Tensor name "bn4b17_branch2c/moving_mean" not found in checkpoint files about deeplab-v2--resnet-101--tensorflow

[John1231983]

Hi. Did you change the encoder name to res101? I guess you are using deeplab encoder

[StevenVerne]

I want to use the deeplab, should I run the res101 first?

P.S. I download the ckpt provided by DrSleep, the code works

[John1231983]

No. Just use encoder is deeplab and pretrain is deeplab_int that pretrained in MSCOCO. That's all. If you want to run encoder is resnet, the pretrain must be resnet_v1_101. The reason is that they must be same name to use pretrained weight. One more thing, the performance using resnet pretrained 70.5% is much lower than deeplab pretrained (by author), 75.5%. I did not find the problem yet. If you find it, lets me know

[StevenVerne]

Thank you for your advise. I am just starting to use the deeplab. Happy to share, if anything interesting found

[John1231983]

@zhengyang-wang: I think the convolution in the fc layer must be initial weight (such as Gaussian), otherwise the training from resnet pretrain will be very low.Do you think so?

[zhengyang-wang]

The architecture for the decoder part (named 'fc' in deeplab and 'decoder' in res101/50) is different from the original network (e.g. ASPP instead of fully connected layers). The number of parameters is different. We cannot transfer weights in this case. That's why we randomly initialized them.

[John1231983]

@zhengyang-wang: Hi. In your convolution weight, it is

w = tf.get_variable('weights', shape=[kernel_size, kernel_size, num_x, num_o])
I did not find the initializer flag is used. Does its default as random initialization? In my opition, I think it should be

w = tf.get_variable('weights', shape=[kernel_size, kernel_size, num_x, num_o],initializer= tf.contrib.layers.xavier_initializer())

[zhengyang-wang]

For convolutional layers, we used pre-trained weights as their initialization. It does not matter what initializers you are using.

[zhengyang-wang]

And, the default initializer is https://www.tensorflow.org/api_docs/python/tf/glorot_uniform_initializer

[John1231983]

@zhengyang-wang : Thanks for your information. Actually, my question about convolutions is in the fc layer (decoder layers-ASPP). We have to initialize weight for these convolutions because we cannot borrow the weight from ResNet pre-train.