tridentx 0.6.1

Multiverse for Deep Learning Developers without Pitfall

Multiverse for Deep Learning Developers without Pitfall

Release Note 0.6.0

New Features:

1. new TextSequenceDataset and TextSequenceDataProvider for prepare dense/ sparse text tensor.

2. rnn/ lstm/ gru is supporting now ( in pytorch backend).

3. reinforcement learning is supporting now ( in pytorch backend).

4. normalization in sequence supporting (automatic permute)

5. bounding box inplace operation. (no need to split first 4 element)

6. New optimizer: RangerLars, and gradient centralization is now supporting.

7. New method ‘trigger_when’ in Model class , make add new callbacks or new event-listening much more easy.

Bug fix: #. min, max, reduce_min, reduce_max bad arguments issue.

0.5.9 #. fix the Model argments ordinal bug. #. fix the deprecated addcdiv_ functions

0.5.8 #. fix the deprecated divide ops and add_ functions

0.5.7 #. Support AMP (automatic mixed precision) with pytorch 1.6. #. attach lots of functions in ops on the tensor and use @numpy_compatible to share same syntax within numpy array or tensor. #. New attribute ‘sequence_rank’ comes to all conv_blocks familay, if sequence_rank=’cna’ means the order is ‘Convolution-Normalization-Activation’. #. Adding EvoNorm, SIREN ..