Skip to content

Latest commit

 

History

History
96 lines (85 loc) · 5.87 KB

experiments.md

File metadata and controls

96 lines (85 loc) · 5.87 KB
Raw

Experiments

Sacred

We use Sacred to help manage experiments and for command line interface. Below are the core features we use from Sacred:

  • Observer interface to collect all experiment related information and store it in a database.
    • files added with ex.add_artifact
    • metrics are logged with ex.log_scalar
    • config added with ex.add_config
  • Command line interface to simplify experiment initialization and configuration
    • each experiment requires the following flag to specify the experiment name --name=experiment_name or -n experiment_name. An output directory containing all experiment files will be created with the experiment name.
    • To ignore oberservers use flag -u which is useful for some quick tests or debugging runs
    • Changes to parameters in the config can be changed on the fly via command line keyword with followed by the key value pair parameters, e.g. with "DATASET.FOLD=$COUNT" "EVAL_ONLY=True". Note, changing the parameter CONFIG_FILE which loads an existing configuration file via command line, needs to have the additional parameter "LOAD_ONLY_CFG_FILE=True", i.e.: 
      python -m seg_3d.train_loop -n experiment_name with "CONFIG_FILE=./config.yaml" "LOAD_ONLY_CFG_FILE=True"

Omniboard

We use Omniboard as the tool for visualizing experiments logged with Sacred. From Omniboard documentation:

Omniboard is a web dashboard for the Sacred machine learning experiment management tool. It connects to the MongoDB database used by Sacred and helps in visualizing the experiments and metrics / logs collected in each experiment.

Follow the installation steps described in the quick start guide. The easiest way of setting up both MongoDB and Omniboard, use docker:

  1. Install docker https://docs.docker.com/get-docker/
  2. Configure docker-compose.yml
  3. Run from the repo root directory sudo docker compose up
  4. Open http://localhost:9000/ in the browser. Port number is specified in docker-compose.yml 
    omniboard:
    ports:
    - 127.0.0.1:9000:9000

Saving and loading existing data into MongoDB

A docker volume is used to persist data from the MongoDB. Volumes are stored on disk at docker/volumes.

To load existing data, first choose the right volume in docker/volumes and then add the following parameter in the docker-compose.yml, replacing /path_to_volume/ with the correct volume path:

  mongo:
    volumes:  # by default volumes are mapped on disk to docker/volumes
      - /path_to_volume/:/data/db

Tensorboard

We also use Tensorboard to visualize the inputs and outputs of the model during train time. It is especially useful to visually monitor the evolution of model predictions and to also confirm data augmentations look reasonable (e.g. elastic deformations stability very sensitive to parameter settings). To bring up the Tensorboard dashboard, run the following command, where output-dir is the path to the directory storing the training runs:

tensorboard --logdir ouput-dir

The following code snippet inside seg_3d/train_loop.py logs the image data to Tensorboard:

# check if need to process masks and images to be visualized in tensorboard
for idx, p in enumerate(patients):
    # hardcoded, only visualize 4 patients from train set
    if p in train_dataset.patient_keys[:4]:
        for name, batch in zip(["img_orig", "img_aug", "mask_gt", "mask_pred"],[orig_imgs, sample, labels, preds]):
            tags_imgs = tensorboard_img_formatter(name=p + "/" + name, batch=batch[idx].unsqueeze(0).detach().cpu())

            # add each tag image tuple to tensorboard
            for item in tags_imgs:
                storage.put_image(*item)

Experiment outputs

Each experiment has an associated output directory containing various files generated during training and evaluation.

output/experiment_name/
│   │
│   ├─ 0/  # fold number for KFold cross validation
│       ├─ config.yaml       # the full config for this experiment
│       ├─ model_best.pth    # weights of best model
│       ├─ model_0000199.pth # weights of model saved during a checkpoint at iteration 199
│       ├─ inference.pk      # predictions from best model (along with inputs and labels)
│       ├─ best_metrics.txt  # results from the best model
│       ├─ metrics.json      # logged metrics during each training/evaluation step
│       ├─ log.txt           # all log messages generated during experiment
│       ├─ last_checkpoint   # used for resuming training if experiment ends abruptly
│       ├─ events.out.tfevents..  # Tensorboard file
│       ├─ masks/  # figures for the prediction masks
│               ├─ cor/  # coronal plane
│               │      ├─ PSMA-01-126/  # directory for each patient
│               │      │      ├─ 56.png  # each figure shows a slice with index based on filename, e.g. 56
│               │      │      ├─ ...
│               │      ├─ ...
│               ├─ sag/  # saggital plane
│               │      ├─ ... # same as above
│               └─ tra/  # transverse/axial plane
│                      ├─ ... # same as above
│       └─ eval_0/   # directory with results from doing EVAL_ONLY mode
│               ├─ config.yaml
│               ├─ inference.pk
│               ├─ log.txt
│               └─ masks/
│
...  # additional directories for the other folds