[PTQ Converter] 'Linear+relu' module conversion failed about tinyneuralnetwork

[peterjc123]

        transpose_13_f = fake_dequant_inner_81_0_0.transpose(1, 2)
        backbone_encoders_0_dropout_2 = self.backbone_encoders_0_dropout(transpose_13_f)
        add_4_f = add_3_f.__add__(backbone_encoders_0_dropout_2)
        backbone_encoders_0_norm_ff = self.backbone_encoders_0_norm_ff(add_4_f)
        backbone_encoders_0_feed_forward_w_1 = self.backbone_encoders_0_feed_forward_w_1(backbone_encoders_0_norm_ff)

  (fake_dequant_inner_81_0_0): DeQuantStub()
  (backbone_encoders_0_norm_ff): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
  (backbone_encoders_0_feed_forward_w_1): LinearReLU(
    (0): Linear(in_features=128, out_features=128, bias=True)
    (1): ReLU()
    (activation_post_process): HistogramObserver(min_val=0.0, max_val=1.3926540613174438)
  )

It seems that we should not fuse backbone_encoders_0_feed_forward_w_1 and ReLU here since they are not quantized.

[peterjc123]

Looking into it further, it seems that backbone_encoders_0_feed_forward_activation is reused so that it fails in the is_quantized_analysis pass.

[Juelianqvq]

is_quantized_analysis

So the easiest way is to rename ReLU op with different names?

[peterjc123]

@Juelianqvq It seems that #301 is enough to fix it, though we should probably duplicate those ops in the future.

[Juelianqvq]

@Juelianqvq It seems that #301 is enough to fix it, though we should probably duplicate those ops in the future.

@peterjc123 Thanks for fixing this :)

[Juelianqvq]

FYI, though conversion succeed but the accuracy is still 0, unless I insert dq+q around relu op, which means "qlinear+float relu",the accuracy behaves normal.

[peterjc123]

@Juelianqvq Let me check how is LinearReLU implemented in PyTorch.

[peterjc123]

@Juelianqvq It seems that after model fusing, the original modules involved will be replaced by nn.Identity. That's why there's so few ReLUs in the target TFLite model. So it seems we still need to duplicate those activation ops.

[peterjc123]

Let's continue the discussion in #308.