bluematrix007 / cn-dpm

Official code for ICLR 2020 paper "A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning."

Continual Neural Dirichlet Process Mixture

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

Paper

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-CIFAR100

System Requirements

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

Installation

$ pip install -r requirements.txt

Usage

$ python main.py --help
usage: main.py [-h] [--config CONFIG] [--episode EPISODE] [--log-dir LOG_DIR]
            [--resume-ckpt RESUME_CKPT] [--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 10 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 below commands to reproduce our experimental results. You can check summaries from 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