#print value_grad[3]
#print value_grad[0].shape
#sys.exit(0)
# warmup
for step in xrange(100):
cur_images, cur_labels = dataset.next_batch()
_ = sess.run(grads,
feed_dict={
images: cur_images,
labels: cur_labels
})
num_iteration = TOTAL_DATA / BATCH_SIZE
print("loop {} data batch_size {} iteration {}".format(
TOTAL_DATA, BATCH_SIZE, num_iteration))
start_time = time.time()
for step in xrange(num_iteration):
cur_images, cur_labels = dataset.next_batch()
_ = sess.run(grads,
feed_dict={
images: cur_images,
labels: cur_labels
})
duration = time.time() - start_time
num_data = num_iteration * BATCH_SIZE
print("duraction {} time/data {}msec".format(
duration, duration * 1000 / num_data))