coreml 0.0.9

Generic Framework for ML projects

CoreML

coreml is an end-to-end machine learning framework aimed at supporting rapid prototyping. It is built on top of PyTorch by combining the several components of any ML pipeline, right from definining the dataset object, choosing how to sample each batch, preprocessing your inputs and labels, iterating on different network architectures, applying various weight initializations, running pretrained models, freezing certain layers, changing optimizers, adding learning rate schedulers, and detailed logging, into a simple model.fit() framework, similar to scikit-learn. The codebase is very modular making it easily extensible for different tasks, modalities and training recipes, avoiding duplication wherever possible. The contents of the README are as follows:

• Features
• Setup
• Quickstart
• How-Tos
  • Optimization
  • Network architectures
  • Datasets
  • Preprocessing
• Testing
• TODOs
• Authors

Features

• Support for end-to-end training using PyTorch with custom training and validation loops.
• Makes every aspect of the training pipeline configurable.
• Data preprocessing on GPU using kornia.
• Provides the ability to define and change architectures right in the config file.
• Built-in support for experiment tracking using Weights & Biases.
• Enables replicability through Docker containers.
• Supports tracking instance-level losses over epochs.
• Logs predictions and metrics over epochs to allow future analysis.
• Supports saving checkpoints and optimizing thresholds based on specific subsets.
• Defines several metrics like PrecisionAtRecall, SpecificityAtSensitivity and ConfusionMatrix.
• Logs several classification curves like PR curve, Sensitivity-Specificity curve, ROC curve.
• Explicitly requires data versioning.
• Supports adding new datasets adhering to a required format.
• Contains unit tests wherever applicable.

Setup

Clone the project:

$ git clone https://github.com/dalmia/coreml.git

Weights & Biases

We use wandb for experiment tracking. You'll need to have that set up:

• Install wandb

$ pip install wandb

2. Login to wandb:

$ wandb login

You will be redirected to a link that will show you your WANDB_API_KEY .

3. Set the WANDB_API_KEY by adding this to your ~/.bashrc file:

export WANDB_API_KEY=YOUR_API_KEY

4. Run source ~/.bashrc.
5. During training, you'll have an option to turn off wandb as well.

Docker

We use Docker containers to ensure replicability of experiments. You can either fetch the Docker image from DockerHub using the following line:

$ docker pull adalmia/coreml:v1.0

OR

You can build the image using the DockerFile:

$ docker build -t adalmia/coreml:v1.0 .

The repository runs inside a Docker container. When creating the container, you need to mount the directory containing data to /data and directory where you want to store the ouptuts to /output on the container. Make the corresponding changes to create_container.sh to mount the respective directories by changing /path/to/coreml, /path/to/data and /path/to/outputs to the appropriate values.

Use the following command to launch a container:

$ bash create_container.sh

Quickstart

CIFAR10

• Download and prepare the data

$ python tasks/data/classification/CIFAR.py

This will create a folder named CIFAR10 under the directory mounted at /data on your container along with default data versions and the required directory structure.

• Since the codebase only supports binary classification for now, the step above also creates a binary.yml data version which converts the 10-class problem into a binary classification problem.

• Run training using the default config:

$ python training/train.py --wandb -v configs/defaults/binary-cifar-classification.yml

The flag --wandb is used to ignore using wandb for this run. You can view other flags that can be passed using -h:

root@ip:/workspace/coreml# python training/train.py -h
usage: train.py [-h] -v VERSION [-n NUM_WORKERS] [--debug] [-o] [--resume]
                [--id ID] [--wandb] [-e ENTITY] [-p PROJECT] [--seed SEED]

Trains a model

optional arguments:
  -h, --help            show this help message and exit
  -v VERSION, --version VERSION
                        path to the experiment config file
  -n NUM_WORKERS, --num_workers NUM_WORKERS
                        number of CPU workers to use
  --debug               specify where a debugging run
  -o, --overfit-batch   specify whether the run is to test overfitting
  --resume              whether to resume experiment in wandb
  --id ID               experiment ID in wandb
  --wandb               whether to ignore using wandb
  -e ENTITY, --entity ENTITY
                        wandb user/org name
  -p PROJECT, --project PROJECT
                        wandb project name
  --seed SEED           seed for the experiment

How-Tos

This section demonstrates the power of everything being parameterized by a config. Refer to configs/defaults/binary-cifar-classification.yml as the base config file on top of which we demonstrate the individual components.

Optimization

model:
    name: classification
    batch_size: 8
    epochs: 1
    optimizer:
      name: SGD
      args:
        lr: 0.001
        momentum: 0.9
        nesterov: true
      scheduler:
        update: epoch
        value: loss
        name: ReduceLROnPlateau
        params:
            mode: 'min'
            factor: 0.1
            patience: 5
            verbose: true
            threshold: 0.0001
            threshold_mode: abs
     loss:
      train:
        name: binary-cross-entropy
        params:
            reduction: none
      val:
        name: binary-cross-entropy
        params:
            reduction: none

The above example shows how to set various hyperparameters like the batch size, number of epochs, optimizer, learning rate scheduler and the loss function. The interesting aspect is how the optimizer, learning rate scheduler and the loss function is directly defined in the config file. This is possible because of the Factory Design Pattern used throughout the codebase. For optimizer, we currently support:

• SGD
• Adam
• AdamW

However, other optimization functions can be simply added by registering their corresponding builders in coreml/optimizers.py. For each optimizer, args contains any parameters required by their corresponding PyTorch definition.

Similarly, we support multiple learning rate schedulers defined in PyTorch along parameterizing whether the scheduler's step should take place after each batch or after each epoch. This is controlled by the key update, which can be one of ['epoch', 'batch']. The key value can be set to decide what parameter should be monitored for the scheduler's step. In the above example, the validation loss is being monitored. Currently, we suport the following schedulers:

• ReduceLROnPlateau
• StepLR
• CyclicLR
• OneCycleLR
• MultiStepLR

We also parameterize the loss function to be used and allow for different loss functions for training and validation. The need for making them different could arise in various situations. One such example is applying label smoothing during training but not during validation.

Network architectures

The network architecture can be completely defined in the config itself:

network:
    name: neural_net
    params:
        config:
        - name: Conv2d
          params:
            in_channels: 3
            out_channels: 64
            kernel_size: 3
        - name: BatchNorm2d
          params:
            num_features: 64
        - name: ReLU
          params: {}
        - name: Conv2d
          params:
            in_channels: 64
            out_channels: 64
            kernel_size: 3
        - name: BatchNorm2d
          params:
            num_features: 64
        - name: ReLU
          params: {}
        - name: AdaptiveAvgPool2d
          params:
            output_size:
            - 1
            - 1
        - name: Flatten
          params: {}
        - name: Linear
          params:
            in_features: 64
            out_features: 64
        - name: ReLU
          params: {}
        - name: Linear
          params:
            in_features: 64
            out_features: 1

The config key takes as input a list of dictionaries, with each dictionary specifying a layer or a backbone network. Yes, if you want to use a pretrained ResNet, you can simply plug it in as a backbone layer:

network:
    name: neural_net
    params:
        config:
        - name: resnet50
          params:
            pretrained: true
            in_channels: 3
        - name: AdaptiveAvgPool2d
          params:
            output_size:
            - 1
            - 1
        - name: Flatten
          params: {}
        - name: Linear
          params:
            in_features: 2048
            out_features: 1

Currently, we support a lot of backbones:

• Resnet variations: resnet18, resnet34, resnet50, resnet101, resnet152, resnext50_32x4d, resnext101_32x8d
• VGGNet variations: vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19_bn, vgg19
• EfficientNet variations: efficientnet-b0, efficientnet-b4, efficientnet-b7

The implementations for the Resnet and VGGNet based backbones have been taken from torchvision.models and those based on EfficientNet are supported by this PyTorch implementation.

Support for other backbones can be similarly added to coreml/networks/backbone.

Datasets

The datasets to use can be specified as follows:

dataset:
    name: classification_dataset
    config:
      - name: CIFAR10
        version: binary
    params: {}

Here, name is used to decide which torch.utils.data.Dataset object to use, as defined in coreml/data/__init__.py. config takes a list of dataset configs. Each dataset config is a dictionary with name as the name of the dataset folder under /data and version being the specific dataset version to use, present in /data/{name}/processed/versions. params contains additional arguments that can be passed to the dataset object, as defined in coreml/data. When avlues are passed to params, they are specified for train/val/test separately:

dataset:
    name: classification_dataset
    config:
      - name: CIFAR10
        version
    params:
      train:
        fraction: 0.1
      val:
        fraction: 0.5

Preprocessing

Input transform

One the input is loaded, the pipeline for processing the input before it is fed into a batch, can be specified in the config as:

signal_transform:
    train:
    - name: Permute
      params:
        order:
          - 2
          - 0
          - 1
    val:
    - name: Permute
      params:
        order:
          - 2
          - 0
          - 1

The pipeline for each split (train/val) is specified separately. For each split, a list of dictonaries are given. Each dictionary represents one transform, as defined in coreml/data/transforms.py, using the name of the transform and the arguments for that transform.

Annotation transform

The raw annotations might have to be transformed before processing as well. This is specified in the config as:

target_transform:
    name: classification
    params:
      classes:
        - 0
        - 1

The specific target transform is selected from annotation_factory inside coreml/data/transforms.py

Testing

We use unittest for all our tests. Simply run the following inside the Docker container:

$ python -m unittest discover tests

TODOs

• Tracking inputs on W&B (currently torch.cat kills the process when the inputs are large).
• Add augmentations
• Support for audio classification
• Support for multi-class classification
• Add benchmarks for multiple datasets
• Add documentation for using new datasets and configuring different parts of the pipeline

Authors

• Aman Dalmia
• Piyush Bagad