Help with implementing YoloV5 and SparseML about sparseml

[markurtz]

@pumetu for the second question, we should be able to inject this SparseML sample code right after the creation of the optimizer here to enable recipes: https://github.com/ZQPei/deep_sort_pytorch/blob/master/deep_sort/deep/train.py#L68

from sparseml.pytorch.optim import ScheduledModifierManager

manager = ScheduledModifierManager.from_yaml(RECIPE_PATH)
optimizer = manager.modify(model, optimizer, steps_per_epoch=len(train_loader))

This will enable using the recipes for the training flows.

[pumetu]

I'm also trying to quantize a library called DeepSort, would you have any suggestions or resources I could look at to quantize it?
https://github.com/ZQPei/deep_sort_pytorch#training-the-re-id-model

[markurtz]

Hi @pumetu, definitely. Here are some short configs for YOLOv5L that haven't been heavily optimized for performance but should get you started in the right direction. Note, for any of the other v5 variants you'll need to adjust the params listed out in the quantization modules to add or subtract appropriately.

Quantization only, run with the fine tuning hyp:

version: 1.1.0
modifiers:
  - !EpochRangeModifier
    start_epoch: 0.0
    end_epoch: 3

  - !SetLearningRateModifier
    start_epoch: 0
    learning_rate: 0.0001

  - !SetLearningRateModifier
    start_epoch: 1
    learning_rate: 0.0032

  - !QuantizationModifier
    start_epoch: 1
    submodules:
      - model.0
      - model.1
      - model.2
      - model.3
      - model.4
      - model.5
      - model.6
      - model.7
      - model.8
      - model.9
      - model.10
      - model.11
      - model.12
      - model.13
      - model.14
      - model.15
      - model.16
      - model.17
      - model.18
      - model.19
      - model.20
      - model.21
      - model.22
      - model.23

Pruning only, run with the from scratch hyp:

version: 1.1.0
modifiers:
  - !EpochRangeModifier
    start_epoch: 0.0
    end_epoch: 240.0
  - !GMPruningModifier
    params: __ALL_PRUNABLE__
    init_sparsity: 0.25
    final_sparsity: 0.90
    update_frequency: 2.0
    start_epoch: 5.0
    end_epoch: 65.0