resnet18_imagenet_post_train_lapq.yaml

# Post-training quantization settings for running in the same conditions as in:
# Nahshan et al., "Loss Aware Post-training Quantization" (https://arxiv.org/abs/1911.07190), according to the
# reference implementation found at: https://github.com/ynahshan/nn-quantization-pytorch/tree/master/lapq
#
# The settings are:
#   * Only fake-quantization is done
#   * Only weights of convolutions and outputs of ReLU are quantized. Pooling, element-wise addition and FC layers
#     are not quantized
#   * The first convolution + relu pair isn't quantized
#   * The last convolution + relu pair ("layer4.1.conv2/relu2") isn't quantized
#
# See example invocations and results after the YAML definition

quantizers:
  post_train_quantizer:
    class: PostTrainLinearQuantizer
    # Don't quantize anything by default, override below for should be quantized
    bits_activations: null
    bits_parameters: null
    bits_accum: 32
    mode:
      activations: ASYMMETRIC_UNSIGNED
      weights: SYMMETRIC
    model_activation_stats: ../quantization/post_train_quant/stats/resnet18_quant_stats.yaml
    per_channel_wts: False
    inputs_quant_auto_fallback: True

    overrides:
      # Conv layers inside the ResNet BasicBlock are quantized (except last one, see below)
      layer.*conv.*:
        bits_activations: null
        bits_weights: 4
      # ReLU layers inside the ResNet BasicBlock are quantized (except last one, see below)
      layer.*relu.*:
        bits_activations: 4
        quantize_inputs: False
      # Conv layers in downsampling residual connections are quantized
      .*downsample\.0.*:
        bits_activations: null
        bits_weights: 4
      # The last conv+relu layers are NOT quantized, we specify them directly
      layer4.1.conv2:
        bits_activations: null
        bits_weights: null
      layer4.1.relu2:
        bits_activations: null

# Example invocations:
#   * Preliminaries:
#       cd <distiller_root>/examples/classifier_compression
#       CONFIG_FILE="../quantization/post_train_quant/resnet18_imagenet_post_train_lapq.yaml"
#       IMAGENET_PATH=<path_to_imagenet>
#
#   * Using L3 initialization:
#     Command:
#       python compress_classifier.py --eval --qe --qe-lapq -a resnet18 --pretrained $IMAGENET_PATH --lapq-eval-size 0.01 --lapq-maxiter 2 --qe-config-file $CONFIG_FILE -b 500 --lapq-init-mode L3 --lapq-init-method powell --lapq-eval-memoize-dataloader --det --lapq-search-clipping
#
#     Excerpts from output:
#       ...
#       Initializing quantizer...
#       Initializing quantization parameters...
#       ...
#       Evaluating initial quantization score...
#       Evaluation set loss after initialization 2.522
#       Test:     loss=2.650, top1=43.816, top5=68.840
#       Using "Powell" minimization algorithm.
#       ...
#       980 evaluations: loss=1.962
#       Iteration 0:   Score=1.956
#       Test:     loss=2.117, top1=51.662, top5=76.606
#       ...
#       2200 evaluations: loss=1.929
#       Iteration 1:   Score=1.926
#       Test:     loss=2.116, top1=51.784, top5=76.712
#       Optimization Done.
#       Arch: resnet18    Test:    top1 = 51.784   top5 = 76.712   loss = 2.116
#
#
#   * Using LAPLACE initialization:
#     Command:
#       python compress_classifier.py --eval --qe --qe-lapq -a resnet18 --pretrained $IMAGENET_PATH --lapq-eval-size 0.01 --lapq-maxiter 2 --qe-config-file $CONFIG_FILE -b 500 --lapq-init-mode LAPLACE --lapq-init-method powell --lapq-eval-memoize-dataloader --det --lapq-search-clipping
#
#     Excerpts from output:
#       ...
#       Initializing quantizer...
#       Initializing quantization parameters...
#       ...
#       Evaluating initial quantization score...
#       Evaluation set loss after initialization 3.376
#       Evaluating on full test set...
#       Test:     loss=3.509, top1=29.492, top5=53.768
#       Using "Powell" minimization algorithm.
#       ...
#       620 evaluations: loss=2.458
#       Iteration 0:   Score=2.458
#       Test:     loss=2.650, top1=42.700, top5=68.138
#       ...
#       1780 evaluations: loss=2.277
#       Iteration 1:   Score=2.274
#       Test:     loss=2.504, top1=45.164, top5=70.400
#       Optimization Done.
#       Arch: resnet18    Test:    top1 = 45.164   top5 = 70.400   loss = 2.504