Is there a way to disable the fused Act(Conv+Bias+Z) operation?

[test3211234]

For me, Act(a1 * Conv + Bias) exists but I don't have the + z operation.

[GreatV]

Yes, you can disable the fused Act(Conv+Bias+Z) operation by controlling how the layers are constructed and fused.

In your case, it seems you're using models like PP-HGNetV2 from PaddleClas, where certain blocks like ConvBNAct or DiverseBranchBlock can support various fusions of convolution, bias, activation, and potentially an additional z term (e.g., identity or residual input). However, by default, these fusions are manually configured during model definition and parameter fusion (e.g., via reparameterization in DiverseBranchBlock or affine scaling in LearnableAffineBlock).

To disable the fused Act(Conv+Bias+Z) pattern, consider the following strategies:

1. Use unfused blocks:

   • In ConvBNAct, disable activation by setting use_act=False.
   • Disable LearnableAffineBlock (LAB) by setting use_lab=False.

   Example:

   ConvBNAct(
   in_channels, out_channels,
   kernel_size=3,
   stride=1,
   padding=1,
   use_act=False, # disables activation
   use_lab=False # disables LAB fusion
   )

2. Avoid reparameterization:
   If you're using DiverseBranchBlock, avoid calling re_parameterize() since it fuses multiple branches (e.g., conv, avg, identity) into a single conv with bias. Keeping is_repped=False ensures branches remain separated and easier to inspect/modify individually.

3. Prevent adding z:
   The Z in Act(Conv + Bias + Z) often refers to residual connections or skip paths (e.g. identity branch). To avoid this, use configurations where such additions do not occur. For example, in DiverseBranchBlock, z is present only when multiple paths including input identity are summed. You can control this by selectively editing which paths are used in forward() or adjusting the block to skip or isolate those paths.

If you are using your own custom module or modifying a model like PPHGNetV2, ensure that in forward passes you avoid:

• calling self.nonlinear(z + conv_out + bias)
• summing extra branches into the convolution result that would represent z

Since you mentioned that you only have Act(a1 * Conv + Bias) and not the extra z, you’re likely using a model configuration without explicit residual or fused affine/identity branches. That's expected if you're not using, for instance, DiverseBranchBlock with all branches enabled.

If needed, you can insert custom code to further isolate the Act(Conv + Bias) by replacing composite layers with atomic layers (Conv2D, BatchNorm2D, ReLU), and controlling exactly what gets fused.

Let me know your specific model definition if you'd like exact code replacements.

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.