New Year Release 1.3.0

Happy new year everyone!

Thanks for the contributions from the following people, TensorLayer has fast development in the pass few months.

@zsdonghao @wagamamaz @luomai @shorxp @angerhang
@sunbohit @sczhengyabin @michuanhaohao @Yugnaynehc
@deepxiangfa @akaraspt @qxin @akaitsuki-ii @todtom
@boscotsang @JingqingZ @lgarithm

---

This is an incomplete list among the many features added.

• Update
  • Support BiDynamicRNNLayer by @akaitsuki-ii
  • Support GaussianNoiseLayer by @zsdonghao
  • Make InputLayer compatible with tf.Variable input by @qxin
  • Make tf.GraphKeys.VARIABLES compatible with TF12 by @sczhengyabin
  • Make nlp/read_words() compatible with Python3.5 by @todtom
• Maintain English documentation by @zsdonghao
• Maintain Chinese documentation by @zsdonghao

This version works well under TF 0.10 and 0.11 backend, it also works well with TF 0.12 in most of the case, feel free to report if you found a bug. Many thanks!

Happy new year again !

Best wishes
TensorLayer contributors

Release 1.2.8

Recommended Update !

This is an incomplete list among the many features added.

• Update
  • Speed up model saving and loading by @sczhengyabin
  • Fix ｀ConcatLayer` bug found by @tobymu and @michuanhaohao
  • Fix BatchNormLayer compatibility to TF12 by @boscotsang
  • Update tutorial_tfrecord3.py by @Yugnaynehc
  • Hard Dice coefficient by @zsdonghao
  • Cosine similarity by @zsdonghao
  • The keep_prob of DropoutLayer can be fixed by @zsdonghao, then instead of changing the keep_prob in placeholder (see tutorial_mnist.py), we can also build inferences for training and testing as follow:

def inference(x, is_train)
    network = ...
    if is_train=True:
        network = DropoutLayer(network, keep=0.8, is_fix=True, name='drop1')
    ....

• Maintain English documentation by @zsdonghao
• Maintain Chinese documentation by @zsdonghao

This version works well under TF 0.10 and 0.11 backend, it also works well with TF 0.12 in most of the case, feel free to report if you found a bug. Many thanks!

Release 1.2.7

Recommended Update !

This is an incomplete list among the many features added.

• Update
  • imresize, crop
  • Better implementation of sample_top_k()
• Example
  • DCGAN by @zsdonghao.
• Bugs
  • Fix Conv1dLayer bug by @michuanhaohao
  • Fix threading_data bug by @luomai
• Developing
  • Simplified CNN APIs
  • State-of-the-art Attention Seq2seq APIs.
• Maintain Chinese documentation by @zsdonghao
• Maintain Chinese documentation by @shorxp @zsdonghao @wagamamaz
• Typo found by @Yugnaynehc

This version work well under TensorFlow 0.10 and 0.11 backend.

Release 1.2.6

Recommended Update !

This is an incomplete list among the many features added.

• New APIs
  • More data augmentation methods like elastic transform. However, for large dataset, we recommend to use TFRecord to speed up the loading time.
  • APIs for geting variables with given name.
  • State-of-the-art cost functions for image segmentation @zsdonghao.
  • APIs for calculating the sequence length for dynamic RNN.
  • APIs for disable printing and empty trash.
• Update
  • Add beta gamma initializer for BatchNormLayer by @JingqingZ.
  • Better implementation of binary cross entropy.
  • FQA on RTD website.
• Developing
  • Simplified CNN APIs.
  • State-of-the-art Attention Seq2seq APIs.
• Example
  • Image Captioning by @zsdonghao.
  • Wild ResNet by @ritchieng.
• Fix bug found by @narrator-wong.

1.2.4 and 1.2.5 can be found on PyPi.

This version work well under TensorFlow 0.10 and 0.11 backend.

Release 1.2.3

Recommended Update !

This is an incomplete list among the many features added

• Release a number of data preprocessing functions using threading and queue (no dependency on TFRecord) by @zsdonghao. However, for large dataset, we recommend to use TFRecord to store images and to speed up the loading time.
• Release and updates layers
  • Update BatchNormLayer by @deepxiangfa @akaraspt , better implementation and better default setting.
  • Update LambdaLayer by @ritchieng
  • Update SlimNetsLayer by @zsdonghao, all TF-Slim pre-trained model can be connected with TensorLayer easily.
  • Release ElementwiseLayer by @zsdonghao
  • Update DeconvLayer by @wagamamaz
• Release state-of-the-art cost functions for image segmentation @zsdonghao
• Fix bug by @deepxiangfa
• FQA on RTD
• Maintain Chinese documentation by @shorxp
• Shortcut : activation -> act

This version work well under TensorFlow 0.10 and 0.11 backend.

Release 1.2.2

This is an incomplete list among the many features added

• Support more layers
  • 3D Deconvolutional layer by @fangde
  • 2D Deconvolutional layer by @fangde @zsdonghao
• Merge with TF-Slim, all Google's pre-trained CNN models can be used in TensorLayer by @zsdonghao
• More examples
  • Inception-V3 by @zsdonghao
  • MultiplexerLayer by @zsdonghao
• Maintain Chinese documentation by @shorxp @zsdonghao @wagamamaz

This version work well under TensorFlow 0.10 backend.

For next version, we need help to add more functions of dynamic RNN, TFRecord and TensorFlow Mobile.