def inference_resnet_own(images, training):
# reference: https://github.com/tensorflow/models/blob/master/official/resnet/imagenet_main.py
resnet_size = FLAGS.resnet_type
choices = {
18: [2, 2, 2, 2],
34: [3, 4, 6, 3],
50: [3, 4, 6, 3],
101: [3, 4, 23, 3],
152: [3, 8, 36, 3],
200: [3, 24, 36, 3]
}
if resnet_size < 50:
bottleneck = False
final_size = 512
else:
bottleneck = True
final_size = 2048
resnet = resnet_model.Model(resnet_size=resnet_size,
bottleneck=bottleneck,
num_classes=NUM_CLASSES + 1,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=choices[resnet_size],
block_strides=[1, 2, 2, 2],
final_size=final_size,
resnet_version=2)
softmax_linear = resnet(images, training)
return softmax_linear
def main(_):
# Specify which gpu to be used
# os.environ["CUDA_VISIBLE_DEVICES"] = '1'
cls_model = resnet_model.Model(resnet_size=50,
bottleneck=True,
num_classes=26,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=[3, 4, 6, 3],
block_strides=[1, 2, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_first',
dtype=resnet_model.DEFAULT_DTYPE)
if FLAGS.input_shape:
input_shape = [
int(dim) if dim != -1 else None
for dim in FLAGS.input_shape.split(',')
]
else:
input_shape = [None, None, None, 3]
exporter.export_inference_graph(FLAGS.input_type, cls_model,
FLAGS.trained_checkpoint_prefix,
FLAGS.output_directory, input_shape)
def inference(images, training):
# reference: https://github.com/tensorflow/models/blob/master/official/resnet/cifar10_main.py
resnet = resnet_model.Model(resnet_size=32,
bottleneck=False,
num_classes=NUM_CLASSES,
num_filters=16,
kernel_size=3,
conv_stride=1,
first_pool_size=None,
first_pool_stride=None,
block_sizes=[5, 5, 5],
block_strides=[1, 2, 2],
final_size=64,
resnet_version=2)
softmax_linear = resnet(images, training)
return softmax_linear
def dnn(image, training):
"""
function which calls resnet model with proper initialization for cifar10 model
:param image: input image tensor
:return: model output tensor node
"""
resnet_object = resnet_model.Model(resnet_size=32,
bottleneck=False,
num_classes=10,
num_filters=16,
kernel_size=3,
conv_stride=1,
first_pool_size=None,
first_pool_stride=None,
block_sizes=[5, 5, 5],
block_strides=[1, 2, 2],
final_size=64,
data_format='channels_last')
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_op):
logits = resnet_object(image, training=training)
return logits
def FCN2(image, mean, variance, phase):
"""
function defining fcn model
:param image: input image tensor in the flattened form
:return: model output tensor node
"""
image = tf.cast(image, tf.float32)
image_reshape = tf.reshape(image, [-1, 512, 512, 3])
image_norm = tf.nn.batch_normalization(image_reshape, mean, variance, None,
None, 0.0001)
model = resnet_model.Model(
resnet_size=50,
bottleneck=False,
num_classes=2,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=[3, 4, 6, 3],
block_strides=[1, 2, 2, 2],
resnet_version=1,
data_format='channels_last',
)
resnet_out = model(image_norm, phase)
intermediate_out = [
v.values()[0] for v in tf.get_default_graph().get_operations()
if 'block_layer' in v.name
]
conv1 = tf.layers.conv2d(intermediate_out[3],
filters=512,
kernel_size=[1, 1],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='conv1_1x1')
conv2 = tf.layers.conv2d(conv1,
filters=512,
kernel_size=[1, 1],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='conv2_1x1')
up1 = tf.layers.conv2d_transpose(conv2,
filters=64,
kernel_size=[16, 16],
strides=(8, 8),
padding='same',
activation=tf.nn.relu,
name='upsample_1')
up2 = tf.layers.conv2d_transpose(intermediate_out[2],
filters=64,
kernel_size=[8, 8],
strides=(4, 4),
padding='same',
activation=tf.nn.relu,
name='upsample_2')
up3 = tf.layers.conv2d_transpose(intermediate_out[1],
filters=64,
kernel_size=[4, 4],
strides=(2, 2),
padding='same',
activation=tf.nn.relu,
name='upsample_3')
up4 = intermediate_out[0]
up5 = tf.layers.conv2d_transpose(up4 + up3 + up2 + up1,
filters=2,
kernel_size=[8, 8],
strides=(4, 4),
padding='same',
activation=None,
name='upsample_5')
return up5
default='/cluster/project/infk/hilliges/lectures/mp20/project2/',
help='path to the dataset')
parser.add_argument('--batch_size', type=int, default=1, help='batch size')
parser.add_argument('--log_dir',
type=str,
default='./example',
help='log storage dir for tensorboard')
opt = parser.parse_args()
with tf.Session() as sess:
# define resnet model
sample = create_test_dataloader(data_root=opt.data_root,
batch_size=opt.batch_size)
with tf.variable_scope('model'):
model = resnet_model.Model()
p3d_out_norm = model(sample['image'], training=False)
p3d_out = unnormalize_pose(p3d_out_norm)
p3d_out = tf.reshape(p3d_out, [-1, 51])
# restore weights
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(opt.log_dir))
predictions = None
with trange(math.ceil(meta_info.NUM_SAMPLES_TEST / opt.batch_size)) as t:
for i in t:
p3d_out_ = sess.run(p3d_out)
if predictions is None:
predictions = p3d_out_
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--input_dir", "-i")
args = parser.parse_args()
data_dir = args.input_dir
sess = tf.Session()
model = resnet_model.Model(resnet_size=RESNET_SIZE,
bottleneck=False,
num_classes=NUM_CLASSES,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(RESNET_SIZE),
block_strides=[1, 2, 2, 2],
resnet_version=RESNET_VERSION,
data_format=None,
dtype=DTYPE,
sess=sess)
print("starting")
dataset = data.custom_input_fn(data_dir, batch_size=BATCH_SIZE)
iterator = dataset.make_initializable_iterator()
it_init = iterator.initializer
sess.run(it_init)
outs, loss = model.network(iterator, training=TRAINING)
def resnet_model_fn(features, labels, mode, params):
# Generate a summary node for the images
tf.compat.v1.summary.image('images', features, max_outputs=6)
# Checks that features/images have same data type being used for calculations.
assert features.dtype == resnet_model.DEFAULT_DTYPE
resnet_size = params['resnet_size']
if resnet_size < 50:
bottleneck = False
else:
bottleneck = True
model = resnet_model.Model(resnet_size=resnet_size,
bottleneck=bottleneck,
num_classes=FLAGS.num_classes,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(resnet_size),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_first',
dtype=resnet_model.DEFAULT_DTYPE)
logits = model(features, mode == tf.estimator.ModeKeys.TRAIN)
# This acts as a no-op if the logits are already in fp32 (provided logits are
# not a SparseTensor). If dtype is is low precision, logits must be cast to
# fp32 for numerical stability.
logits = tf.cast(logits, tf.float32)
predictions = {
'classes': tf.argmax(input=logits, axis=1),
'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
}
if mode == tf.estimator.ModeKeys.PREDICT:
# Return the predictions and the specification for serving a SavedModel
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
export_outputs={
'predict': tf.estimator.export.PredictOutput(predictions)
})
# Calculate loss, which includes softmax cross entropy and L2 regularization.
if FLAGS.label_smoothing != 0.0:
one_hot_labels = tf.one_hot(labels, 1001)
cross_entropy = tf.losses.softmax_cross_entropy(
logits=logits,
onehot_labels=one_hot_labels,
label_smoothing=FLAGS.label_smoothing)
else:
cross_entropy = tf.compat.v1.losses.sparse_softmax_cross_entropy(
logits=logits, labels=labels)
# Create a tensor named cross_entropy for logging purposes.
tf.identity(cross_entropy, name='cross_entropy')
tf.compat.v1.summary.scalar('cross_entropy', cross_entropy)
# If no loss_filter_fn is passed, assume we want the default behavior,
# which is that batch_normalization variables are excluded from loss.
def loss_filter_fn(_):
return True
# Add weight decay to the loss.
weight_decay = FLAGS.weight_decay
l2_loss = weight_decay * tf.add_n(
# loss is computed using fp32 for numerical stability.
[
tf.nn.l2_loss(tf.cast(v, tf.float32))
for v in tf.compat.v1.trainable_variables()
if loss_filter_fn(v.name)
])
tf.compat.v1.summary.scalar('l2_loss', l2_loss)
loss = cross_entropy + l2_loss
scaffold = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_or_create_global_step()
learning_rate = configure_learning_rate(FLAGS.decay_steps, global_step)
# Create a tensor named learning_rate for logging purposes
tf.identity(learning_rate, name='learning_rate')
tf.compat.v1.summary.scalar('learning_rate', learning_rate)
momentum = FLAGS.momentum
if FLAGS.enable_lars:
optimizer = tf.contrib.opt.LARSOptimizer(
learning_rate,
momentum=momentum,
weight_decay=weight_decay,
skip_list=['batch_normalization', 'bias'])
else:
optimizer = tf.compat.v1.train.MomentumOptimizer(
learning_rate=learning_rate, momentum=momentum)
# loss_scale = FLAGS.loss_scale
# fp16_implementation = FLAGS.fp16_implementation
# if fp16_implementation == 'graph_rewrite':
# optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
# optimizer, loss_scale=loss_scale)
def _dense_grad_filter(gvs):
"""Only apply gradient updates to the final layer.
This function is used for fine tuning.
Args:
gvs: list of tuples with gradients and variable info
Returns:
filtered gradients so that only the dense layer remains
"""
return [(g, v) for g, v in gvs if 'dense' in v.name]
# if loss_scale != 1 and fp16_implementation != 'graph_rewrite':
# # When computing fp16 gradients, often intermediate tensor values are
# # so small, they underflow to 0. To avoid this, we multiply the loss by
# # loss_scale to make these tensor values loss_scale times bigger.
# scaled_grad_vars = optimizer.compute_gradients(loss * loss_scale)
#
# if fine_tune:
# scaled_grad_vars = _dense_grad_filter(scaled_grad_vars)
#
# # Once the gradient computation is complete we can scale the gradients
# # back to the correct scale before passing them to the optimizer.
# unscaled_grad_vars = [(grad / loss_scale, var)
# for grad, var in scaled_grad_vars]
# minimize_op = optimizer.apply_gradients(unscaled_grad_vars, global_step)
# else:
grad_vars = optimizer.compute_gradients(loss)
# if fine_tune:
# grad_vars = _dense_grad_filter(grad_vars)
minimize_op = optimizer.apply_gradients(grad_vars, global_step)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
train_op = tf.group(minimize_op, update_ops)
# keep_checkpoint_every_n_hours = FLAGS.keep_checkpoint_every_n_hours
# saver = tf.train.Saver(
# sharded=True,
# keep_checkpoint_every_n_hours=keep_checkpoint_every_n_hours,
# save_relative_paths=True)
# # if not tf.get_collection(tf.GraphKeys.SAVERS):
# tf.add_to_collection(tf.GraphKeys.SAVERS, saver)
# scaffold = tf.train.Scaffold(saver=saver)
else:
train_op = None
accuracy = tf.compat.v1.metrics.accuracy(labels, predictions['classes'])
accuracy_top_5 = tf.compat.v1.metrics.mean(
tf.nn.in_top_k(predictions=logits,
targets=labels,
k=5,
name='top_5_op'))
metrics = {'accuracy': accuracy, 'accuracy_top_5': accuracy_top_5}
# Create a tensor named train_accuracy for logging purposes
tf.identity(accuracy[1], name='train_accuracy')
tf.identity(accuracy_top_5[1], name='train_accuracy_top_5')
tf.compat.v1.summary.scalar('train_accuracy', accuracy[1])
tf.compat.v1.summary.scalar('train_accuracy_top_5', accuracy_top_5[1])
return tf.estimator.EstimatorSpec(mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op,
eval_metric_ops=metrics,
scaffold=scaffold)
with tf.Session(config=config) as sess:
# Read data
if opt.val_subject in TRAINING_SUBJECT:
TRAINING_SUBJECT.remove(opt.val_subject)
sample = create_H36_dataloader(data_root=opt.data_root,
batch_size=opt.batch_size,
subjects=TRAINING_SUBJECT)
image, pose3d_gt = sample['image'], sample['pose3d']
# Normalize pose
pose3d_gt_norm = normalize_pose(pose3d_gt)
# Predict pose
with tf.variable_scope("model", reuse=False):
model = resnet_model.Model()
pose3d_out_norm = model(image, training=True)
# Compare with GT
loss = tf.losses.absolute_difference(pose3d_gt_norm, pose3d_out_norm)
# Optimize network parameters
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) # for batch norm
with tf.control_dependencies(update_ops):
train_op = tf.train.AdamOptimizer(learning_rate=opt.lr).minimize(loss)
# Unnormalize pose
pose3d_out = unnormalize_pose(pose3d_out_norm)
# Validation graph
if not opt.val_subject == -1:
def build_graph(self, model='mnist'):
xinit = tf.contrib.layers.xavier_initializer
binit = tf.constant_initializer(0.0)
relu = tf.nn.relu
g = tf.get_default_graph()
self.lr = tf.placeholder(tf.float32, name='lr')
self.beta1 = tf.placeholder(tf.float32, name='beta1')
self.eps_t = tf.placeholder(tf.float32, name='eps_t')
self.trn_ph = tf.placeholder(tf.bool, name='train_ph')
self.weight_decay = tf.placeholder(tf.float32, name='weight_decay')
if model == 'mnist':
self.x = tf.placeholder(tf.float32,
shape=(None, 28, 28, 1),
name='input_ph')
self.y = tf.placeholder(tf.int32, shape=(None, ), name='label_ph')
out = self.x
out = tf.layers.conv2d(out,
32,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.layers.conv2d(out,
64,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.layers.conv2d(out,
128,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.reshape(out, [-1, np.prod(out.get_shape().as_list()[1:])])
out = tf.layers.dropout(out, rate=0.5, training=self.trn_ph)
logits = tf.layers.dense(out,
10,
kernel_initializer=xinit(),
bias_initializer=binit)
self.pred = tf.argmax(logits, axis=1, name='pred_op')
elif model == 'mnist_binary':
self.x = tf.placeholder(tf.float32,
shape=(None, 28, 28, 1),
name='input_ph')
self.y = tf.placeholder(tf.int32, shape=(None, ), name='label_ph')
out = self.x
out = tf.layers.conv2d(out,
32,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.layers.conv2d(out,
64,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.layers.conv2d(out,
128,
3,
strides=2,
activation=relu,
padding='same',
kernel_initializer=xinit(),
bias_initializer=binit)
out = tf.reshape(out, [-1, np.prod(out.get_shape().as_list()[1:])])
out = tf.layers.dropout(out, rate=0.5, training=self.trn_ph)
logits = tf.layers.dense(out,
2,
kernel_initializer=xinit(),
bias_initializer=binit)
self.pred = tf.argmax(logits, axis=1, name='pred_op')
elif model == 'cifar10':
self.x = tf.placeholder(tf.float32,
shape=(None, 32, 32, 3),
name='input_ph')
self.y = tf.placeholder(tf.int32, shape=(None, ), name='label_ph')
resnet_size = 20
num_blocks = (resnet_size - 2) // 6
conv = resnet_model.Model(
resnet_size=resnet_size,
bottleneck=False,
num_classes=10,
num_filters=16,
kernel_size=3,
conv_stride=1,
first_pool_size=None,
first_pool_stride=None,
block_sizes=[num_blocks] * 3,
block_strides=[1, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_last')
logits = conv(self.x, self.trn_ph)
self.pred = tf.argmax(logits, axis=1, name='pred_op')
elif model == 'cifar100':
self.x = tf.placeholder(tf.float32,
shape=(None, 32, 32, 3),
name='input_ph')
self.y = tf.placeholder(tf.int32, shape=(None, ), name='label_ph')
resnet_size = 20
num_blocks = (resnet_size - 2) // 6
conv = resnet_model.Model(
resnet_size=resnet_size,
bottleneck=False,
num_classes=100,
num_filters=16,
kernel_size=3,
conv_stride=1,
first_pool_size=None,
first_pool_stride=None,
block_sizes=[num_blocks] * 3,
block_strides=[1, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_last')
logits = conv(self.x, self.trn_ph)
self.pred = tf.argmax(logits, axis=1, name='pred_op')
loss = tf.losses.sparse_softmax_cross_entropy(logits=logits,
labels=self.y)
# Add weight decay to the loss.
def exclude_batch_norm(name):
return 'BatchNorm' not in name
loss_filter_fn = exclude_batch_norm
l2_loss = self.weight_decay * tf.add_n([
tf.nn.l2_loss(tf.cast(v, tf.float32))
for v in tf.trainable_variables() if loss_filter_fn(v.name)
])
loss += l2_loss
# optimizer update
opt = self.default_hparams()['opt']
if opt.lower() == 'adam':
self.opt = tf.train.AdamOptimizer(learning_rate=self.lr,
beta1=tf.reshape(self.beta1, []))
elif opt.lower() == 'momentum':
self.opt = tf.train.MomentumOptimizer(learning_rate=self.lr,
momentum=tf.reshape(
self.beta1, []))
self.opt_op = self.opt.minimize(loss)
# langevin updates
grads, varlist = list(zip(*self.opt.compute_gradients(loss)))
grads = [(g + tf.random_normal(
g.get_shape().as_list(), mean=0.0, stddev=self.eps_t))
for g in grads]
self.lang_op = self.opt.apply_gradients(list(zip(grads, varlist)))
def tower_loss(scope, images, labels, is_training, human_to_label):
"""Calculate the total loss on a single tower running the model.
Args:
scope: unique prefix string identifying the CIFAR tower, e.g. 'tower_0'
images: Images. 4D tensor of shape [batch_size, height, width, 3].
labels: Labels. 2D tensor of shape [batch_size, num_classes].
Returns:
Tensor of shape [] containing the total loss for a batch of data
"""
# TODO scope? weight decay? validation_test? embedding? accuracy? saver? restored?
# TODO Calculate loss, which includes softmax cross entropy and L2 regularization.
# Build inference Graph.
model = resnet_model.Model(
resnet_size=18,
bottleneck=False,
num_classes=data_utils.NUM_CLASSES,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(18),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_last',
dtype=tf.float32)
logits = model(images, training=is_training)
# weights = tf.multiply(tf.cast(labels, tf.float32), 1000.0) + 1.0 # example: labels[0,0,1,1,0] -> weights[1,1,7179,7179,1]
cross_entropy = tf.losses.sigmoid_cross_entropy(
logits=logits,
multi_class_labels=labels,
weights=1.0,
label_smoothing=0.0)
# log summary of single gpu on tensorboard
if 'tower_0' in scope:
tf.summary.histogram('logits', logits, collections=LOG_COLLECTIONS)
tf.summary.scalar(
'loss_cross_entropy',
cross_entropy,
collections=LOG_COLLECTIONS,
family='iic')
logits_round = tf.cast(tf.round(tf.sigmoid(logits)), tf.int32)
CM = tf_utils.Confusion_Matrix(logits_round, labels)
tf.summary.scalar(
'MEAN/f2_score',
CM.f2_score,
collections=LOG_COLLECTIONS,
family='iic')
tf.summary.scalar(
'MEAN/precision',
CM.precision,
collections=LOG_COLLECTIONS,
family='iic')
tf.summary.scalar(
'MEAN/recall',
CM.recall,
collections=LOG_COLLECTIONS,
family='iic')
for label_name in TARGET_LABELS:
label_int = human_to_label[label_name]
CM_target = tf_utils.Confusion_Matrix(logits_round[:, label_int],
labels[:, label_int])
tf.summary.scalar(
'{}/f2_score'.format(label_name),
CM_target.f2_score,
collections=LOG_COLLECTIONS,
family=label_name)
tf.summary.scalar(
'{}/precision'.format(label_name),
CM_target.precision,
collections=LOG_COLLECTIONS,
family=label_name)
tf.summary.scalar(
'{}/recall'.format(label_name),
CM_target.recall,
collections=LOG_COLLECTIONS,
family=label_name)
# # Assemble all of the losses for the current tower only.
# losses = tf.get_collection('losses', scope)
# # Calculate the total loss for the current tower.
# total_loss = tf.add_n(losses, name='total_loss')
# # Attach a scalar summary to all individual losses and the total loss; do the
# # same for the averaged version of the losses.
# for l in losses + [total_loss]:
# # Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
# # session. This helps the clarity of presentation on tensorboard.
# loss_name = re.sub('%s_[0-9]*/' % cifar10.TOWER_NAME, '', l.op.name)
# tf.summary.scalar(loss_name, l)
return cross_entropy
def eval():
LABEL_TO_CLASS_PATH = '../inputs/label_to_class.json'
with open(LABEL_TO_CLASS_PATH, 'r') as infile:
label_class_mapping = json.load(infile)
with tf.Graph().as_default() as graph, tf.device('/cpu:0'):
data_dict = data_generator()
batch_filenames = data_dict['name']
batch_images = data_dict['image']
# Calculate the gradients for each model tower.
is_training = tf.placeholder(tf.bool)
with tf.variable_scope(tf.get_variable_scope()):
with tf.device('/gpu:0'):
model = resnet_model.Model(
resnet_size=18,
bottleneck=False,
num_classes=data_utils.NUM_CLASSES,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(18),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_model.DEFAULT_VERSION,
data_format='channels_last',
dtype=tf.float32)
logits = model(batch_images, training=is_training)
logits_round = tf.cast(tf.round(tf.sigmoid(logits)), tf.int32)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
init_local = tf.local_variables_initializer()
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True)) as sess:
sess.run([init, init_local])
# sess.run(iter_init)
if FLAGS.restore_path is not None:
saver.restore(sess, FLAGS.restore_path)
print('successfully restore model from checkpoint: %s' %
(FLAGS.restore_path))
# Create a coordinator and run all QueueRunner objects
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
predicted = []
start_time = time.time()
while True:
try:
filenames, predictions = sess.run(
[batch_filenames, logits_round],
feed_dict={is_training: False})
conds = np.not_equal(predictions, 0)
code_result = []
for cond in conds:
results = np.where(cond)
results = list(
map(
lambda x: label_class_mapping[
'label_to_code'][str(x)], results[0]))
str_result = ''
for res in results:
str_result = str_result + ' ' + res
code_result.append(str_result)
filenames = list(
map(lambda x: os.path.split(x)[1][:-4], filenames))
for fname, result in zip(filenames, code_result):
predicted.append({
'image_id': fname.decode('utf-8'),
'labels': result
})
print(len(predicted))
except tf.errors.OutOfRangeError:
duration = time.time() - start_time
print('OutOfRangeError, and time cost: {}'.format(
duration))
submission = pd.read_csv(
'../labels/stage_1_sample_submission.csv',
index_col='image_id')
tuning_labels = pd.read_csv(
'../labels/tuning_labels.csv',
names=['id', 'labels'],
index_col=['id'])
predicted_df = pd.DataFrame.from_dict(predicted,
orient='columns')
predicted_df = predicted_df.set_index('image_id')
submission['labels'] = None
submission.update(predicted_df)
submission.update(tuning_labels)
submission.to_csv(FLAGS.result_path)
break
# Stop the threads
coord.request_stop()
# Wait for threads to stop
coord.join(threads)