def preload():
ckpt_path = 'ckpt-alexnet'
x_holder = tf.placeholder(tf.float32, [None, 150, 150, 3])
keep_prob = tf.placeholder(tf.float32)
pred, prob = alexnet.classifier(x_holder, keep_prob)
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
return keep_prob, prob, sess, x_holder
def classify(
image,
top_k,
k_patches,
ckpt_path,
imagenet_path):
""" Procedure to classify the image given through the command line
Args:
image: path to the image to classify
top_k: integer representing the number of predictions with highest probability
to retrieve
k_patches: number of crops taken from an image and to input to the model
ckpt_path: path to model's tensorflow checkpoint
imagenet_path: path to ILSRVC12 ImageNet folder containing train images,
validation images, annotations and metadata file
"""
wnids, words = tu.load_imagenet_meta(os.path.join(imagenet_path, 'data/meta.mat'))
# taking a few crops from an image
image_patches = tu.read_k_patches(image, k_patches)
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
_, pred = alexnet.classifier(x, dropout=1.0)
# calculate the average precision through the crops
avg_prediction = tf.div(tf.reduce_sum(pred, 0), k_patches)
# retrieve top 5 scores
scores, indexes = tf.nn.top_k(avg_prediction, k=top_k)
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto()) as sess:
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
s, i = sess.run([scores, indexes], feed_dict={x: image_patches})
s, i = np.squeeze(s), np.squeeze(i)
print('AlexNet saw:')
for idx in range(top_k):
print ('{} - score: {}'.format(words[i[idx]], s[idx]))
def test(top_k, k_patches, display_step, imagenet_path, ckpt_path):
"""
Procedure to evaluate top-1 and top-k accuracy (and error-rate) on the
ILSVRC2012 validation (test) set.
Args:
top_k: integer representing the number of predictions with highest probability
to retrieve
k_patches: number of crops taken from an image and to input to the model
display_step: number representing how often printing the current testing accuracy
imagenet_path: path to ILSRVC12 ImageNet folder containing train images,
validation images, annotations and metadata file
ckpt_path: path to model's tensorflow checkpoint
"""
test_images = sorted(
os.listdir(os.path.join(imagenet_path, 'ILSVRC2012_img_val')))
test_labels = tu.read_test_labels(
os.path.join(imagenet_path,
'data/ILSVRC2012_validation_ground_truth.txt'))
test_examples = len(test_images)
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
y = tf.placeholder(tf.float32, [None, 1000])
_, pred = alexnet.classifier(x, 1.0)
# calculate the average precision of the crops of the image
avg_prediction = tf.div(tf.reduce_sum(pred, 0), k_patches)
# accuracy
top1_correct = tf.equal(tf.argmax(avg_prediction, 0), tf.argmax(y, 1))
top1_accuracy = tf.reduce_mean(tf.cast(top1_correct, tf.float32))
topk_correct = tf.nn.in_top_k(tf.stack([avg_prediction]),
tf.argmax(y, 1),
k=top_k)
topk_accuracy = tf.reduce_mean(tf.cast(topk_correct, tf.float32))
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto()) as sess:
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
total_top1_accuracy = 0.
total_topk_accuracy = 0.
for i in range(test_examples):
# taking a few patches from an image
image_patches = tu.read_k_patches(
os.path.join(imagenet_path, 'ILSVRC2012_img_val',
test_images[i]), k_patches)
label = test_labels[i]
top1_a, topk_a = sess.run([top1_accuracy, topk_accuracy],
feed_dict={
x: image_patches,
y: [label]
})
total_top1_accuracy += top1_a
total_topk_accuracy += topk_a
if i % display_step == 0:
print(
'Examples done: {:5d}/{} ---- Top-1: {:.4f} -- Top-{}: {:.4f}'
.format(i + 1, test_examples,
total_top1_accuracy / (i + 1), top_k,
total_topk_accuracy / (i + 1)))
print('---- Final accuracy ----')
print('Top-1: {:.4f} -- Top-{}: {:.4f}'.format(
total_top1_accuracy / test_examples, top_k,
total_topk_accuracy / test_examples))
print('Top-1 error rate: {:.4f} -- Top-{} error rate: {:.4f}'.format(
1 - (total_top1_accuracy / test_examples), top_k,
1 - (total_topk_accuracy / test_examples)))
def train(epochs, batch_size, learning_rate, dropout, momentum, lmbda, resume,
imagenet_path, display_step, test_step, ckpt_path, summary_path):
""" Procedure to train the model on ImageNet ILSVRC 2012 training set
Args:
resume: boolean variable, true if want to resume the training, false to train from scratch
imagenet_path: path to ILSRVC12 ImageNet folder containing train images,
validation images, annotations and metadata file
display_step: number representing how often printing the current training accuracy
test_step: number representing how often make a test and print the validation accuracy
ckpt_path: path where to save model's tensorflow checkpoint (or from where resume)
summary_path: path where to save logs for TensorBoard
"""
train_img_path = os.path.join(imagenet_path, 'ILSVRC2012_img_train')
ts_size = tu.imagenet_size(train_img_path)
num_batches = int(float(ts_size) / batch_size)
wnid_labels, _ = tu.load_imagenet_meta(
os.path.join(imagenet_path, 'data/meta.mat'))
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
y = tf.placeholder(tf.float32, [None, 1000])
lr = tf.placeholder(tf.float32)
keep_prob = tf.placeholder(tf.float32)
# queue of examples being filled on the cpu
with tf.device('/cpu:0'):
q = tf.FIFOQueue(batch_size * 3, [tf.float32, tf.float32],
shapes=[[224, 224, 3], [1000]])
enqueue_op = q.enqueue_many([x, y])
x_b, y_b = q.dequeue_many(batch_size)
pred, _ = alexnet.classifier(x_b, keep_prob)
# cross-entropy and weight decay
with tf.name_scope('cross_entropy'):
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=pred,
labels=y_b,
name='cross-entropy'))
with tf.name_scope('l2_loss'):
l2_loss = tf.reduce_sum(
lmbda *
tf.stack([tf.nn.l2_loss(v) for v in tf.get_collection('weights')]))
tf.summary.scalar('l2_loss', l2_loss)
with tf.name_scope('loss'):
loss = cross_entropy + l2_loss
tf.summary.scalar('loss', loss)
# accuracy
with tf.name_scope('accuracy'):
correct = tf.equal(tf.argmax(pred, 1), tf.argmax(y_b, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
tf.summary.scalar('accuracy', accuracy)
global_step = tf.Variable(0, trainable=False)
epoch = tf.div(global_step, num_batches)
# momentum optimizer
with tf.name_scope('optimizer'):
optimizer = tf.train.MomentumOptimizer(learning_rate=lr,
momentum=momentum).minimize(
loss,
global_step=global_step)
# merge summaries to write them to file
merged = tf.summary.merge_all()
# checkpoint saver
saver = tf.train.Saver()
coord = tf.train.Coordinator()
#init = tf.initialize_all_variables()
init = tf.global_variables_initializer()
with tf.Session(config=tf.ConfigProto()) as sess:
if resume:
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
else:
sess.run(init)
# enqueuing batches procedure
def enqueue_batches():
while not coord.should_stop():
im, l = tu.read_batch(batch_size, train_img_path, wnid_labels)
sess.run(enqueue_op, feed_dict={x: im, y: l})
# creating and starting parallel threads to fill the queue
num_threads = 3
for i in range(num_threads):
t = threading.Thread(target=enqueue_batches)
t.setDaemon(True)
t.start()
# operation to write logs for tensorboard visualization
train_writer = tf.summary.FileWriter(
os.path.join(summary_path, 'train'), sess.graph)
start_time = time.time()
for e in range(sess.run(epoch), epochs):
for i in range(num_batches):
summary_str, _, step = sess.run(
[merged, optimizer, global_step],
feed_dict={
lr: learning_rate,
keep_prob: dropout
})
train_writer.add_summary(summary_str, step)
# decaying learning rate
if step == 170000 or step == 350000:
learning_rate /= 10
# display current training informations
if step % display_step == 0:
c, a = sess.run([loss, accuracy],
feed_dict={
lr: learning_rate,
keep_prob: 1.0
})
print(
'Epoch: {:03d} Step/Batch: {:09d} --- Loss: {:.7f} Training accuracy: {:.4f}'
.format(e, step, c, a))
# make test and evaluate validation accuracy
if step % test_step == 0:
val_im, val_cls = tu.read_validation_batch(
batch_size,
os.path.join(imagenet_path, 'ILSVRC2012_img_val'),
os.path.join(
imagenet_path,
'data/ILSVRC2012_validation_ground_truth.txt'))
v_a = sess.run(accuracy,
feed_dict={
x_b: val_im,
y_b: val_cls,
lr: learning_rate,
keep_prob: 1.0
})
# intermediate time
int_time = time.time()
print('Elapsed time: {}'.format(
tu.format_time(int_time - start_time)))
print('Validation accuracy: {:.04f}'.format(v_a))
# save weights to file
save_path = saver.save(
sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
print('Variables saved in file: %s' % save_path)
end_time = time.time()
print('Elapsed time: {}'.format(tu.format_time(end_time - start_time)))
save_path = saver.save(sess, os.path.join(ckpt_path,
'alexnet-cnn.ckpt'))
print('Variables saved in file: %s' % save_path)
coord.request_stop()
def train(
epochs,
batch_size,
learning_rate,
dropout,
momentum,
lmbda,
resume,
display_step,
test_step,
ckpt_path,
summary_path
):
train_img_path = '/var/data/bias_data/image/train'
evaluate_path = '/var/data/bias_data/image/train'
num_whole_images = 60000
num_batches = int(float(num_whole_images) / batch_size)
wnid_labels = ['cheer_out', 'fearful_out', 'happy_out', 'joy_out', 'rage_out', 'sorrow_out']
x = tf.placeholder(tf.float32, [None, 150, 150, 3])
y = tf.placeholder(tf.float32, [None, 6])
lr = tf.placeholder(tf.float32)
keep_prob = tf.placeholder(tf.float32)
# queue of examples being filled on the cpu
with tf.device('/cpu:0'):
q = tf.FIFOQueue(batch_size * 3, [tf.float32, tf.float32], shapes=[[150, 150, 3], [6]])
enqueue_op = q.enqueue_many([x, y])
x_b, y_b = q.dequeue_many(batch_size)
pred, prob = alexnet.classifier(x_b, keep_prob)
# cross-entropy and weight decay
with tf.name_scope('cross_entropy'):
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y_b, name='cross-entropy'))
with tf.name_scope('l2_loss'):
l2_loss = tf.reduce_sum(lmbda * tf.stack([tf.nn.l2_loss(v) for v in tf.get_collection('weights')]))
tf.summary.scalar('l2_loss', l2_loss)
with tf.name_scope('loss'):
loss = cross_entropy + l2_loss
tf.summary.scalar('loss', loss)
# accuracy
with tf.name_scope('accuracy'):
correct = tf.equal(tf.argmax(pred, 1), tf.argmax(y_b, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
tf.summary.scalar('accuracy', accuracy)
global_step = tf.Variable(0, trainable=False)
epoch = tf.div(global_step, num_batches)
# momentum optimizer
with tf.name_scope('optimizer'):
optimizer = tf.train.MomentumOptimizer(learning_rate=lr, momentum=momentum).minimize(loss,
global_step=global_step)
merged = tf.summary.merge_all()
saver = tf.train.Saver()
coord = tf.train.Coordinator()
init = tf.global_variables_initializer()
with tf.Session(config=tf.ConfigProto()) as sess:
if resume:
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
else:
sess.run(init)
# enqueuing batches procedure
def enqueue_batches():
while not coord.should_stop():
im, l = tu.read_batch(batch_size, train_img_path, wnid_labels)
sess.run(enqueue_op, feed_dict={x: im, y: l})
# creating and starting parallel threads to fill the queue
num_threads = 3
threads = []
for i in range(num_threads):
t = threading.Thread(target=enqueue_batches)
t.setDaemon(True)
t.start()
threads.append(t)
# operation to write logs for tensorboard visualization
train_writer = tf.summary.FileWriter(os.path.join(summary_path, 'train'), sess.graph)
valid_batch_size = 126
val_im, val_cls = tu.read_batch(valid_batch_size, evaluate_path, wnid_labels)
start_time = time.time()
for e in range(sess.run(epoch), epochs):
for i in range(num_batches):
_, step = sess.run([optimizer, global_step], feed_dict={lr: learning_rate, keep_prob: dropout})
# train_writer.add_summary(summary, step)
# decaying learning rate
if step == 170000 or step == 350000:
learning_rate /= 10
# display current training informations
if step % display_step == 0:
c, a = sess.run([loss, accuracy], feed_dict={lr: learning_rate, keep_prob: 1.0})
print (
'Epoch: {:03d} Step/Batch: {:09d} --- Loss: {:.7f} Training accuracy: {:.4f}'.format(e, step, c,
a))
# make test and evaluate validation accuracy
if step % test_step == 0:
v_a = sess.run(accuracy, feed_dict={x_b: val_im, y_b: val_cls, lr: learning_rate, keep_prob: 1.0})
# intermediate time
int_time = time.time()
print ('Elapsed time: {}'.format(tu.format_time(int_time - start_time)))
print ('Validation accuracy: {:.04f}'.format(v_a))
# save weights to file
save_path = saver.save(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
print('Variables saved in file: %s' % save_path)
end_time = time.time()
print ('Elapsed time: {}'.format(tu.format_time(end_time - start_time)))
save_path = saver.save(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
print('Variables saved in file: %s' % save_path)
coord.request_stop()
coord.join(threads)
os.makedirs(layer_spec_dir)
print('Create dir: %s' % layer_spec_dir)
""" get imagenet images """
test_images = sorted(
os.listdir(os.path.join(imagenet_path, 'ILSVRC2012_img_val')))
test_labels = tu.read_test_labels(
os.path.join(imagenet_path,
'data/ILSVRC2012_validation_ground_truth.txt'))
n_imgs = len(test_images)
print(n_imgs)
""" build graph """
# set up the computational graph
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
y = tf.placeholder(tf.float32, [None, 1000])
_, pred = alexnet.classifier(x, 1.0)
# load the model back
saver = tf.train.Saver()
""""""
sess = tf.InteractiveSession()
saver.restore(sess, os.path.join(ckpt_path, 'alexnet-cnn.ckpt'))
# get layer name
conv1 = tf.get_default_graph().get_tensor_by_name(
'alexnet_cnn/alexnet_cnn_conv1/Relu:0')
conv2 = tf.get_default_graph().get_tensor_by_name(
'alexnet_cnn/alexnet_cnn_conv2/Relu:0')
conv3 = tf.get_default_graph().get_tensor_by_name(
'alexnet_cnn/alexnet_cnn_conv3/Relu:0')
conv4 = tf.get_default_graph().get_tensor_by_name(
