cifar100_ls1_kd.yaml

# Validation set evaluation metrics:
# Top-1 Accuracy: 71.5%
# Top-5 Accuracy: 92.0%
seed: null
environment:
    platform: local
    ngpus: 1
    cuda:
        cudnn_deterministic: false
        cudnn_benchmark: true
data:
    dataset_path: data/cifar100/
    train_batch_size: 128
    test_batch_size: 100
    workers: 16
model:
    architecture: resnet
    loss: cross_entropy
    arch_config:
        moving_average_mode: 'off'
        moving_average_momentum: 0.99
        block: xnor
        layer0:
            n_in_channels: 64
            kernel_size: 3
            stride: 1
            padding: 1
            bias: false
            maxpool:
                type: identity
        layer1:
            x_quant: ls-1
            w_quant: ls-1
            clamp:
                kind: symmetric
                alpha: 2
            double_shortcut: true
        layer2:
            x_quant: ls-1
            w_quant: ls-1
            clamp:
                kind: symmetric
                alpha: 2
            double_shortcut: true
        layer3:
            x_quant: ls-1
            w_quant: ls-1
            clamp:
                kind: symmetric
                alpha: 2
            double_shortcut: true
        layer4:
            x_quant: ls-1
            w_quant: ls-1
            clamp:
                kind: symmetric
                alpha: 2
            double_shortcut: true
        nonlins: ['relu', 'relu']
        num_blocks: [2, 2, 2, 2]
        output_classes: 100
    kd_config:
        teacher_config_path: experiments/cifar100-teacher/config.yaml
        teacher_checkpoint_path: experiments/cifar100-teacher/checkpoints/checkpoint_200.pt
        freeze_teacher: true
        train_mode: true
        criterion_config:
            temperature: 5
            teacher_correction: true
optimization:
    epochs: 350
    optimizer:
        algorithm: adam
        lr: 0.0002
        weight_decay: 0.0
    lr_scheduler:
        scheduler: multi_step_lr
        milestones:
            - 150
            - 250
            - 320
        gamma: 0.1
log:
    level: INFO
    interval: 100
    tensorboard: true
    tensorboard_root: runs/
    root_experiments_dir: experiments/
    save_model_freq: 80