def run_vgg_training(model_name, data_directory, path_to_train_file,
path_to_val_file, path_to_labels_file, bsize, num_steps,
train_log_dir, optimizer, initial_checkpoint):
graph = tf.Graph()
sess = tf.InteractiveSession(graph=graph)
with graph.as_default():
name_dict, nclass = gen_dict(data_directory, path_to_labels_file)
label, image = getImage(path_to_train_file, nclass)
vlabel, vimage = getImage(path_to_val_file, nclass)
imageBatch, labelBatch = tf.train.shuffle_batch([image, label],
batch_size=bsize,
capacity=2000,
min_after_dequeue=1000)
vimageBatch, vlabelBatch = tf.train.shuffle_batch(
[vimage, vlabel],
batch_size=bsize,
capacity=2000,
min_after_dequeue=1000)
with sess.as_default():
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sess.run(tf.global_variables_initializer())
batch_xs, batch_ys = sess.run([imageBatch, labelBatch])
vbatch_xs, vbatch_ys = sess.run([vimageBatch, vlabelBatch])
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
if model_name == 'vgg_16':
with tf.variable_scope("vgg_16") as scope:
logits, end_points = vgg.vgg_16(batch_xs,
num_classes=2,
is_training=True)
scope.reuse_variables()
vlogits, vend_points = vgg.vgg_16(vbatch_xs,
num_classes=2,
is_training=True)
elif model_name == 'vgg_19':
with tf.variable_scope("vgg_19") as scope:
logits, end_points = vgg.vgg_19(batch_xs,
num_classes=2,
is_training=True)
scope.reuse_variables()
vlogits, vend_points = vgg.vgg_19(vbatch_xs,
num_classes=2,
is_training=True)
# predictions = end_points['Predictions']
# predictions_validation = vend_points['Predictions'] # -- for inception model use Predictions
predictions = tf.nn.softmax(logits)
predictions_validation = tf.nn.softmax(vlogits)
correct_prediction = tf.equal(tf.argmax(predictions, 1),
tf.argmax(batch_ys, 1))
vcorrect_prediction = tf.equal(
tf.argmax(predictions_validation, 1), tf.argmax(vbatch_ys, 1))
# get mean of all entries in correct prediction, the higher the better
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
accuracy_validation = tf.reduce_mean(
tf.cast(vcorrect_prediction, tf.float32))
logits = tf.reshape(logits, [bsize, 2])
tf.losses.softmax_cross_entropy(batch_ys, logits)
total_loss = tf.losses.get_total_loss()
train_tensor = slim.learning.create_train_op(total_loss, optimizer)
def train_step_fn(sess, *args, **kwargs):
total_loss, should_stop = train_step(sess, *args, **kwargs)
accuracy = sess.run([train_step_fn.accuracy])
if train_step_fn.step % 50 == 0:
# sess.run(assignment)
accuracy_validation = sess.run(
[train_step_fn.accuracy_validation])
# print('Step %s - Loss: %.2f Validation Accuracy: %.2f%%' %
# (str(train_step_fn.step).rjust(6, '0'), total_loss, accuracy * 100))
# saver.save(sess, os.path.join(train_log_dir, "model.ckpt"), train_step_fn.step)
train_step_fn.step += 1
return [total_loss, should_stop]
train_step_fn.step = 0
train_step_fn.accuracy = accuracy
train_step_fn.accuracy_validation = accuracy_validation
summaries.add(tf.summary.scalar('accuracy', accuracy))
# tf.summary.scalar('accuracy', accuracy)
summaries.add(tf.summary.scalar('loss', total_loss))
# tf.summary.scalar('loss', total_loss)
for end_point in end_points:
x = end_points[end_point]
summaries.add(tf.summary.histogram('activations/' + end_point, x))
for variable in slim.get_model_variables():
summaries.add(tf.summary.histogram(variable.op.name, variable))
summary_op = tf.summary.merge(list(summaries), name='summary_op')
slim.learning.train(train_tensor,
train_log_dir,
number_of_steps=num_steps,
summary_op=summary_op,
train_step_fn=train_step_fn,
save_summaries_secs=20)
print('completed training')
coord.request_stop()
coord.join(threads)
def run_inception_training(model_name, data_directory, path_to_train_file,
path_to_val_file, path_to_labels_file, bsize,
num_steps, train_log_dir, optimizer,
initial_checkpoint):
graph = tf.Graph()
sess = tf.InteractiveSession(graph=graph)
with graph.as_default():
name_dict, nclass = gen_dict(data_directory, path_to_labels_file)
label, image = getImage(path_to_train_file, nclass)
vlabel, vimage = getImage(path_to_val_file, nclass)
imageBatch, labelBatch = tf.train.shuffle_batch([image, label],
batch_size=bsize,
capacity=2000,
min_after_dequeue=1000)
vimageBatch, vlabelBatch = tf.train.shuffle_batch(
[vimage, vlabel],
batch_size=bsize,
capacity=2000,
min_after_dequeue=1000)
with sess.as_default():
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sess.run(tf.global_variables_initializer())
batch_xs, batch_ys = sess.run([imageBatch, labelBatch])
vbatch_xs, vbatch_ys = sess.run([vimageBatch, vlabelBatch])
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
if model_name == 'inception_v1':
with tf.variable_scope("InceptionV1") as scope:
logits, end_points = inception.inception_v1(
batch_xs, num_classes=2, is_training=True)
scope.reuse_variables()
vlogits, vend_points = inception.inception_v1(
vbatch_xs, num_classes=2, is_training=True)
elif model_name == 'inception_v2':
with tf.variable_scope("InceptionV2") as scope:
logits, end_points = inception.inception_v2(
batch_xs, num_classes=2, is_training=True)
scope.reuse_variables()
vlogits, vend_points = inception.inception_v2(
vbatch_xs, num_classes=2, is_training=True)
elif model_name == 'inception_v3':
with tf.variable_scope("InceptionV3") as scope:
logits, end_points = inception.inception_v3(
batch_xs, num_classes=2, is_training=True)
scope.reuse_variables()
vlogits, vend_points = inception.inception_v3(
vbatch_xs, num_classes=2, is_training=True)
predictions = end_points['Predictions']
predictions_validation = vend_points[
'Predictions'] # -- for inception model use Predictions
correct_prediction = tf.equal(tf.argmax(predictions, 1),
tf.argmax(batch_ys, 1))
vcorrect_prediction = tf.equal(
tf.argmax(predictions_validation, 1), tf.argmax(vbatch_ys, 1))
# get mean of all entries in correct prediction, the higher the better
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
accuracy_validation = tf.reduce_mean(
tf.cast(vcorrect_prediction, tf.float32))
logits = tf.reshape(logits, [bsize, 2])
tf.losses.softmax_cross_entropy(batch_ys, logits)
total_loss = tf.losses.get_total_loss()
train_tensor = slim.learning.create_train_op(total_loss, optimizer)
embedding_size = 1024
embedding_input = end_points
# print(tf.shape(embedding_input))
embedding = tf.Variable(tf.zeros([bsize, embedding_size]),
name="Embedding_Tensor")
assignment = embedding.assign()
writer = tf.summary.FileWriter(train_log_dir + '/', sess.graph)
config = tf.contrib.tensorboard.plugins.projector.ProjectorConfig()
embedding_config = config.embeddings.add()
embedding_config.tensor_name = embedding.name
embedding_config.sprite.image_path = train_log_dir + '/sprite.png'
embedding_config.metadata_path = train_log_dir + '/metadata.tsv'
print(embedding_config.sprite.image_path)
print(embedding_config.metadata_path)
# Specify the width and height of a single thumbnail.
embedding_config.sprite.single_image_dim.extend([224, 224])
tf.contrib.tensorboard.plugins.projector.visualize_embeddings(
writer, config)
def train_step_fn(sess, *args, **kwargs):
total_loss, should_stop = train_step(sess, *args, **kwargs)
accuracy = sess.run([train_step_fn.accuracy])
if train_step_fn.step % 50 == 0:
# sess.run(assignment)
accuracy_validation = sess.run(
[train_step_fn.accuracy_validation])
# print('Step %s - Loss: %.2f Validation Accuracy: %.2f%%' %
# (str(train_step_fn.step).rjust(6, '0'), total_loss, accuracy * 100))
# saver.save(sess, os.path.join(train_log_dir, "model.ckpt"), train_step_fn.step)
train_step_fn.step += 1
return [total_loss, should_stop]
train_step_fn.step = 0
train_step_fn.accuracy = accuracy
train_step_fn.accuracy_validation = accuracy_validation
tf.summary.scalar('accuracy', accuracy)
tf.summary.scalar('loss', total_loss)
for end_point in end_points:
x = end_points[end_point]
summaries.add(tf.summary.histogram('activations/' + end_point, x))
for variable in slim.get_model_variables():
summaries.add(tf.summary.histogram(variable.op.name, variable))
summary_op = tf.summary.merge(list(summaries), name='summary_op')
slim.learning.train(train_tensor,
train_log_dir,
number_of_steps=num_steps,
summary_op=summary_op,
train_step_fn=train_step_fn,
save_summaries_secs=20)
print('completed training')
coord.request_stop()
coord.join(threads)