Styjun

I'm working with a a neural network for image segmentation using Tensorflow.

The training phase and inference run are ok if the is_traning parameter of the slim.batch_norm layer is set to True.
But when i run the session using is_training as false, meaning (in what i understand) just infering/forwarding data through the network the result segmentation image data comes out empty.

I believe it has to do with batchNorm layers but i've already lost my mind over it and i just can't make it work.

I'm using a code based in Semantic Segmentation Suite in TensorFlow.
Below is a simplified version of what works and what fails.

.....
def ConvBlock(inputs, n_filters, kernel_size=[3, 3],is_training=True):
 net = slim.conv2d(inputs, n_filters, kernel_size=[1, 1], activation_fn=None)
 net = slim.batch_norm(net, fused=True, is_training=is_training)
 net = tf.nn.relu(net)
 return net

def DepthwiseSeparableConvBlock(inputs, n_filters, kernel_size=[3, 3],is_training=True):
 net = slim.separable_convolution2d(inputs, num_outputs=None, depth_multiplier=1, kernel_size=[3, 3], activation_fn=None)
 net = slim.batch_norm(net, fused=True, is_training=is_training)
 net = tf.nn.relu(net)
 ....
 return net

def ConvTransposeBlock(inputs, n_filters, kernel_size=[3, 3],is_training=True):
 net = slim.conv2d_transpose(inputs, n_filters, kernel_size=[3, 3], stride=[2, 2], activation_fn=None)
 net = slim.batch_norm(net,is_training=is_training)
 net = tf.nn.relu(net)
 return net

def build_mobile_unet(inputs, .... ,is_training=True):
 net = ConvBlock(inputs, 64, is_training=is_training)
 net = DepthwiseSeparableConvBlock(net, 64, is_training=is_training)
 net = slim.pool(net, [2, 2], stride=[2, 2], pooling_type='MAX')
 ....
 net = ConvTransposeBlock(net, 64, is_training=is_training)
 net = DepthwiseSeparableConvBlock(net, 64, is_training=is_training)
 net = DepthwiseSeparableConvBlock(net, 64, is_training=is_training)
 net = slim.conv2d(net, num_classes, [1, 1], activation_fn=None, scope='logits')
 return net



# Define the param placeholders 
net_input_image = tf.placeholder(tf.float32,shape=[None,None,None,3], name="input") 
net_input_label = tf.placeholder(tf.int32, [None,None,None])
# Training phase placeholder
net_training = tf.placeholder(tf.bool, name='phase_train')

model, _ = build_mobile_unet(
 net_input=net_input_image, 
 ....
 is_training=net_training)

model = tf.nn.softmax(model, name="softmax_output")

with tf.name_scope('loss'):
 cross_entropy =tf.losses.sparse_softmax_cross_entropy(logits=model, labels=net_input_label)
cross_entropy = tf.reduce_mean(cross_entropy)

# use RMSProp to optimize
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
 optimizer = tf.train.RMSPropOptimizer(learning_rate=0.0001,decay=0.995)
 train_step = optimizer.minimize(cross_entropy)

# create train OP
total_loss = tf.losses.get_total_loss()
train_op = slim.learning.create_train_op(total_loss,optimizer)


# Do the training here
for epoch in range(args.epoch_start_i, args.num_epochs):

 input_image_batch = ...
 label_image_batch = ...

 # Do the training
 train_dict=
 net_input_image:input_image_batch,
 net_input_label:label_image_batch,
 net_training: True
 
 train_loss=sess.run(train_op, feed_dict=train_dict)


 # Do the validation on a small set of validation images
 for ind in val_indices:

 input_image = np.expand_dims(np.float32(utils.load_image(val_input_names[ind])[:args.crop_height, :args.crop_width]),axis=0)/255.0
 gt = utils.load_image(val_output_names[ind])[:args.crop_height, :args.crop_width]
 gt = helpers.reverse_one_hot(helpers.one_hot_it(gt, label_values))

 # THIS WORKS : Image segmentation result is OK
 output_image = sess.run(
 model,
 feed_dict=
 net_input_image:input_image,
 net_training: True
 )
 # THIS FAILS : Image segmentation result is all Zeros....
 output_image = sess.run(
 model,
 feed_dict=
 net_input_image:input_image,
 net_training: False
 )

The training works well, and the net converges and all....
and If i allways keep the placeholder net_training as True, all is well.

But is i invoke the sess.run(model,...net_training: False) as you can see in the code above while testing some images the output result comes out empty.

What i'm i doing wrong guys?
Any help would be highly appreciated.
Thank you for your time.