Update README.md

README.md

@@ -1,6 +1,6 @@
-# DA-Transformer
+# FA-DAT
 
-Implementation for the ICML2022 paper "**Directed Acyclic Transformer for Non-Autoregressive Machine Translation**".
+Implementation for the ICLR2023 paper "**Fuzzy Alignments in Directed Acyclic Graph for Non-autoregressive Machine Translation**".
 
 **Abstract**: Directed Acyclic Transformer (DA-Transformer) represents the hidden states in a Directed Acyclic Graph (DAG), where each path of the DAG corresponds to a specific translation. The whole DAG simultaneously captures multiple translations and facilitates fast predictions in a non-autoregressive fashion.
 
@@ -12,8 +12,6 @@ Implementation for the ICML2022 paper "**Directed Acyclic Transformer for Non-Au
 
 
 
-![model](model.png)
-
 
 
 This repo is modified from [``fairseq:5175fd``](https://github.com/pytorch/fairseq/tree/5175fd5c267adceec9445bf067597686e159e7e7), please see [fairseq documentation](https://fairseq.readthedocs.io/en/latest/) for more information.
@@ -35,50 +33,16 @@ This repo is modified from [``fairseq:5175fd``](https://github.com/pytorch/fairs
 
 ## Main Files
 
-Most codes of the framework are from Fairseq. We mainly add the following files.
+Most codes of the framework are from Fairseq and DA-Transformer. We mainly add the following file.
 
 ### fs_plugins
 
 ```
 fs_plugins
 ├── criterions
-│   └── nat_dag_loss.py                   # DA-Transformer loss
-├── cub                                   # Requirements: Nvidia CUDA programming model
-├── custom_ops                            # operations implementations and cuda kernels
-│   ├── dag_best_alignment.cu
-│   ├── logsoftmax_gather.cu
-│   ├── dag_loss.cu
-│   ├── dag_loss.py
-│   └── dag_loss.cpp
-├── models
-│   ├── glat_decomposed_with_link.py      # A PyTorch implementation of DA-Transformer
-│   ├── ls_glat_decomposed_with_link.py   # A lightseq implementation of DA-Transformer
-│   └── ls_*                              # Other files required for lightseq
-├── optimizer
-│   └── ls_adam.py                        # Lightseq Adam
-├── scripts
-│   ├── test_tradeoff.py                  # Parameter search script used in BeamSearch
-│   ├── average_checkpoints.py            # Average checkpoints tricks
-|   └── convert_ls_to_fairseq.py          # Converting lightseq model to fairseq model
-└── tasks
-    └── translation_lev_modified.py
+   └── nat_dag_loss_ngram.py                   # fuzzy alignment loss
 ```
 
-### Modified LightSeq for NAT
-
-We include a customized [LightSeq](https://github.com/thu-coai/lightseq-nat/). The changes include:
-
-* Implement the non-autoregressive decoder based on the LightSeq autoregressive decoder
-* Increase the supported max length (1024 for now)
-* Align the parameters and model architectures with Fairseq implementation, and provide a script for checkpoint conversion.
-
-### BeamSearch on DAG
-
-We include [dag_search](https://github.com/thu-coai/DAG-Search) to implement the BeamSearch algorithm.
-
-## Data Preprocessing
-
-Please follow the [instruction](https://github.com/facebookresearch/fairseq/tree/main/examples/translation#wmt14-english-to-german-convolutional) in Fairseq to prepare the data.
 
 The training require binarized data generated by the following script: