Texar Versions Save

New features

• Support only Python 3.6 and 3.7. Drop support of older Python versions. (#211)
• Add Tokenizers including tokenizers for pretrained models (BERTTokenizer, XLNetTokenizer, etc). (#225)
• Add GPT2 modules (GPT2Encoder, GPT2Decoder, GPT2Classifier, etc). (#228)

Feature improvements

• Update embedder modules dropout_strategy=='item' to support TensorFlow v1.15. (#231)
• Update .gitignore and add .gitignore files to all examples. (#233)
• Polish code style according to flake8. (#234)
• Add GPT2 XL pretrained checkpoint. (#243)

Fixes

• Fix examples/transformer/scripts/wmt14_en_de.sh to create output dir automatically. (#238)
• Fix variable scope issue in texar.tf.modules.decoders.dynamic_decode. (#246)

New features

• Texar Tensorflow now imported by import texar.tf as tx. (#197)
• Add pretrained modules [docs], e.g., BERT, XLNet, etc. (#206)
• Add new tensor shape related utils, e.g., reduce_with_weights, varlength_*. (#201)

Feature improvements

Fixes

• Fix docstring of connector _mlp_transform. (#192)
• Fix dynamic_decode out-of-range error when max-decoding-length is reached. (#208)

New features

• Enable installation from PyPI. (#186)
• Update dependency to TF>=1.10.0. (#186)

Feature improvements

• Use lazy import to be compatible with texar-pytorch. (#183)

Fixes

• Fix or toggle of all warnings (mainly TF deprecation warnings). (#186)

New features

• Add support for GPT-2 345M model in examples/gpt-2. (#156)
• Add BERT modules, including texar.modules.BERTEncoder (doc) and texar.modules.BERTClassifier (doc). (#167)

Feature improvements

• Refactor TransformerEncoder and TransformerDecoder to separate position embeddings from the modules. (#126)
• Allow passing a Tensor to output_layer of decoders' constructors -- used for weight tie b/w the output layer and input embedding matrix. (#126)
• TransformerDecoder constructor interface made exact the same with RNN decoders constructor interfaces. (#126)
• Refactor decoder Helpers to allow two-argument embedding_fn (supporting for position embedding). (#126)
• Refactor SinusoidsPositionEmbedder to enable infinite large or negative position indexes. (#176)

Fixes

• Fix texar.losses.reduce_batch_time when sequence has dtype other than tf.float32. (#143)
• Fix texar.losses.reduce_dimensions when average_axes or sum_axes is int. (#141)
• Fix GPT-2 tokenization loading path. (#165)
• Fix examples/vae_text EOS bug. (#168)
• Fix transformer bleu_tool.py when translation_length is 0. (#176)
• Fix StochasticConnector and ReparameterizedStochasticConnector when transform=False. (#179)

New features

• TFRecordData: A new data module for reading and processing TFRecord data, with support for, e.g., image data, feature data, etc. (#107)
• GPT-2: OpenAI pretrained language model. (#91, example)
• TopKSampleEmbeddingHelper to perform top_k random sample decoding. (baa09ff)

Feature improvements

• Refactor BERT example using TFRecordData data module.
• TransformerDecoder supports helper arguments to specify decoding strategy. (#76)

Fixes

• Fix variable collection bug in examples/seqgan. (#110)
• Fix error when beam_search_decode with output_layer=tf.identity (#77)
• Fix readthedocs compilation error (#85)

Texar v0.1.0

Initial release