Skip to main content

PyTorch Pneumonia Detection Training and Inference Pipeline

This tutorial is designed to demonstrate a typical machine learning development workflow using proxiML GPUs. It walks through how to create a proxiML Dataset, build an initial model using a proxiML Notebook, run parallel hyperparameter tuning experiments using proxiML Training Jobs, save the results of a marathon training job to a reusable proxiML Model, use that model to run a proxiML Inference Job on a batch of new images, receive the results on your local computer, and (optionally) deploy the proxiML Model as an Inference Endpoint to run real-time inference. The data used by this tutorial consists of DICOM files of chest radiographs and their associated labels from the Kaggle RSNA Pneumonia Detection Challenge. The model code was largely adapted from Kaggle notebooks by Guilia Savorgnan. The following changes were made to the original code to better facilitate the tutorial:

  • Changed all directory path references to match the location of the data, temp, and output directories in the proxiML job environment.
  • Converted the notebooks into python scripts with an argparse command line interface.
  • Exposed and implemented additional hyperparameter settings.
  • Added rudimentary tensorboard logging.
  • Changed prediction logic to save the predictions as JSON and save annotated images as PNGs.
caution

This code is for illustrative purposes only. It is not meant as an example of a high performing or efficient model. It's only purpose is to show the various ways the proxiML capabilities can be utilized in the model development process. Do NOT use this in production.

Running the full tutorial as documented should cost approximately 10 credits ($10) if your account is still under the 50 GB free storage allotment.

Prerequisites​

Before beginning this tutorial, ensure that you have satisfied the following prerequisites.

Model Development​

Dataset Staging​

The first step in building an initial model is to load the training data as a proxiML Dataset. The training data can be viewed here. In order to create a proxiML dataset from Kaggle competition data, login to the proxiML web interface and navigate to the Datasets Dashboard. Click the Create button from the dashboard. Enter a memorable name in the name field (e.g. RSNA Pneumonia Detection), select Kaggle as the source type, select Competition as the type, and enter rsna-pneumonia-detection-challenge as the path. Click the Create button to begin the dataset creation. Once the dataset changes to the ready status, it can be used in the subsequent steps.

tip

If the Kaggle option is disabled, you have not yet configured your Kaggle API keys in your proxiML account. Follow the instructions here to proceed.

The RSNA Pneumonia Detection dataset has two sets of images stage_2_train_images and stage_2_test_images. Only the "train" images have labels. We will use the "train" images during the training process and the "test" images to demonstrate the inference process. For the inference process, you should also download the Kaggle data to your local computer using the command:

kaggle competitions download -c rsna-pneumonia-detection-challenge

Once the download completes, unzip the file and save the contents of the stage_2_test_images folder to a memorable location (the tutorial uses ~/rsna-pneumonia-detection-challenge/new_images). The rest of the data can be deleted.

Initial Model Creation​

The easiest way to start a new project is with a proxiML Notebook. Navigate to the Notebook Dashboard and click the Create button. Input a memorable name as the job name and select an available GPU Type (the code in this tutorial assumes a RTX 2080 Ti). Expand the Data section and click Add Dataset. Select My Dataset as the dataset type and select the dataset you created in the previous section from the list (e.g. RSNA Pneumonia Detection). Expand the Model section and specify the tutorial code git repository url https://github.com/proxiML/pytorch-pneumonia-detection.git to automatically download the tutorial's model code. Click Next to view a summary of the new notebook and click Create to start the notebook.

Once the notebook reaches the running state, click the Open to access the notebook instance. Inside the Jupyter Lab environment, the file browser pane on the left will show two directories, input and models. The input folder contains the RSNA Pneumonia dataset and the models folder contains the code from the git repository. Double click on the models folder and open the eda-adapted notebook. This notebook contains some exploratory data analysis on the dataset. The original source is located here, it was only modified to direct file path variables to the correct location with the proxiML job environment. It also generates a features file that is required for the model training notebook. Either run this notebook to generate the file or run python data_processing.py from a terminal tab.

Warning

The train.csv file generated from the notebook or script must be present in the models directory in order for the subsequent steps to succeed.

Once the train.csv appears in the file explorer in the models folder, open the pytorch-pneumonia-detection notebook. This notebook contains the model training and evaluation code. You can find the original here. You can either review the stored results or run all cells to observe the training yourself. To shorten the duration, change the Debug variable to True

tip

Take note of how the proxiML environment variables are used to define the different data paths in both the notebooks and as the default arguments in the data_processing.py script. This is the recommended way to define file locations in code when using the proxiML job environment.

Continue to explore the notebook design and job environment as desired. In most real projects, the objective of the notebook stage is to ensure the data is being loaded correctly and the model code is executing correctly before moving on to longer duration training experiments.

Model Training​

Adapting the Notebook for Training​

Notebooks are great for exploratory data analysis and stepping through code once block at a time. However, to run large scale experiments or productionalize a model, it should be converted into callable python scripts. During this process, a key consideration is which variables should be exposes as "inputs" to the script during training and inference. Typically, these variables fall into at least 3 categories; environment settings, features switches, and hyperparameters. Environment settings allow the script to adapt to different execution environments (e.g. your laptop versus Kaggle versus proxiML). Common environment settings are the file path for the input data, how many GPUs are available, and where to save the model checkpoints or outputs. Features switches control if certain logic is executed or not. Common feature switches include enabling tensorboard logging, placing the script in debug mode, or running a particular data processing step. Hyperparameters are whichever model parameters you wish to expose for training experiments. Typical hyper parameters are learning rate, number of epochs, etc. but can also be setup to change the model's optimizer, pre-trained backbone, number of layers, or anything else.

Once you have selected the interesting variables to set at runtime, you must also implement a method for reading those variables during execution. There are numerous methods to accomplish this. Options include reading from a yaml or json file, setting environment variables, parsing command line arguments, or using another library. This tutorial is designed using the python built-in argparse library to specify variable as command line parameters to the training and prediction scripts.

The .py files in this code repository show an example of how the notebook could be converted into python scripts. The dataset, model, and metrics .py files contain many of the core functions related to those parts of the original notebook. The train and predict scripts contain the rest of the logic and serve as the entry points for running the model. The data_processing files is a converted version of the exploratory data analysis notebook that generates a csv file required for the rest of the scripts to work.

In addition to breaking out some functions and classes into their own files, the two main changes are the implementation of a main function and the argparse setup:

if __name__ == "__main__":
    parser = make_parser()
    args = parser.parse_args()

All the required parameters are then defined with their types, defaults, and allowed options in the make_parser function. At runtime, the args object gets passed to the function the script calls and its values can be passed to subsequent methods. This allows the user to run the script like this:

python train.py --debug --optimizer adamax

This will set args.debug to True and args.optimizer to adamax, which can then be referenced in the training loop code.

The scripts in the repository additionally implement basic tensorboard logging, saving annotated images as files, and allowing the user to specify the optimizer at runtime. You should verify that your scripts successful execute and start training using the terminal window in a Notebook instance before moving on to the next step.

Now that the model code is organized with a scriptable interface, running training experiments becomes very simple. The example code exposes the type of optimizer as a hyperparameter, with 5 different options. Rather than run each of them one at a time and having to wait 2+ hours for each, proxiML training jobs allow you to run these experiments in parallel. The easiest way to start a training job is by converting a working notebook into one. From the Notebook Dashboard, select the notebook job created in the previous section. If it is either running or stopped, the Copy button is enabled at the top of the dashboard. Click the button and the copy notebook form will appear. Select Convert to Training Job as the copy type and give the job a memorable name. Leave the current dataset selected and select Local as the output type. Create an empty directory on your computer and specify the path as the Output storage path (e.g. ~/rsna-pneumonia-detection-challenge/output). This will instruct the job workers to send their training results to that directory in on your local computer.

caution

Using the Local storage type requires prerequisites to be satisfied. Ensure your computer is properly configured before using this setting.

Since we have 5 optimizer types to try, set the Number of Workers to 5. Enter the following commands, one for each worker:

python train.py --optimizer adam
python train.py --optimizer adamw
python train.py --optimizer adamax
python train.py --optimizer sgd
python train.py --optimizer adagrad --lr 0.1
tip

If you choose to run this training job from the Create button on the Training Job Dashboard instead of the copying the notebook, you must prefix the above five commands with python data_processing.py && in order to generate the train.csv file the training script requires.

Click the Copy button to start the training. Note that this does not require you to stop working in the notebook to run the experiment, however, any changes to the notebook after the copy process is complete will not affect the new training job.

Once the training job enters the running state, you can monitor the training progress by clicking the View button. Log message from the workers will stream to this view. The job will run for approximately 5 hours on RTX 2080 Ti GPUs. Before the job ends, you must click the Connect button and follow the instructions to allow the workers to upload their results to your computer.

Warning

When using the Local output type, workers will wait indefinitely for you to connect, and you will continue to billed until the upload completes.

Once connected, you will also see the worker logs streaming in your terminal window. Once the all workers complete, you should see 5 new files in the directory you specified earlier. If you extract these files you will find the saved model checkpoints, a folder containing the validation images with annotations, and a logs folder containing the tensorboard logs. Additionally, to download an extract of the workers' output logs, go back to the Training Dashboard and click the name of the training job. On the job details page, you will see a Download Logs button next to each finished worker. Click this to download the logs.

Review the results of all the worker runs to decide which optimizer was most effective. You can also run additional experiment changing other hyperparameters. In the worker log output, the last line prior to the messages pertaining to the data upload shows the Total Average Precision: at the end of the training run. Your results may vary, but in the example run used to build this tutorial, the adamw optimizer had the highest precision of 0.043. As a result, we will select this option when doing the final training run in the next step.

Marathon Training​

Once the best hyperparameters have been selected, the next step is to run a longer training run to get the best model possible. Training jobs are the best job type for this as well. To start this process, navigate to the Training Job Dashboard and click the Create button. Choose a memorable name for the job and select RTX 2080 Ti as the GPU Type. Click Add Datasets in the Input Data section and select the same dataset as used in the previous sections. Since we intend to use the results of this training run for future inference jobs, select proxiML as the Output Type. When this is selected, the ML_OUTPUT_PATH is redirected to the same location as the ML_MODEL_PATH and the entire contents of that directory are saved as a proxiML model.

Specify the project's git repository url https://github.com/proxiML/pytorch-pneumonia-detection.git as the Model Code Location and use the following command for the job worker.

python data_processing.py && python train.py --optimizer adamw --epochs 10 --train-threshold --no-save-images

Note that because this job is not being copied from a notebook that already generated the train.csv file, this worker will have to run the data_processing.py script prior to running the training script. Click Next to review the job and Create to start the training. Since the local data option is not used, it is not necessary to connect to this job. You can monitor the job progress through the real-time log viewer on the website, or connect to the job if you prefer the log output in a terminal window. The job will run for approximately 13 hours on an RTX 2080 Ti GPU.

Once the job is complete, navigate to the Models Dashboard. There you will fine the model that was created from the jobs output with the name Job - <name of the training job>. You can click the model name to view additional details about the model, including a tree view that shows the directory structure, the number files in each directory, and the size of the files in each directory. Now that the model is ready, the final step is to use the trained model to generate annotated images from new data.

Inference Pipeline​

Inference is typically done either in batches, when you have several new inputs to process, or one-by-one in real-time. The choice of the method depends on how the new data is received as well as the urgency of the predicted results. Batch inference is typically more efficient in terms of resource usage, but will provide predictions in minutes or hours. Inference endpoints will provide prediction results in seconds or milliseconds, but may be idle for long periods of time while waiting for new requests.

Since this data is from a Kaggle competition, batch inference is the most appropriate method to generate predictions for the test set. However, we will still demonstrate how an inference endpoint can be deployed to using the same trained model to provide one-by-one prediction results.

Batch Inference​

In order to demonstrate running an inference job on new DICOM files, this tutorial uses the stage_2_test_images images from the same Kaggle RSNA Pneumonia Detection dataset. If you have not already staged these files on your local computer, return to the Dataset Staging and complete the second half of the instructions. In order to proceed, you should have a folder on your local computer (e.g. ~/rsna-pneumonia-detection-challenge/new_images) that contains all the *.dcm files from the stage_2_test_images folder directly inside (e.g. ~/rsna-pneumonia-detection-challenge/new_images/0000a175-0e68-4ca4-b1af-167204a7e0bc.dcm, ~/rsna-pneumonia-detection-challenge/new_images/0005d3cc-3c3f-40b9-93c3-46231c3eb813.dcm, etc.)

To create a new inference job, navigate to the Inference Job Dashboard and click the Create button. Enter a memorable name for a job and select RTX 2080 Ti as the GPU Type. In the data section, select Local as both the Input Type and the Output Type. Specify the location of the new DICOM files in the Input storage path field (e.g. ~/rsna-pneumonia-detection-challenge/new_images) and the location you want the job to place the annotated images in the Output storage path field (e.g. ~/rsna-pneumonia-detection-challenge/output). In the Model section, select proxiML as the Model Type and select the model created in the previous step from the Model list. Enter the following command as the worker command:

python data_processing.py --type prediction --images $ML_DATA_PATH && python predict.py

Click Next to review the job and Create to start inference. Since this job is using the local data option for input data as well as output data, you must connect to the job before it will start. The job will wait indefinitely for you to connect, and you continue to be charged while it is waiting. Once you connect and the download completes, the job will take approximately 30 minutes to finish on an RTX 2080 Ti GPU.

Once it is finished, open the directory you specified as the output storage path and unzip the file. It will contain an images folder with the annotated images based on the new dataset as well as an annotations.json file that contains the predicted boxes for each images in text form.

Inference Endpoint​

To create a new endpoint, navigate to the Endpoint Dashboard and click the Create button. Enter a memorable name for a job and select RTX 2080 Ti as the GPU Type. In the Model section, select proxiML as the Model Type and select the model created in the previous step from the Model list. In the Endpoint section, click Add Route to add a new endpoint route. Keep the HTTP Verb as POST, enter /predict as the Path, endpoint.py as the File Name, and get_prediction as the Function Name. Click the Add Parameter button in the Request Body Template section twice. Set the first parameter's name to pId and the second parameter's name to dicom. Both parameters should be Strings and are not optional. Keep Function Uses Positional Arguments checked.

Click Next to review the job and Create to start inference. Once the job reaches the running state, click the Connect button to view the endpoint URL.

Using the Command LIne​

To get a prediction for the endpoint from the command line, you can use the predict_scan.sh bash script in the repository. Simply specify the endpoint address from above and the path to a file you want to predict as the two arguments. For example:

./predict_scan.sh <endpoint_address> ~/rsna-pneumonia-detection-challenge/new_images/234add61-53f4-4780-9fb3-10bfcf22e84a.dcm

If you have jq installed, the annotated image will automatically be written to a file called annotated_image.png in the current working directory. Open this image to view the results.

Using a Browser​

info

A current version of Node.js must be installed for this section to work.

Open the file front-end/src/config.js with a text editor. Change the api_address value to your endpoint URL and save the file. If you used a different the route path when creating the endpoint, you must also update that here.

Go to the front-end folder of the repository in a terminal window and type npm start. This will open a web browser to http://localhost:3000 and load the example front end. Click the Upload File button and select one of the DICOM from the dataset. Click Get Prediction to send the file to the endpoint. When the response comes back, the annotated image will display with the box coordinates (if any). Click Upload New File and Get Prediction on additional DICOM files as desired.

Once you are done with the endpoint, be sure to return to the Endpoint Dashboard to stop it. Endpoints, like Notebooks, will stay running until you stop them.

Congratulations, you have trained a custom model on a custom dataset and used it to run inference on brand new data for less than $10 on the proxiML platform!