Continual Neural Dirichlet Process Mixture

May 17, 2020 ยท View on GitHub

Official PyTorch implementation of ICLR 2020 paper: A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning.

Paper

CN-DPM

Experimental Results

Summarization of the main experiments

Method Split-MNIST
Acc. (%)		 Split-MNIST (Gen.)
bits/dim		 MNIST-SVHN
Acc.(%)		 Split-CIFAR10
Acc.(%)		 Split-CIFAR100
Acc.(%)
iid-offline 98.63 0.1806 96.69 93.17 73.80
iid-online 96.18 0.2156 95.24 62.79 20.46
Fine-tune 19.43 0.2817 83.35 18.08 2.43
Reservoir 85.69 0.2234 94.12 44.00 10.01
CN-DPM 93.23 0.2110 94.46 45.21 20.10

Training Graphs

Split-CIFAR10 (0.2 Epoch)

Split-CIFAR10 (0.2 Epoch)

Split-CIFAR100

Split-CIFAR100

System Requirements

  • Python >= 3.6.1
  • CUDA >= 9.0 supported GPU with at least 10GB memory

Installation

  1. Install PyTorch 1.0.1 and TorchVision 0.2.2 for your environment. Follow the instructions in HERE.

  2. Install other required packages.

$ pip install -r requirements.txt

Usage

$ python main.py --help
usage: main.py [-h] [--config CONFIG] [--episode EPISODE] [--log-dir LOG_DIR] [--override OVERRIDE]

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG, -c CONFIG
  --episode EPISODE, -e EPISODE
  --log-dir LOG_DIR, -l LOG_DIR
  --override OVERRIDE

Composing Continual Learning Episodes

We provide a quick and easy solution to compose continual learning scenarios. You can configure a scenario by writing a YAML file. Here is an example of Split-CIFAR10 where each stage is repeated for ten epochs:

- subsets: [['cifar10', 0], ['cifar10', 1]]
  epochs: 10
- subsets: [['cifar10', 2], ['cifar10', 3]]
  epochs: 10
- subsets: [['cifar10', 4], ['cifar10', 5]]
  epochs: 10
- subsets: [['cifar10', 6], ['cifar10', 7]]
  epochs: 10
- subsets: [['cifar10', 8], ['cifar10', 9]]
  epochs: 10

Basic rules:

  • Each scenario consists of a list of stages.
  • Each stage defines a list of subsets.
  • A subset is a two-element list [dataset_name, subset_name]. By default, each class is defined as a subset with the class number as its name.
  • Each stage may optionally define one of epochs, steps, and samples to set the length of the stage. Otherwise, the default length is set to 1 epoch.

The main logic is implemented in the DataScheduler in data.py.

Reproducing Experiments

Run the commands below to reproduce our experimental results. You can check the summaries on TensorBoard.

1. MNIST Generation

iid Offline

$ python main.py \
    --config configs/mnist_gen-iid_offline.yaml \
    --episode episodes/mnist-iid-100epochs.yaml \
    --log-dir log/mnist_gen-iid_offline

iid Online

$ python main.py \
    --config configs/mnist_gen-iid_online.yaml \
    --episode episodes/mnist-iid-online.yaml \
    --log-dir log/mnist_gen-iid_online

Finetune

$ python main.py \
    --config configs/mnist_gen-iid_online.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist_gen-finetune

Reservoir

$ python main.py \
    --config configs/mnist_gen-reservoir.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist_gen-reservoir

CN-DPM

$ python main.py \
    --config configs/mnist_gen-cndpm.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist_gen-cndpm

2. MNIST Classification

iid Offline

$ python main.py \
    --config configs/mnist-iid_offline.yaml \
    --episode episodes/mnist-iid-100epochs.yaml \
    --log-dir log/mnist-iid_offline

iid Online

$ python main.py \
    --config configs/mnist-iid_online.yaml \
    --episode episodes/mnist-iid-online.yaml \
    --log-dir log/mnist-iid_online

Finetune

$ python main.py \
    --config configs/mnist-iid_online.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist-finetune

Reservoir

$ python main.py \
    --config configs/mnist-reservoir.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist-reservoir

CN-DPM

$ python main.py \
    --config configs/mnist-cndpm.yaml \
    --episode episodes/mnist-split-online.yaml \
    --log-dir log/mnist-cndpm

3. MNIST-SVHN Classification

iid Offline

$ python main.py \
    --config configs/mnist_svhn-iid_offline.yaml \
    --episode episodes/mnist_svhn-iid-10epochs.yaml \
    --log-dir log/mnist_svhn-iid_offline

iid Online

$ python main.py \
    --config configs/mnist_svhn-iid_online.yaml \
    --episode episodes/mnist_svhn-iid-online.yaml \
    --log-dir log/mnist_svhn-iid_online

Finetune

$ python main.py \
    --config configs/mnist_svhn-iid_online.yaml \
    --episode episodes/mnist_svhn-online.yaml \
    --log-dir log/mnist_svhn-finetune

Reservoir

$ python main.py \
    --config configs/mnist_svhn-reservoir.yaml \
    --episode episodes/mnist_svhn-online.yaml \
    --log-dir log/mnist_svhn-reservoir

CN-DPM

$ python main.py \
    --config configs/mnist_svhn-cndpm.yaml \
    --episode episodes/mnist_svhn-online.yaml \
    --log-dir log/mnist_svhn-cndpm

4. CIFAR10 Classification

iid Offline

$ python main.py \
    --config configs/cifar10-iid_offline.yaml \
    --episode episodes/cifar10-iid-100epochs.yaml \
    --log-dir log/cifar10-iid_offline

iid Online

$ python main.py \
    --config configs/cifar10-iid_online.yaml \
    --episode episodes/cifar10-iid-online.yaml \
    --log-dir log/cifar10-iid_online

Finetune

$ python main.py \
    --config configs/cifar10-iid_online.yaml \
    --episode episodes/cifar10-split-online.yaml \
    --log-dir log/cifar10-finetune

Reservoir

$ python main.py \
    --config configs/cifar10-reservoir.yaml \
    --episode episodes/cifar10-split-online.yaml \
    --log-dir log/cifar10-reservoir

CN-DPM

$ python main.py \
    --config configs/cifar10-cndpm.yaml \
    --episode episodes/cifar10-split-online.yaml \
    --log-dir log/cifar10-cndpm

CN-DPM (0.2 Epoch)

$ python main.py \
    --config configs/cifar10-cndpm.yaml \
    --episode episodes/cifar10-split-0.2epoch.yaml \
    --log-dir log/cifar10-cndpm-0.2epoch

CN-DPM (10 Epochs)

$ python main.py \
    --config configs/cifar10-cndpm.yaml \
    --episode episodes/cifar10-split-10epochs.yaml \
    --log-dir log/cifar10-cndpm-10epoch

5. CIFAR100 Classification

iid Offline

$ python main.py \
    --config configs/cifar100-iid_offline.yaml \
    --episode episodes/cifar100-iid-100epochs.yaml \
    --log-dir log/cifar100-iid_offline

iid Online

$ python main.py \
    --config configs/cifar100-iid_online.yaml \
    --episode episodes/cifar100-iid-online.yaml \
    --log-dir log/cifar100-iid_online

Finetune

$ python main.py \
    --config configs/cifar100-iid_online.yaml \
    --episode episodes/cifar100-split-online.yaml \
    --log-dir log/cifar100-finetune

Reservoir

$ python main.py \
    --config configs/reservoir-resnet_classifier-cifar100.yaml \
    --episode episodes/cifar100-split-online.yaml \
    --log-dir log/cifar100-reservoir

CN-DPM

$ python main.py \
    --config configs/cifar100-cndpm.yaml \
    --episode episodes/cifar100-split-online.yaml \
    --log-dir log/cifar100-cndpm