TheanoLayer is a light library written with theano. It is conveninet for fast building and training DL models. Since MILA has stop developing theano and tensorflow is the most dominant static graph library， I still find that theano is more speed and memory efficiency for some models, especially seq2seq model in NLP which heavily use for-loop structure.
-
Convenient shared variables creation, sharing and serilization.
-
Sparse gradient update for lookup table.
-
Memory efficient optimizer
-
functional: Build compatition graph given input tensors and shared variables. Like functions in python.
-
Module: A combination of functionals and shared variables. Modules are all python classes which inherit a base class Module.
-
layer: A combination of different modules. It is a high-level interface for some frequent-used layers when building networks, such as dense layer or bidirectional gru layer.
-
utils: utility functions
-
nnet: funtionals (build computation) and modules
-
layers: neural layers
-
parameters.py: A global shared variables management mechanism.
-
variable.py: A wrapper of theano shared variable for gradients accumulation and sparse update.
-
optimizers.py: Gradient optimizer algorithms.
This library borrows some codes from Lasagne and Tensorflow.
Haoran Wei(whr94621@163.com, whr94621@gmail.com)