0.2.0 (WIP)

[0x00b1]

Features

• [ ] One-dimensional models
• [ ] Three-dimensional models
• [ ] Time-distributed one-dimensional blocks and models
• [ ] Time-distributed two-dimensional blocks and models
• [ ] Time-distributed three-dimensional blocks and models
• [ ] Group normalization layer
• [ ] Upsample layer
• [ ] Two-dimensional feature pyramids
• [ ] Updated benchmark suite
• [ ] Weights for two-dimensional ResNet-50

Depreciations

• [ ] There’s a few. I’ll update this soon.

To-do

• [ ] Upsample layer tests
• [ ] resize tests
• [ ] CNTK resize implementation
• [ ] Theano resize implementation
• [ ] Update CIFAR-10 benchmark hyperparameters
• [ ] Update CIFAR-100 benchmark hyperparameters

[0x00b1]

@hgaiser Re: use_bias. That’s a mistake. I’ll fix shortly. 😆

[0x00b1]

@hgaiser And thanks for the feedback!

[0x00b1]

@hgaiser I think the BatchNormalization layer requires a training=False in the call, otherwise it will update its weights.

I missed this. I assumed it was in the constructor, but you’re right it’s in the call. Thanks, dude!

Edit: And it’s training not trainable! 😳

Edit: And call doesn’t take **kwargs. Woof.