[ANN] Mocha.jl v0.0.7 - Deep Learning framework for Julia

[Chiyuan Zhang]

Mocha.jl v0.0.6: Mocha is an efficient and flexible deep learning framework for Julia. This time we include some small but very useful updates.

v0.0.7 2015.02.27

• Infrastructure
  • Clearer Nesterov solver (@the-moliver)
  • Staged momentum policy
  • Learning rate policy to decay dynamically based on performance on validation set
• Network
• Async HDF5 data layer: faster and with chunking to support fast data shuffling
• Softlabel-softmax-loss layer allows training with posterior (instead of hard labels) as labels
• Weight loss layers to combine multiple loss functions in one network
• Square loss layer is now capable of propagating gradients to both sides

Best,

Chiyuan

[Chiyuan Zhang]

Please also check Boltzmann.jl (https://github.com/dfdx/Boltzmann.jl) for DBN pre-training integration example with Mocha.jl!