ppyolo_mbv3_large_qat.yml

pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams
slim: QAT

QAT:
  quant_config: {
    'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
    'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.99,
    'quantizable_layer_type': ['Conv2D', 'Linear']}
  print_model: True

PPYOLOFPN:
  in_channels: [160, 368]
  coord_conv: true
  conv_block_num: 0
  spp: true
  drop_block: false