PTQ calibration shows bad results.

[taestaes]

I followed the https://github.com/alibaba/TinyNeuralNetwork/blob/main/examples/quantization/post_error_anaylsis.py

calibrated with 10 iterations on test dataloader.

But my network PTQ outputs really bad results as below.

why the following layers have really low cosine similarity?

float_functional_simple_8                          cosine: 0.4407, scale: 0.0016, zero_point: 133
output                                             cosine: 0.6237, scale: 0.0004, zero_point: 154

and found the result image have values of multiple of 0.004 (maybe this is because 8 bit?)

WARNING (tinynn.util.quantization_analysis_util) Quantization error report:

Activations (cosine sorted ):
fake_quant_0                                       cosine: 1.0000, scale: 0.0043, zero_point: 5
float_functional_simple_0                          cosine: 1.0000, scale: 0.0043, zero_point: 5
patch_embed_proj                                   cosine: 0.9980, scale: 0.0133, zero_point: 125
body_0_ffn_project_in                              cosine: 0.9858, scale: 0.0177, zero_point: 126
body_0_ffn_dwconv                                  cosine: 0.9823, scale: 0.0127, zero_point: 124
body_0_ffn_act.quant                               cosine: 1.0000, scale: 0.0005, zero_point: 255
body_0_ffn_act.f_mul_alpha                         cosine: 0.9720, scale: 0.0008, zero_point: 0
body_0_ffn_act.f_add                               cosine: 0.9801, scale: 0.0051, zero_point: 0
body_0_ffn_project_out                             cosine: 0.9130, scale: 0.0134, zero_point: 128
float_functional_simple_2                          cosine: 0.9400, scale: 0.0142, zero_point: 139
body_1_ffn_project_in                              cosine: 0.9636, scale: 0.0285, zero_point: 123
body_1_ffn_dwconv                                  cosine: 0.9774, scale: 0.0499, zero_point: 188
body_1_ffn_act.quant                               cosine: 1.0000, scale: 0.0000, zero_point: 255
body_1_ffn_act.f_mul_alpha                         cosine: 0.9806, scale: 0.0001, zero_point: 0
body_1_ffn_act.f_add                               cosine: 0.8992, scale: 0.0133, zero_point: 0
body_1_ffn_project_out                             cosine: 0.9002, scale: 0.0131, zero_point: 127
float_functional_simple_4                          cosine: 0.8736, scale: 0.0064, zero_point: 135
body_2_ffn_project_in                              cosine: 0.9458, scale: 0.0125, zero_point: 130
body_2_ffn_dwconv                                  cosine: 0.9597, scale: 0.0170, zero_point: 177
body_2_ffn_act.quant                               cosine: 1.0000, scale: 0.0001, zero_point: 0
body_2_ffn_act.f_mul_alpha                         cosine: 0.9837, scale: 0.0003, zero_point: 238
body_2_ffn_act.f_add                               cosine: 0.9069, scale: 0.0067, zero_point: 49
body_2_ffn_project_out                             cosine: 0.9250, scale: 0.0070, zero_point: 122
float_functional_simple_6                          cosine: 0.9139, scale: 0.0061, zero_point: 122
body_3_ffn_project_in                              cosine: 0.9626, scale: 0.0085, zero_point: 126
body_3_ffn_dwconv                                  cosine: 0.9880, scale: 0.0102, zero_point: 147
body_3_ffn_act.quant                               cosine: 1.0000, scale: 0.0001, zero_point: 255
body_3_ffn_act.f_mul_alpha                         cosine: 0.9933, scale: 0.0001, zero_point: 0
body_3_ffn_act.f_add                               cosine: 0.9638, scale: 0.0048, zero_point: 0
body_3_ffn_project_out                             cosine: 0.9703, scale: 0.0051, zero_point: 148
float_functional_simple_8                          cosine: 0.4407, scale: 0.0016, zero_point: 133
output                                             cosine: 0.6237, scale: 0.0004, zero_point: 154
float_functional_simple_9                          cosine: 0.9894, scale: 0.0040, zero_point: 2

WARNING (tinynn.util.quantization_analysis_util) Quantization error report:

Weights (cosine sorted 20):
output                                   cosine: 0.9996, scale: 0.0029, zero_point: 0
patch_embed_proj                         cosine: 0.9998, scale: 0.0056, zero_point: 0
body_3_ffn_dwconv                        cosine: 0.9999, scale: 0.0113, zero_point: 0
body_1_ffn_dwconv                        cosine: 0.9999, scale: 0.0114, zero_point: 0
body_0_ffn_dwconv                        cosine: 0.9999, scale: 0.0132, zero_point: 0
body_2_ffn_dwconv                        cosine: 0.9999, scale: 0.0110, zero_point: 0
body_0_ffn_project_in                    cosine: 0.9999, scale: 0.0118, zero_point: 0
body_1_ffn_project_in                    cosine: 0.9999, scale: 0.0104, zero_point: 0
body_2_ffn_project_in                    cosine: 0.9999, scale: 0.0094, zero_point: 0
body_3_ffn_project_out                   cosine: 0.9999, scale: 0.0052, zero_point: 0
body_0_ffn_project_out                   cosine: 0.9999, scale: 0.0051, zero_point: 0
body_1_ffn_project_out                   cosine: 0.9999, scale: 0.0058, zero_point: 0
body_3_ffn_project_in                    cosine: 0.9999, scale: 0.0085, zero_point: 0
body_2_ffn_project_out                   cosine: 0.9999, scale: 0.0048, zero_point: 0

Activations (cosine sorted 20):
body_1_ffn_act.f_add                               cosine: 0.9432, scale: 0.0133, zero_point: 0
body_0_ffn_project_in                              cosine: 0.9546, scale: 0.0177, zero_point: 126
body_0_ffn_act.f_mul_alpha                         cosine: 0.9757, scale: 0.0008, zero_point: 0
body_0_ffn_dwconv                                  cosine: 0.9781, scale: 0.0127, zero_point: 124
body_1_ffn_project_in                              cosine: 0.9788, scale: 0.0285, zero_point: 123
body_0_ffn_act.f_add                               cosine: 0.9803, scale: 0.0051, zero_point: 0
body_1_ffn_dwconv                                  cosine: 0.9900, scale: 0.0499, zero_point: 188
patch_embed_proj                                   cosine: 0.9907, scale: 0.0133, zero_point: 125
body_1_ffn_project_out                             cosine: 0.9928, scale: 0.0131, zero_point: 127
float_functional_simple_2                          cosine: 0.9939, scale: 0.0142, zero_point: 139
body_0_ffn_project_out                             cosine: 0.9941, scale: 0.0134, zero_point: 128
body_1_ffn_act.f_mul_alpha                         cosine: 0.9949, scale: 0.0001, zero_point: 0
body_2_ffn_project_out                             cosine: 0.9991, scale: 0.0070, zero_point: 122
float_functional_simple_8                          cosine: 0.9993, scale: 0.0016, zero_point: 133
float_functional_simple_4                          cosine: 0.9993, scale: 0.0064, zero_point: 135
body_2_ffn_act.f_add                               cosine: 0.9994, scale: 0.0067, zero_point: 49
body_3_ffn_project_out                             cosine: 0.9995, scale: 0.0051, zero_point: 148
body_3_ffn_act.f_add                               cosine: 0.9995, scale: 0.0048, zero_point: 0
body_2_ffn_project_in                              cosine: 0.9995, scale: 0.0125, zero_point: 130
float_functional_simple_6                          cosine: 0.9996, scale: 0.0061, zero_point: 122

QMLBNR_Video_Images_CleanedImgOut(
  (fake_quant_0): QuantStub(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0043], device='cuda:0'), zero_point=tensor([5], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-0.060858067125082016, max_val=1.065263271331787)
    )
  )
  (patch_embed_proj): Conv2d(
    34, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0056], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.7156392335891724, max_val=0.6727595925331116)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0133], device='cuda:0'), zero_point=tensor([125], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.7740585803985596, max_val=2.808042287826538)
    )
  )
  (body_0_ffn_project_in): Conv2d(
    32, 64, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0118], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.50102961063385, max_val=1.4929399490356445)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0177], device='cuda:0'), zero_point=tensor([126], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-8.30683708190918, max_val=9.019135475158691)
    )
  )
  (body_0_ffn_dwconv): Conv2d(
    64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0132], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.6797765493392944, max_val=1.2519268989562988)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0127], device='cuda:0'), zero_point=tensor([124], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-8.81155776977539, max_val=8.465405464172363)
    )
  )
  (body_0_ffn_act): QPReLU(
    (relu1): ReLU()
    (relu2): ReLU()
    (f_mul_neg_one1): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_neg_one2): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_alpha): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0008], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=0.0, max_val=1.1990134716033936)
      )
    )
    (f_add): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0051], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=1.9476141185914564e-11, max_val=8.480807304382324)
      )
    )
    (quant): QuantStub(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0005], device='cuda:0'), zero_point=tensor([255], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=-0.1360728144645691, max_val=-0.1360728144645691)
      )
    )
  )
  (body_0_ffn_project_out): Conv2d(
    64, 32, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0051], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.656369149684906, max_val=0.5680031180381775)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0134], device='cuda:0'), zero_point=tensor([128], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-4.129014492034912, max_val=2.9563019275665283)
    )
  )
  (body_1_ffn_project_in): Conv2d(
    32, 64, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0104], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.1867985725402832, max_val=1.3303239345550537)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0285], device='cuda:0'), zero_point=tensor([123], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-10.604576110839844, max_val=11.774846076965332)
    )
  )
  (body_1_ffn_dwconv): Conv2d(
    64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0114], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.218837857246399, max_val=1.4485722780227661)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0499], device='cuda:0'), zero_point=tensor([188], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-26.407068252563477, max_val=9.284073829650879)
    )
  )
  (body_1_ffn_act): QPReLU(
    (relu1): ReLU()
    (relu2): ReLU()
    (f_mul_neg_one1): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_neg_one2): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_alpha): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([6.1041e-05], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=0.0, max_val=0.04620003327727318)
      )
    )
    (f_add): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0133], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=8.125244184073455e-13, max_val=9.228955268859863)
      )
    )
    (quant): QuantStub(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([6.8477e-06], device='cuda:0'), zero_point=tensor([255], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=-0.0017461760435253382, max_val=-0.0017461760435253382)
      )
    )
  )
  (body_1_ffn_project_out): Conv2d(
    64, 32, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0058], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.6410632133483887, max_val=0.7345181703567505)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0131], device='cuda:0'), zero_point=tensor([127], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.9626619815826416, max_val=3.1261534690856934)
    )
  )
  (body_2_ffn_project_in): Conv2d(
    32, 64, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0094], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.9506298303604126, max_val=1.1930553913116455)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0125], device='cuda:0'), zero_point=tensor([130], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-4.432260036468506, max_val=4.595865249633789)
    )
  )
  (body_2_ffn_dwconv): Conv2d(
    64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0110], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.1248000860214233, max_val=1.408692479133606)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0170], device='cuda:0'), zero_point=tensor([177], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-12.326212882995605, max_val=4.741044044494629)
    )
  )
  (body_2_ffn_act): QPReLU(
    (relu1): ReLU()
    (relu2): ReLU()
    (f_mul_neg_one1): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_neg_one2): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_alpha): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0003], device='cuda:0'), zero_point=tensor([238], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=-0.32712748646736145, max_val=0.005777135491371155)
      )
    )
    (f_add): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0067], device='cuda:0'), zero_point=tensor([49], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=-0.32712748646736145, max_val=4.730205535888672)
      )
    )
    (quant): QuantStub(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0001], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=0.02653917297720909, max_val=0.02653917297720909)
      )
    )
  )
  (body_2_ffn_project_out): Conv2d(
    64, 32, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0048], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.6111413240432739, max_val=0.6124895811080933)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0070], device='cuda:0'), zero_point=tensor([122], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.243239164352417, max_val=2.098046064376831)
    )
  )
  (body_3_ffn_project_in): Conv2d(
    32, 64, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0085], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.9145596027374268, max_val=1.082226276397705)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0085], device='cuda:0'), zero_point=tensor([126], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-4.47300910949707, max_val=3.173271656036377)
    )
  )
  (body_3_ffn_dwconv): Conv2d(
    64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=64
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0113], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-1.4376935958862305, max_val=1.413638710975647)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0102], device='cuda:0'), zero_point=tensor([147], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-4.438266754150391, max_val=4.107792377471924)
    )
  )
  (body_3_ffn_act): QPReLU(
    (relu1): ReLU()
    (relu2): ReLU()
    (f_mul_neg_one1): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_neg_one2): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
      )
    )
    (f_mul_alpha): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0001], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=0.0, max_val=0.12268836051225662)
      )
    )
    (f_add): FloatFunctional(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0048], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=2.958098876959525e-09, max_val=4.043459415435791)
      )
    )
    (quant): QuantStub(
      (activation_post_process): PTQFakeQuantize(
        fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0001], device='cuda:0'), zero_point=tensor([255], device='cuda:0', dtype=torch.int32)
        (activation_post_process): HistogramObserver(min_val=-0.02744886465370655, max_val=-0.02744886465370655)
      )
    )
  )
  (body_3_ffn_project_out): Conv2d(
    64, 32, kernel_size=(1, 1), stride=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0052], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.553494393825531, max_val=0.6569843292236328)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0051], device='cuda:0'), zero_point=tensor([148], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.0227982997894287, max_val=2.459625244140625)
    )
  )
  (output): Conv2d(
    32, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)
    (weight_fake_quant): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=-128, quant_max=127, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric, ch_axis=-1, scale=tensor([0.0029], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): MinMaxObserver(min_val=-0.23968760669231415, max_val=0.3721259534358978)
    )
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0004], device='cuda:0'), zero_point=tensor([154], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-0.38361242413520813, max_val=0.10394330322742462)
    )
  )
  (fake_dequant_0): DeQuantStub()
  (float_functional_simple_0): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0043], device='cuda:0'), zero_point=tensor([5], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-0.060858067125082016, max_val=1.065263271331787)
    )
  )
  (float_functional_simple_1): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
    )
  )
  (float_functional_simple_2): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0142], device='cuda:0'), zero_point=tensor([139], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-5.174468040466309, max_val=2.951368808746338)
    )
  )
  (float_functional_simple_3): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
    )
  )
  (float_functional_simple_4): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0064], device='cuda:0'), zero_point=tensor([135], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.5720856189727783, max_val=2.035804271697998)
    )
  )
  (float_functional_simple_5): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
    )
  )
  (float_functional_simple_6): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0061], device='cuda:0'), zero_point=tensor([122], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-2.3314638137817383, max_val=1.8853096961975098)
    )
  )
  (float_functional_simple_7): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([1.], device='cuda:0'), zero_point=tensor([0], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=inf, max_val=-inf)
    )
  )
  (float_functional_simple_8): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0016], device='cuda:0'), zero_point=tensor([133], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-0.6245776414871216, max_val=0.5727238655090332)
    )
  )
  (float_functional_simple_9): FloatFunctional(
    (activation_post_process): PTQFakeQuantize(
      fake_quant_enabled=tensor([1], device='cuda:0', dtype=torch.uint8), observer_enabled=tensor([0], device='cuda:0', dtype=torch.uint8), quant_min=0, quant_max=255, dtype=torch.quint8, qscheme=torch.per_tensor_affine, ch_axis=-1, scale=tensor([0.0040], device='cuda:0'), zero_point=tensor([2], device='cuda:0', dtype=torch.int32)
      (activation_post_process): HistogramObserver(min_val=-0.06914715468883514, max_val=1.1036951541900635)
    )
  )
)

[zk1998]

您好，我已经收到您的邮件，我会尽快查看并在第一时间回复您。

[zk1998]

Hi @taestaes, you can add a pseudo-quantization switch behind this example code to analyze where the quantization error mainly comes.
The usage looks like:

    # Disable observer and enable fake quantization to validate model with quantization error
    ptq_model.apply(torch.quantization.disable_observer)
    ptq_model.apply(torch.quantization.enable_fake_quant)

    if torch.cuda.device_count() > 1:
        ptq_model = ptq_model.module
    # Disable the activation or weight quantization to check whether the quantization error comes from activation or weight.
    from tinynn.graph.quantization.fake_quantize import PTQFakeQuantize
    for name, module in ptq_model.named_modules():
        if isinstance(module, PTQFakeQuantize):
            # disable all activation weight quantization
            # if name.endswith('weight_fake_quant'):
            #     module.apply(torch.quantization.disable_fake_quant)
            # disable all activation quantization
            if name.endswith('activation_post_process') :
                module.apply(torch.quantization.disable_fake_quant)

    # Or, you could profile to identify which operator is causing the quantization to fail.
    # First, disable all quantization, then gradually enable quantization for certain layers/Op, and pinpoint the layer with the highest quantization loss.
    # The following is just an example, you should comment or modify it before profiling.
    # ptq_model.apply(torch.quantization.disable_fake_quant)
    # for name, child in ptq_model.named_modules():
    #     # if name in ['body_0_ffn_act']:
    #     if 'ffn_act' in name:
    #         child.apply(torch.quantization.enable_fake_quant)

    dummy_input.to(device)
    ptq_model(dummy_input)
    print(ptq_model)

All quantized value ranges look normal. I think it is the loss of accuracy caused by Prelu's activation value quantization. You can firstly try to only quantize Prelu and check the quantization error.

[taestaes]

    # Get the first batch from the context.train_loader
    first_batch = next(iter(context.val_loader))

    # Set the history frame.
    model._set_history_frame(first_batch, 0)

    # Update the info from the cfg.
    model._update_cfg(first_batch['cfg'])

    # Send the input data to the buffer.
    model.feed_data(first_batch)

    dummy_input_real = model.lq[:1]

    ptq_model.apply(torch.quantization.disable_observer)
    ptq_model.apply(torch.quantization.enable_fake_quant)
    output = ptq_model(dummy_input_real)
    torch.save(output[:,:,:500,:500].clone(), 'output_ptq.pt')

    ptq_model.apply(torch.quantization.disable_fake_quant)
    ptq_model.apply(torch.quantization.enable_observer)
    output_real = ptq_model(dummy_input_real)
    torch.save(output_real[:,:,:500,:500].clone(), 'output_real.pt')

The result files are output_ptq.zip
You can see that the values are changing as:

output
tensor([[[[ 0.0040,  0.0000, -0.0080,  ...,  0.0040,  0.0000,  0.0040],
          [-0.0080,  0.0040,  0.0040,  ..., -0.0040,  0.0040,  0.0040],
          [ 0.0040,  0.0040,  0.0000,  ...,  0.0040,  0.0000,  0.0000],
          ...,
          [-0.0080, -0.0040,  0.0040,  ...,  0.0040,  0.0080,  0.0000],
          [ 0.0080,  0.0080,  0.0040,  ...,  0.0119,  0.0000,  0.0040],
          [-0.0040,  0.0040, -0.0040,  ...,  0.0040,  0.0040, -0.0040]],
         [[-0.0080, -0.0080,  0.0000,  ...,  0.0000,  0.0000, -0.0040],
          [-0.0040, -0.0080,  0.0040,  ..., -0.0040,  0.0040,  0.0040],
          [ 0.0040, -0.0040,  0.0040,  ..., -0.0040,  0.0199,  0.0199],
          ...,
          [-0.0080,  0.0040, -0.0040,  ..., -0.0080, -0.0080,  0.0040],
          [ 0.0119, -0.0080,  0.0040,  ..., -0.0080,  0.0040,  0.0159],
          [-0.0040, -0.0080,  0.0000,  ...,  0.0040, -0.0080, -0.0080]],
         [[-0.0080, -0.0080, -0.0040,  ...,  0.0040,  0.0000,  0.0000],
          [ 0.0080,  0.0080, -0.0080,  ...,  0.0040,  0.0080, -0.0080],
          [-0.0040,  0.0000,  0.0000,  ...,  0.0000,  0.0119,  0.0040],
          ...,
          [ 0.0080, -0.0080,  0.0040,  ..., -0.0040,  0.0000, -0.0080],
          [-0.0040,  0.0000,  0.0000,  ..., -0.0080,  0.0040,  0.0040],
          [-0.0080,  0.0000,  0.0040,  ..., -0.0040, -0.0040,  0.0040]],
         [[-0.0040, -0.0040, -0.0040,  ...,  0.0040, -0.0040,  0.0040],
          [-0.0040, -0.0080, -0.0080,  ...,  0.0000, -0.0040, -0.0040],
          [-0.0080, -0.0080, -0.0040,  ...,  0.0040,  0.0000,  0.0040],
          ...,
          [ 0.0000,  0.0040,  0.0040,  ...,  0.0000, -0.0040,  0.0040],
          [ 0.0000,  0.0040, -0.0040,  ...,  0.0000,  0.0000,  0.0040],
          [ 0.0000, -0.0080,  0.0000,  ..., -0.0080,  0.0000, -0.0080]]]],
       device='cuda:0', grad_fn=<PixelShuffleBackward0>)

output_real
tensor([[[[ 5.3058e-04,  2.8980e-04, -6.9127e-06,  ...,  1.7132e-03,
            2.6455e-03,  1.7680e-03],
          [-4.6593e-05, -6.8595e-05,  1.1221e-05,  ...,  1.9627e-03,
            2.2377e-03,  1.9130e-03],
          [-3.1052e-04, -3.8708e-04, -7.7196e-05,  ...,  1.2334e-03,
            2.0632e-03,  1.3827e-03],
          ...,
          [-6.5564e-04, -4.1493e-04,  6.6315e-06,  ..., -1.3240e-04,
            1.8151e-04,  7.3130e-05],
          [-2.3227e-04, -6.4236e-04, -3.9087e-04,  ..., -4.1661e-05,
            8.5274e-04,  5.5308e-04],
          [-2.3724e-04, -8.0742e-04, -4.4858e-04,  ..., -9.9482e-05,
            3.8351e-04,  2.2529e-04]],
         [[ 4.0678e-04,  5.6869e-04,  2.8066e-04,  ...,  3.4742e-03,
            3.5885e-03,  3.3987e-03],
          [ 2.0502e-04,  7.0215e-04, -1.2013e-04,  ...,  3.2012e-03,
            3.5111e-03,  3.1280e-03],
          [ 4.3886e-05, -2.1096e-04, -2.7194e-05,  ...,  2.8814e-03,
            3.1086e-03,  3.0355e-03],
          ...,
          [ 1.5980e-04,  1.1029e-04,  1.8060e-04,  ..., -3.0157e-05,
           -6.5934e-05, -1.6346e-04],
          [-2.2554e-04,  1.7096e-04, -1.5797e-04,  ...,  6.5495e-05,
           -3.4225e-04, -1.5740e-04],
          [-2.4068e-04,  2.9705e-04,  5.6365e-05,  ..., -2.3401e-04,
           -1.4961e-04, -6.9319e-04]],
         [[ 5.5293e-04,  4.7936e-04,  4.6354e-04,  ...,  3.4175e-03,
            3.9984e-03,  3.8437e-03],
          [ 5.9630e-04,  3.6567e-04,  3.6413e-04,  ...,  3.5018e-03,
            3.4483e-03,  3.5330e-03],
          [ 1.8462e-05, -8.4457e-05,  1.7958e-04,  ...,  3.0389e-03,
            3.4476e-03,  2.9877e-03],
          ...,
          [ 5.8573e-06,  1.9718e-04,  1.9586e-04,  ...,  3.8674e-05,
           -1.9411e-04,  1.3282e-04],
          [-1.6424e-04, -3.5088e-04,  1.9186e-04,  ..., -1.2239e-05,
            1.2320e-04, -9.2608e-05],
          [-9.5273e-05,  4.7107e-05, -9.0065e-05,  ...,  1.3077e-04,
           -8.2172e-04,  1.7154e-04]],
         [[ 1.1913e-04,  5.1337e-05,  9.3932e-05,  ...,  1.2537e-03,
            1.6869e-03,  1.1538e-03],
          [ 2.3396e-05, -4.5896e-04, -1.7053e-05,  ...,  1.3006e-03,
            1.2846e-03,  1.0711e-03],
          [-1.0827e-04, -1.2818e-04,  4.2531e-05,  ...,  1.2620e-03,
            1.6253e-03,  1.2912e-03],
          ...,
          [ 2.0148e-04, -2.4638e-04,  4.6862e-05,  ..., -2.4427e-05,
           -1.3010e-04, -3.9609e-06],
          [-1.6371e-04, -2.9308e-04,  3.4205e-05,  ..., -1.4886e-05,
           -1.6856e-04, -3.4492e-04],
          [ 6.9447e-05, -2.2164e-04,  4.1239e-04,  ..., -3.4211e-05,
            2.2587e-04, -1.5262e-04]]]], device='cuda:0',
       grad_fn=<PixelShuffleBackward0>)

output.mean()
tensor(0.0041, device='cuda:0', grad_fn=<MeanBackward0>)
output_real.mean()
tensor(0.0036, device='cuda:0', grad_fn=<MeanBackward0>)
torch.mean((output - output_real) ** 2)
tensor(3.5770e-05, device='cuda:0', grad_fn=<MeanBackward0>)

Isn't it severe degradation? I don't know whether the PTQ worked well or not. How can I solve this issue?

[taestaes]

When I only apply quantize for 'ffn_act' in name:

ptq_model.apply(torch.quantization.disable_fake_quant)
for name, child in ptq_model.named_modules():
    # if name in ['body_0_ffn_act']:
    if 'ffn_act' in name:
        # if 'ffn_act' not in name:
        # if True:
        child.apply(torch.quantization.enable_fake_quant)
output = ptq_model(dummy_input_real)
print(output)
print(torch.mean((output - output_real) ** 2))
tensor([[[[ 2.0368e-03,  1.2448e-03,  7.7955e-04,  ...,  9.3544e-04,
            4.0698e-03,  3.5494e-03],
          [ 2.9961e-03,  3.0125e-04, -2.0483e-04,  ...,  3.1312e-03,
            2.2275e-03,  3.7057e-03],
          [-3.7076e-04,  5.1835e-04, -3.0740e-03,  ...,  1.3497e-03,
           -1.0703e-04, -3.2131e-04],
          ...,
          [ 1.3461e-03, -2.3721e-04,  2.3214e-03,  ..., -1.2453e-03,
            1.6108e-03,  4.3571e-05],
          [ 1.7997e-03,  1.3568e-04,  1.8453e-04,  ...,  9.4718e-04,
            4.1021e-03,  2.7804e-03],
          [ 1.6654e-03, -3.8672e-04, -1.2939e-04,  ..., -8.9052e-04,
            3.0044e-03,  1.0265e-03]],
         [[ 9.2924e-04, -2.5432e-04,  4.8849e-04,  ...,  2.8125e-03,
            4.1685e-03,  4.0453e-03],
          [-5.0195e-04,  1.5800e-03, -6.3357e-04,  ...,  3.6965e-03,
            1.8738e-03,  4.5212e-03],
          [ 4.2118e-04,  4.2148e-04,  8.4925e-04,  ...,  1.5781e-03,
            2.5055e-03,  2.4197e-03],
          ...,
          [-1.0149e-03, -5.6088e-04, -1.0164e-03,  ...,  4.2484e-04,
            8.1021e-04, -1.1831e-03],
          [-6.1880e-04, -1.8684e-03, -1.1293e-04,  ...,  9.5189e-04,
           -1.8500e-03, -3.0048e-03],
          [-1.5383e-03, -4.5454e-04, -1.2841e-04,  ..., -7.9780e-05,
           -3.5130e-03, -5.7345e-04]],
         [[ 1.7464e-03, -5.9591e-04,  1.9699e-03,  ...,  2.0564e-03,
            3.8366e-03,  4.4443e-03],
          [ 1.3778e-03, -2.1180e-03,  4.7921e-05,  ...,  3.7347e-03,
            3.6731e-03,  2.1038e-03],
          [ 1.4942e-03,  7.0955e-04,  1.7510e-03,  ...,  4.1659e-03,
            5.1922e-03,  1.6140e-03],
          ...,
          [-1.2614e-03, -4.5123e-03, -3.2840e-03,  ...,  2.0923e-03,
            9.8494e-05,  1.2904e-03],
          [ 2.5750e-04, -1.1238e-03,  3.8382e-04,  ..., -1.5174e-03,
            1.3054e-03, -1.4098e-03],
          [ 1.1181e-03, -3.0475e-04,  1.1111e-03,  ..., -4.7602e-04,
           -2.2319e-03, -1.6639e-03]],
         [[-2.1641e-04,  7.5085e-04, -8.1187e-04,  ...,  7.1958e-04,
            1.5348e-03,  1.6091e-03],
          [ 1.1027e-04,  4.4316e-04,  6.8363e-04,  ...,  3.2131e-03,
            1.0030e-03,  5.6958e-05],
          [-1.2995e-03,  3.3173e-04,  1.1150e-03,  ...,  2.8133e-03,
            1.6314e-05, -8.4240e-05],
          ...,
          [ 9.0627e-04, -7.3723e-04, -1.5959e-03,  ...,  7.7767e-04,
            2.6287e-04, -1.0412e-03],
          [-8.7875e-04, -7.2003e-05, -8.8811e-04,  ..., -5.6203e-04,
           -5.9063e-04, -7.2781e-04],
          [ 9.3326e-04, -7.9842e-06,  9.7631e-04,  ..., -9.0815e-04,
            6.0385e-04,  9.3084e-04]]]], device='cuda:0',
       grad_fn=<PixelShuffleBackward0>)
tensor(5.0668e-05, device='cuda:0', grad_fn=<MeanBackward0>)

When I only apply quantize for 'ffn_act' not in name:

ptq_model.apply(torch.quantization.disable_fake_quant)
for name, child in ptq_model.named_modules():
    # if name in ['body_0_ffn_act']:
    # if 'ffn_act' in name:
    if 'ffn_act' not in name:
        # if True:
        child.apply(torch.quantization.enable_fake_quant)
output = ptq_model(dummy_input_real)
print(output)
print(torch.mean((output - output_real) ** 2))
tensor([[[[-0.0085,  0.0042,  0.0042,  ...,  0.0000,  0.0000, -0.0042],
          [-0.0085, -0.0085,  0.0000,  ...,  0.0000,  0.0000, -0.0042],
          [-0.0042,  0.0000,  0.0042,  ...,  0.0000,  0.0127,  0.0127],
          ...,
          [-0.0085, -0.0169,  0.0000,  ...,  0.0042, -0.0042, -0.0085],
          [ 0.0000,  0.0000, -0.0042,  ...,  0.0127,  0.0000,  0.0127],
          [ 0.0042,  0.0000, -0.0042,  ..., -0.0085, -0.0042, -0.0085]],
         [[ 0.0000, -0.0085,  0.0127,  ...,  0.0085,  0.0000,  0.0000],
          [ 0.0000,  0.0000,  0.0042,  ...,  0.0085, -0.0042,  0.0000],
          [ 0.0042,  0.0085,  0.0085,  ...,  0.0042,  0.0042,  0.0085],
          ...,
          [-0.0042,  0.0042, -0.0085,  ..., -0.0085, -0.0085, -0.0085],
          [ 0.0042,  0.0000,  0.0169,  ...,  0.0042,  0.0169,  0.0000],
          [ 0.0042,  0.0042, -0.0042,  ...,  0.0127, -0.0127, -0.0042]],
         [[-0.0042,  0.0085,  0.0000,  ..., -0.0042, -0.0085,  0.0042],
          [-0.0042, -0.0042,  0.0085,  ..., -0.0042,  0.0085,  0.0042],
          [ 0.0000,  0.0042,  0.0000,  ..., -0.0085,  0.0085,  0.0085],
          ...,
          [ 0.0085,  0.0085,  0.0042,  ..., -0.0042, -0.0085,  0.0000],
          [-0.0042,  0.0042,  0.0000,  ...,  0.0085,  0.0042,  0.0000],
          [ 0.0000, -0.0042,  0.0000,  ...,  0.0000,  0.0042,  0.0042]],
         [[ 0.0042, -0.0042, -0.0042,  ..., -0.0042,  0.0042,  0.0042],
          [ 0.0042, -0.0042, -0.0042,  ...,  0.0042,  0.0000,  0.0042],
          [ 0.0042,  0.0000,  0.0085,  ...,  0.0127, -0.0042, -0.0085],
          ...,
          [ 0.0042, -0.0085, -0.0042,  ...,  0.0000,  0.0000,  0.0042],
          [ 0.0042,  0.0000,  0.0042,  ...,  0.0000,  0.0042,  0.0042],
          [ 0.0000, -0.0042,  0.0000,  ...,  0.0085,  0.0042,  0.0000]]]],
       device='cuda:0', grad_fn=<PixelShuffleBackward0>)
tensor(4.7243e-05, device='cuda:0', grad_fn=<MeanBackward0>)

It seems two errors are similar but the result values are quantized with 0.0042 for 'ffn_act' not in name.
How can I make the result better...?

[zk1998]

The result looks bad, whether it is only quantizing prelu or not quantizing prelu, you can determine the quantization error by checking the result of graph_error_analysis to see whether the consine of the corresponding output is close to 1.

BTW, if allowed, you could provide the xx_q.py, xx_q.pth generated in the out directory and provide several example input tensors, I can help you check the quantization Source of error.

[taestaes]

@zk1998 What is your email? I can send it. The dummy input files have large filesize.

[zk1998]

@taestaes zhouye.zk@alibaba-inc.com

[taestaes]

zhouye.zk@alibaba-inc.com

I have sent the input files and the output folder. Can you find them? Thank you for your help, and I really need your assistance. I really want to resolve this issue.

[zk1998]

I have received it and will help you analyze it ASAP.

[zk1998]

Hi @taestaes ，could you please provide a simple post-processing function that outputs pixel values? I found that your model output contains a large number of values ​​close to 0, which makes it difficult for me to compare numerical accuracy. The comparison of pixel values ​​will be more meaningful.

[taestaes]

Hi @taestaes ，could you please provide a simple post-processing function that outputs pixel values? I found that your model output contains a large number of values ​​close to 0, which makes it difficult for me to compare numerical accuracy. The comparison of pixel values ​​will be more meaningful.

What do you mean by post-processing function? my model's input are image values and outputs are also image values.

[taestaes]

Also I want to say that without calibration, the output values are all 0 or 1.
After calibration with the real validation images (sent to you), the output is shown as above. So I think if you find all values are 0 or 1, it will have values after calibration.

[zk1998]

The input and output value range of your model is float[0,1], but the pixel value range of the image is int[0,255].It seems that your model should be image input and image output, so your dataloader should have preprocess and postprocess to process the image.

[zk1998]

@taestaes Maybe your work relies on an open source project, this will make it easier for me to reproduce your problem. I currently find that it is difficult to evaluate the quality of your model by only observing numerical accuracy. It needs to be combined with actual tasks.