torchasn's Introduction

torchASN

A pytorch implementation of "Abstract Syntax Networks for Code Generation and Semantic Parsing".

Prerequisites

pytorch >= 1.4.0

Run on Demo Dataset

Instructions of running torchASN on DeepRegex, an NL-to-Regex dataset, are included. This implementation achieves 61.6% DFA-accuracy on DeepRegex. For comparison, the performance of a seq-to-seq model with attention is 58.2%.

To run the code,

# Preprocess dataset
python dataset/turk/make_dataset.py

# Train a model. A pretrained model is included at checkpoints/turk/pretrained.bin.
./scripts/turk/train.sh

# Test a model. <model_file> is the pointer to the model, e.g., the pretrained one mentioned above.
./scripts/turk/test.sh <model_file>

Addapting to New Programming Language

Adapting to a new programming langauge requires following steps:

Write down the domain specific language in the Abstract Syntax Description Language (ASDL) form. (refer to data/turk/turk_asdl.txt)
Implement the TransitionSystem for the new DSL to (1) convert the programs between logical forms and ASTs (2) check if a partial program (incomplete AST) is correct during beam search decoding. Refer to grammar/turk/turk_trainsition_system.py.
Modify the evaluation code accordingly.

Credit

Part of the codes and system design are modified from TranX.

torchasn's People

Contributors

Stargazers

Watchers

torchasn's Issues

subprocess.CalledProcessError: Command '['java', '-cp', './external/datagen.jar:./external/lib/*', '-ea', 'datagen.Main', 'equiv', 'endwith(or(<vow>,<cap>))', 'endwith(or(<cap>,<vow>))']' returned non-zero exit status 1.

When running ./scripts/turk/test.sh checkpoints/turk/pretrained.bin The following error is outputted:

Decoding: 100%|##########| 2500/2500 [06:07<00:00, 6.79it/s]
Traceback (most recent call last):
File "test.py", line 82, in
test(args)
File "test.py", line 72, in test
eq_res = check_equiv(pred_code, gt_code)
File "test.py", line 30, in check_equiv
out = subprocess.check_output(
File "C:\Users\AviVarma\AppData\Local\Programs\Python\Python38\lib\subprocess.py", line 415, in check_output
return run(popenargs, stdout=PIPE, timeout=timeout, check=True,
File "C:\Users\AviVarma\AppData\Local\Programs\Python\Python38\lib\subprocess.py", line 516, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['java', '-cp', './external/datagen.jar:./external/lib/', '-ea', 'datagen.Main', 'equiv', 'endwith(or(,))', 'endwith(or(,))']' returned non-zero exit status 1.

Question: Trying to replace LuongAttention with dot_prod_attention.

I am trying to understand the scaled dot product implementation provided in nn.utils.py on line 16. why are you adding a batch_size dimension and what is hidden_size? can you describe the function's input parameters and outputs in more detail please.

Recommend Projects

xiye17 / torchasn Goto Github PK

torchasn's Introduction

torchASN

Prerequisites

Run on Demo Dataset

Addapting to New Programming Language

Credit

torchasn's People

Contributors

Stargazers

Watchers

Forkers

torchasn's Issues

subprocess.CalledProcessError: Command '['java', '-cp', './external/datagen.jar:./external/lib/*', '-ea', 'datagen.Main', 'equiv', 'endwith(or(<vow>,<cap>))', 'endwith(or(<cap>,<vow>))']' returned non-zero exit status 1.

Question: Trying to replace LuongAttention with dot_prod_attention.

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent