Some minor bugs. - Githubissues

HasnainRaz / FC-DenseNet-TensorFlow

Fully Convolutional DenseNet (A.K.A 100 layer tiramisu) for semantic segmentation of images implemented in TensorFlow.

MIT License

123 stars 41 forks source link

Some minor bugs. #12

Closed Ouwen closed 5 years ago

Ouwen commented 5 years ago

weight init In the original paper, he initialization is used instead of xavier, this would translate to using: tf.contrib.layers.variance_scaling_initializer
- max pooling In the original paper, the transition down should be a 2x2 max pool that is non-overlapping. Checking the original Lasagne code it seems they use a 2x2 max pool with a window of 2x2 and stride of 2x2. Wasn't sure if your current 4x4 window as intentional.

Great work on this btw, very readable!

Ouwen commented 5 years ago

https://github.com/HasnainRaz/FC-DenseNet-TensorFlow/blob/e709d6376c2427fd4701e5b3ad7fcd48231fa5ab/model.py#L162

vs original code

HasnainRaz commented 5 years ago

Thank you for the feedback.

Yeah sorry about those, I made this because of a project I was working on, so I may have changed the pooling window and weight inits. Good catch, I'll restore them to the originals.