Datamodels

Compute datamodels easily with datamodels!

[Overview] [Examples] [Tutorial] [Citation]
Figure: Datamodels are a versatile tool for understanding datasets through the lens of the model class. Above figure illustrates one application of finding conflicting subpopulations; we show examples from a top principal component of datamodel embeddings on CIFAR-10.

This repository contains the code to compute datamodels, as introduced in our paper:

Datamodels: Predicting Predictions with Training Data
Andrew Ilyas, Sung Min Park, Logan Engstrom, Guillaume Leclerc, Aleksander Madry
Paper: https://arxiv.org/abs/2112.01008

Overview

Use the datamodels library to compute datamodels for any choice of model and dataset:

1. Train a large number of models (any model!) on different subsets of the training set of choice [1];
2. Store their outputs (any output! e.g., margins, confidences, or even model weights); and
3. Compute datamodels from the above data using our fast_l1 library for large-scale sparse linear regressions.

[1] While we focus on varying the training set composition, you can vary any desired parameters within each training run and can also use our code for grid search, for example.

Examples

The easiest way to learn how to use datamodels is by example, but we also provide a general tutorial below.

Simple examples for a toy setting and CIFAR-10 that run out of the box!

• Toy Setting: See examples/minimal for a dummy example that illustrates how the entire pipeline works together. Simply run example.sh.
• CIFAR10: See examples/cifar:
  • To train, run example.sh, which will create necessary data storage, train models, and log their outputs. You must run this script from the root of the datamodels directory for it to work properly. The script also assumes you are running with a machine with access to 8 GPUs (you can modify as appropriate).
  • To compute datamodels, run example_reg.sh after modifying the variable tmp_dir to be the directory output by the training script.

Pre-computed data on CIFAR-10

For pre-computed datamodels on CIFAR-10 (from traing hundreds of thousands of models!), check out here.

Tutorial

Computing datamodels with the datamodels library has two main stages:

1. Model training: Training models and storing arbitrary data in the form of tensors (these could include predictions, logits, training data masks, or even model weights);
2. Regression: Fitting a sparse linear model from any binary covariates (e.g., training masks, as in our paper) to a continuous outcome (e.g., margins, as in our paper).

The first stage takes up the bulk of compute, but is readily parallelizable by nature.

We describe each stage in detail below.

I. Training models and storing data

There are four key steps here:

1. Write a specification file (in JSON) that pre-declares the data you want to save for each model and the total number of models.
2. Create a "store directory" with empty memory-mapped numpy arrays based on the above a spec.
3. Make a Python file that implements a main function: given an index i, the main function should execute the appropriate training task (potentially using parameters specific to that index), and return a dictionary whose keys map to the datatypes declared in the specification file.
4. Run a worker for each model that you want to train, passing the above Python file and a different index i to each one. Each worker will call the specified file with the given index.

Detailed overview

(Follow along in examples/minimal/ and in particular examples/cifar10/example.sh)

1. Making a spec file: The first step is to make a JSON specification file, which pre-declares what data to record during model training. The spec file has two fields, "num_models" (the number of models that will be trained) and "schema", which maps keys to dictionaries containing shape and dtype information:

   {
       "num_models": 100,
       "schema": {
           "masks": {
               "shape": [50000],
               "dtype": "bool_"
           },
           "margins": {
               "shape": [10000],
               "dtype": "float16"
           }
       }
   }
   

   The dtype field is any attribute of numpy (i.e., float or uint8 or bool_, the numpy boolean type). num_models specifies how many models will be trained.

2. Setting up the data store: Use the spec to make a store directory containing contiguous arrays that will store all model outputs:

   python -m datamodels.training.initialize_store \
       --logging.logdir=$LOG_DIR \
       --logging.spec=examples/cifar10/spec.json
   

   The output directory after initialization:

   > ls $LOG_DIR
   _completed.npy
   masks.npy
   margins.npy
   

   (_completed.npy is used internally to keep track of which models are trained.)

   Each array is of size ($N$ x ...), where $N$ is the number of models to train, and ... is the shape of the data that you want to write (could be scalar, could be vector, could be any shape but there will be $N$ of them concatenated together). That is, it will contain a numpy array with shape: (num_models, schema[key]) for each key in the schema field.

3. Writing a training script: Next, make a training script with a main function that just takes an index index and a logging directory logdir. The training script should return a Python dictionary mapping the keys declared in the schema dictionary above to numpy arrays of the correct shape and dtype.

   Note that you probably don't have to use logdir unless you want to write something other than numpy arrays. Here is a basic example of a main.py file:

   def main(*_, index, logdir):
       # Assume 50,000 training examples and datamodel alpha = 0.5
       mask = np.random.choice(a=[False, True], size=(50_000,), p=[0.5, 0.5])
       # Assume a function train_model_on_mask that takes in a binary
       # mask and trains on the corresponding subset of the training set
       model = train_model_on_mask(mask)
       margins = evaluate_model(model)
       return {
           'masks': mask,
           'margins': margins
       }
   

   The worker will automatically log each key, value in this dictionary as a row <value> in logdir/<key>.npy. Note that (as you'll see in the example scripts), you need to manually create and save training masks (datamodels can log it for you but you will need to generate your own masks and then apply them to the training data).

4. Running the workers. The last step is to pass the training script to the workers. Each worker run will write into a single index of each contiguous array with the datamodel results. Train num_models models, each outputting matrices corresponding to the (n, shape)-shape matrices found in the spec file.

   python -m datamodels.training.worker \
       --worker.index=0 \
       --worker.main_import=examples.cifar10.train_cifar \
       --worker.logdir=$LOG_DIR
   

   Note all the arguments not specified in the datamodels/worker.py file (e.g., --trainer.multiple above) will be passed directly to the training script that your main method is located in.

   You can decide how to best run thousands of workers for your compute environment. An example in GNU parallel with 8 workers at once would be:

   parallel -j 8 CUDA_VISIBLE_DEVICES='{=1 $_=$arg[1] % 8 =}' \
                       python -m datamodels.training.worker \
                       --worker.index={} \
                       --worker.main_import=examples.minimal.train_minimal \
                       --worker.logdir=$LOG_DIR ::: {0..999}

   For our projects, we use GNU parallel in conjunction with our SLURM cluster.

II. Running sparse linear regression

After training the desired number of models, you can use datamodels.regression.compute_datamodels to fit sparse linear models from training masks (or any other binary output saved in the last step) to model outputs (any other continuous output saved in the last stage). There are only three steps here:

1. Writing a dataset: First, we convert the data (stored in memory-mapped arrays from the previous stage) to FFCV format (.beton):

   python -m datamodels.regression.write_dataset \
                --cfg.data_dir $LOG_DIR \
                --cfg.out_path OUT_DIR/reg_data.beton \
                --cfg.y_name NAME_OF_YVAR \
                --cfg.x_name NAME_OF_XVAR
   

2. Making a config file: The next step is to make a config file (.yaml) specifying the data as well as hyperparameters for the regression. For example:

   data:
       data_path: 'OUT_DIR/reg_data.beton' # Path to FFCV dataset
       num_train: 90_000 # Number of models to use for training
       num_val: 10_000 # Number of models to use for validation
       seed: 0 # Random seed for picking the validation set
       target_start_ind: 0 # Select a slice of the data to run on
       target_end_ind: -1 # Select a slice of the data to run on (-1 = end)
       # If target_start_ind and target_end_ind are specified, will slice the
       # output variable accordingly
   cfg:
       k: 100 # Number of lambdas along the regularization path
       batch_size: 2000 # Batch size for regression, must divide both num_train and num_val
       lr: 5e-3 # Learning rate
       eps: 1e-6 # Multiplicative factor between the highest and lowest lambda
       out_dir: 'OUT_DIR/reg_results' # Where to save the results
       num_workers: 8 # Number of workers for dataloader
   early_stopping:
       check_every: 3 # How often to check for inner and outer-loop convergence
       eps: 5e-11 # Tolerance for inner-loop convergence (see GLMNet docs)

3. Running the regression: The last step is to just run the regression script:

   python -m datamodels.regression.compute_datamodels -C config_file.yml
   

Citation

If you use this code in your work, please cite using the following BibTeX entry:

@inproceedings{ilyas2022datamodels,
  title = {Datamodels: Predicting Predictions from Training Data},
  author = {Andrew Ilyas and Sung Min Park and Logan Engstrom and Guillaume Leclerc and Aleksander Madry},
  booktitle = {ICML},
  year = {2022}
}