안녕하세요. 조금 혼란 스러운 부분이 있어서 질문을 드립니다.
해당 라인에서 한 배치마다 Test error를 계산한 것을 볼 수 있었습니다. validation 처럼 사용하신 것 같은데요. 그래서 작성하신 test_loss, test_accuracy를 전 val_loss, val_accuracy로 이해했습니다. 그런데 데이터 전처리에서 validation 데이터를(실제로test 데이터로 이름 붙이신) Train 데이터 mean, std로 노말라이즈 한것을 볼 수 있었는데요.

제가 알기론 validation 데이터는 validation mean, std로 노말라이즈 하는 걸로 알고 있는데 제가 잘 못 알고 있는 건가요?

아니면 코드 그대로 배치마다 테스트 데이터를 돌려 보신건가요?. 궁금합니다.

I ran this project on ubuntu server 16.04.
I installed this below requirements for this project
tensorflow==1.14.0 keras==2.2.5 matplotlib==3.1.1 Pillow==6.1.0 scipy==1.1.0

I just cloned this project and then ran as the instruction in README.md file by following command:
python main.py --phase train --dataset tiny --res_n 18 --lr 0.1.
After ran for a while encountered this issue [*] Failed to find a checkpoint . The error came from this statement tf.train.get_checkpoint_state(checkpoint_dir).
So what is the correct argument

checkpiont_dir

for this function?

When I set --res_n 50, the code will report a error.

def resblock(x_init, channels, is_training=True, use_bias=True, downsample=False, scope='resblock') :
with tf.variable_scope(scope) :

    x = batch_norm(x_init, is_training, scope='batch_norm_0')
    x = relu(x)

    if downsample :
        x = conv(x, channels, kernel=3, stride=2, use_bias=use_bias, scope='conv_0')
        x_init = conv(x_init, channels, kernel=1, stride=2, use_bias=use_bias, scope='conv_init')

    else :
        x = conv(x, channels, kernel=3, stride=1, use_bias=use_bias, scope='conv_0')

    x = batch_norm(x, is_training, scope='batch_norm_1')
    x = relu(x)
    x = conv(x, channels, kernel=3, stride=1, use_bias=use_bias, scope='conv_1')

    return x + x_init

Dimensions of x and x_init are not equal.

Hi,taki0112:
i had read you code for learning. Great code to read and learn.
there is a mistake in the README.md file. "--ren_n" should be "--res_n".

i look your commit,and could you explain why you chang
----x_init = avg_pooling(x_init)
----x_init = conv(x_init, channels, kernel=1, stride=1, use_bias=use_bias, scope='conv_init')
to
----x_init = conv(x_init, channels*4, kernel=1, stride=2, use_bias=use_bias, scope='conv_init')
??

I tried this code and it seems that the resulting train/test accuracies are not changed no matter how high the regularization parameter is (in ops.py). A little digging in tensorflow documentation reveals that in order for regularization to work, you had to add the following line in build_model() in class ResNet.

self.train_loss+=tf.losses.get_regularization_loss()

In the original paper, there are 7x7 conv and 3x3 max pooling layer before residual block.

I wonder how to load pretrained network, thanks