NMR_Upscale_UW_DIRECT

Check out the project description!

---

Read it at: project_plans/Airforce_VV.pdf

Setup your python environment

---

Using isotope.yml, create a copy of the python environment with required packages to run all code

• If using Anaconda or Miniconda, you can run: conda env create -f isotope.yml

Generate some simulated NMR data

---

1. Conda activate your environment (look at the section above)
2. Run the preprocessing/nmrgenerate.py script with the following custom command line flags:

• --low_resolution [default=60]: the lower resolution spectrum frequency
• --high_resolution [default=400]: the higher resolution spectrum frequency
• --num_spec [default=100]: The number of spectra samples to generate
• --data_dir [default=./data]: the directory to save the data

3. Look at nmrgenerate.sh for an example of how to run the script

Train some models!

---

1. Conda activate your environment (look at the section above)
2. Run the train.py script with the following custom command line flags:

• --model_name [default=mlp]: what model to use. You can look at available models in the models.py file.
• --num_epochs [default=30]: how many epochs to train the model
• --high_resolution_frequency [default=400]: the higher resolution spectrum frequency of the dataset
• --data_dir [default=./data]: directory that contains the generated NMR data
• --random_key [default=12345]: a random seed integer for reproducibility
• --train_split [default=0.7]: The fraction of the entire dataset to use for training
• --valid_split [default=0.15]: The fraction of the entire dataset to use for validation
• --save_dir [default=./results]: directory to save results

3. Look at train.sh for an example of how to run the script

Interpret the results!

---

1. Once you've run the train.py script, you'll notice that results for the model name will be generated on the save_dir specified. The files saved include:
   • best_model.pt: the best model weights + epoch number + average validation loss + model hyperparameters
   • loss_curves.pt: the loss_curves (training and validation )for the training run
   • predictions.pt: the predictions for each test set spectra
2. You can find an example of how to visualize these results here: noteboks/visualize_results.ipynb

Implement your own models!

---

You can either:

1. Change the model parameters in model_parameters.py for each of the already implemented models (easiest).
2. Implement your own model in models.py and put your class initialization parameters into the name2params dictionary in model_parameters.py. Then you should add the model class spec + name into the name2model dictionary in models.py.

The base implementation includes unittests

---

We included unittests to the base model development implementation. You can run them with the following commands:

• pytest test/test_datasets.py Should return something like:

====================================== test session starts =======================================
platform darwin -- Python 3.6.12, pytest-6.2.4, py-1.11.0, pluggy-0.13.1
rootdir: /Users/davinan/Dropbox/classes/dscience/NMR_Upscale_UW_DIRECT
collected 1 item

test/test_datasets.py .                                                                    [100%]

======================================= 1 passed in 3.98s ========================================

• pytest test/test_train.py Should return something like:

====================================== test session starts =======================================
platform darwin -- Python 3.6.12, pytest-6.2.4, py-1.11.0, pluggy-0.13.1
rootdir: /Users/davinan/Dropbox/classes/dscience/NMR_Upscale_UW_DIRECT
collected 4 items

test/test_train.py ....                                                                    [100%]

======================================= 4 passed in 8.12s ========================================

• pytest test/test_models.py Should return something like:

====================================== test session starts =======================================
platform darwin -- Python 3.6.12, pytest-6.2.4, py-1.11.0, pluggy-0.13.1
rootdir: /Users/davinan/Dropbox/classes/dscience/NMR_Upscale_UW_DIRECT
collected 3 items

test/test_models.py ...                                                                    [100%]

======================================= 3 passed in 0.94s ========================================

Our own explorations:

---

We also included some exploratory approaches + hyperparameter optimization code in the exploratory/ directory. It includes notebooks that generate data and even our initial tests with using Generative Adversarial Networks (GANs) to our problem.