Implementation of Attention is All You Need in Sonnet/Tensorflow

Architecture:

Paper: https://arxiv.org/abs/1706.03762

Usage

1. Install requirements pip install -r requirements.txt
2. Install sonnet
3. Run run_micro_services.sh

Organisation of the repository

Transformer's architecture is composed of blocks. The uses of Sonnet makes the implementation very modular, and reusable. I tried to keep the blocks as much decouple as possible, following the paper:

• attention/algorithms/transformers: Create an tf.contrib.learn.Experiment, and the tf.contrib.data.Dataset

• attention/modules/cores: Implementation of the core blocks of Transformer such as MultiHeadAttention, PointWiseFeedForward

• attention/modules/decoders: Implementation of a Decoder block, and a Decoder

• attention/modules/encoders: Implementation of an Encoder block, and an Encoder.

• attention/models: Implementation of a full Transformer Block. This Module is responsible to create the Encoder and the Decoder

• attention/services: Micro Services that create the dataset, or train the model

• attention/utils: Some classes uses as utility (recursive namespace, mocking object)

• attention/*/tests/: Test of the Module/Algorithm/MicroService

Training Task implemented

• [X] Copy inputs
• [ ] Dialogue generation

Road Map

• [X] Code modules
• [X] Test modules
• [X] Construct input function
• [X] Build Estimator
• [X] Run estimator
• [X] Plug into a workflow
• [X] Add validation queue
• [ ] Iterate over model improvements