imagenet_fp.yaml

# Validation set evaluation metrics:
# Top-1 Accuracy: 69.8%
# Top-5 Accuracy: 89.3%
seed: null
environment:
    platform: local
    ngpus: 8
    cuda:
        cudnn_deterministic: false
        cudnn_benchmark: true
data:
    dataset_path: data/imagenet/
    train_batch_size: 256
    test_batch_size: 256
    workers: 16
model:
    architecture: resnet
    loss: cross_entropy                         # select from {'cross_entropy', 'nll_loss', 'kl_div'}, see get_loss_fn() docs for details.
    arch_config:
        moving_average_mode: 'off'              # select from {'off', 'eval_only', 'train_and_eval'}, see ActivationQuantizer docs for details.
        moving_average_momentum: 0.99
        block: regular                          # select from {'regular', 'xnor'}, see QResNet docs for details.
        layer0:
            n_in_channels: 64
            kernel_size: 7
            stride: 2
            padding: 3
            bias: false
            maxpool:
                type: maxpool2d
                kernel_size: 3
                stride: 2
                padding: 1
        layer1:
            x_quant: fp                         # select from {'fp', 'ls-1', 'ls-T', 'ls-2', 'gf-1', 'gf-2', 'gf-3' (any `gf-k`)}, see QuantConv2d docs for details.
            w_quant: fp                         # select from {'fp', 'ls-1', 'ls-T', 'ls-2', 'gf-1', 'gf-2', 'gf-3' (any `gf-k`)}, see QuantConv2d docs for details.
            clamp:
                kind: identity                  # select from {'identity', 'symmetric'}, see QuantConv2d docs for details.
        layer2:
            x_quant: fp
            w_quant: fp
            clamp:
                kind: identity
        layer3:
            x_quant: fp
            w_quant: fp
            clamp:
                kind: identity
        layer4:
            x_quant: fp
            w_quant: fp
            clamp:
                kind: identity
        nonlins: ['relu', 'relu']               # A list of 2 strings where each string is in {'relu', 'prelu', 'identity'}.
        num_blocks: [2, 2, 2, 2]
        output_classes: 1000
optimization:
    epochs: 100
    optimizer:
        algorithm: sgd                          # select from {'sgd', 'adam', 'adadelta'}, see get_optimizer() docs for details.
        lr: 0.1
        momentum: 0.9
        nesterov: true
        weight_decay: 0.0001
    lr_scheduler:
        scheduler: step_lr                      # select from {'step_lr', 'multi_step_lr', 'linear_lr', 'lambda_lr'}, see get_lr_scheduler() docs for details.
        gamma: 0.1
        step_size: 30
log:
    level: INFO
    interval: 80
    tensorboard: true
    tensorboard_root: runs/
    root_experiments_dir: experiments/
    save_model_freq: 20