Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/bvsam/ml-translation
Sequence to sequence models for language translation.
https://github.com/bvsam/ml-translation
deep-learning machine-learning python pytorch rnn translation
Last synced: 4 days ago
JSON representation
Sequence to sequence models for language translation.
- Host: GitHub
- URL: https://github.com/bvsam/ml-translation
- Owner: bvsam
- License: mit
- Created: 2023-12-04T13:30:19.000Z (12 months ago)
- Default Branch: main
- Last Pushed: 2023-12-05T01:15:23.000Z (12 months ago)
- Last Synced: 2023-12-05T14:59:57.961Z (12 months ago)
- Topics: deep-learning, machine-learning, python, pytorch, rnn, translation
- Language: Jupyter Notebook
- Homepage:
- Size: 25.4 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# ML Translation
## About
- Sequence to sequence models (seq2seq) for language translation with Pytorch
- bidirectional encoder-decoder LSTM
- notebook at `src/en-fr_rnn.ipynb`
- English to French translation using the Anki dataset
- Model deployment on AWS Sagemaker for inference using [bentoml](https://www.bentoml.com/)
## Deployment
Steps to deploy the trained model on AWS Sagemaker for inference. The final result will include the model being accessible for inference at a specific API endpoint on AWS.
### Prerequisites:
- Install Terraform
- Install Docker
### Steps
1. Install `requirements.txt`
```
pip install -r requirements.txt
```
2. Run the `src/en-fr_rnn.ipynb` notebook to train the model and save its weights at `src/weights/en-fr_rnn_lstm_512.pt`. Modify `src/create_model.py` with any changes to the final model's architecture or the paths to its weights, if necessary.
Note: pretrained weights are currently not included in this repo due to file size limits imposed by Github
3. Create the bentoml model
```
cd src
py create_model.py
```
3. Build the bento
```
bentoml build
```
4. Install the aws-sagemaker bentoctl operator
```
bentoctl operator install aws-sagemaker
```
5. Initialize a bentoctl deployment. You may want to delete files including `src/deployment_config.yaml` and `src/main.tf` before doing so (so that they can be overwritten without any conflict).
Note: If other operators are installed, running `bentoctl init` may not work as expected, specifically for Windows hosts. Remove all existing bentoctl operators by deleting the bentoctl folder at `~/bentoctl`. Then install the operator you'd like to work with.
```
bentoctl init
```
6. Run `bentoctl build` with the deployment config and the built bento. You can view existing bentos with the command `bentoml list`.
```
bentoctl build -f deployment_config.yaml -b
```
7. Initialize Terraform and apply the terraform config/plan. Use `bentoctl.tfvars`, which should have been created when running `bentoctl init`, for the var. file.
```
terraform init
terraform apply --var-file=bentoctl.tfvars
```
### Teardown
Once done, destroy all resources created (including the AWS ecr repository) with:
```
bentoctl destroy
```
## TODO (Improvements)
- Implement attention with the RNNs, or even a transformer
- Try using a larger dataset, such as WMT 2014
- Test and implement different methods to reduce text degeneration
## References
- Pytorch tutorial by [Sean Robertson](https://github.com/spro) introducing seq2seq networks: https://pytorch.org/tutorials/intermediate/seq2seq_translation_tutorial.html
- Youtube video by [@mildlyoverfitted](https://www.youtube.com/@mildlyoverfitted) explaining the bentoml deployment process: https://youtu.be/Zci_D4az9FU