History
version 0.9.4 [3-30-2021]
- NEW: add
util.torch_safe_run() for recovering train/inference process from CUDA OOM
version 0.9.3 [2-26-2021]
- MODIFIED: add
mask input/output support for functional.dropout()
version 0.9.2 [12-3-2020]
- MODIFIED: replace
torch.clamp() with in-place Tensor.clamp_() for util.grad_clip()
version 0.9.1 [11-25-2020]
- NEW:
functional.dropout() now supports 1) varying dropout probability 2) replace dropped values with given fill_value
version 0.9.0 [11-24-2020]
- NEW: add online documentation
- NEW: add
functional.dropout()
version 0.8.0 [9-8-2020]
- NEW: add
BatchNorm module for replacing pytorch's implementation
version 0.7.6 [5-7-2020]
- MODIFIED: add support for gzip compression for
gpickle.loads() & gpickle.dumps() functions
version 0.7.5 [1-8-2020]
- NEW: add
util.get_file_md5() for file md5 check
version 0.7.4 [1-3-2020]
- NEW: add
util.set_value() for tensor initialization with numpy array
version 0.7.3 [12-26-2019]
- NEW: Add
util.freeze_module()/unfreeze_module()/get_trainable_parameters() for training parameters handling
version 0.7.2 [11-5-2019]
- NEW: Add
util.verbose_print() for print with verbose level filtering