More RL and callbacks

[Borda]

[0.4.0] - 2021-09-09

Added

• Added Soft Actor Critic (SAC) Model (Soft Actor Critic (SAC) Model #627)
• Added EMNISTDataModule, BinaryEMNISTDataModule, and BinaryEMNIST dataset (Add EMNISTDataModule #676)
• Added Advantage Actor-Critic (A2C) Model (Advantage Actor Critic (A2C) Model #598)
• Added Torch ORT Callback (Add Torch ORT Callback, re-write callback section #720)
• Added SparseML Callback (Add the sparseml callback #724)

Changed

• Changed the default values pin_memory=False, shuffle=False and num_workers=16 to pin_memory=True, shuffle=True and num_workers=0 of datamodules (Use pin_memory=True, shuffle=True and num_workers=0 by default #701)
• Supporting deprecated attribute usage (Support deprecated attribute usage #699)

Fixed

• Fixed ImageNet val loader to use val transform instead of train transform (Fix ImageNet val loader to use val transform #713)
• Fixed the MNIST download giving HTTP 404 with torchvision>=0.9.1 (Fix MNIST 404 with torchvision>=0.9.1 #674)
• Removed momentum updating from val step and add separate val queue (Remove momentum updating from val step and add separate val queue #631)
• Fixed moving the queue to GPU when resuming checkpoint for SwAV model (Move queue to gpu when resuming checkpoint - SWAV self supervised model #684)
• Fixed FP16 support with vision GPT model (Set the dtype correctly for vision GPT model #694)
• Removing bias from linear model regularisation (Removing regularisation applied to linear model bias #669)
• Fixed CPC module issue (fixed CPC module arg name #680)

---

This discussion was created from the release More RL and callbacks.