solution for padding on binary (+1, -1) in resnet-18 #1332

aininot260 · 2025-05-10T09:46:07Z

aininot260
May 10, 2025

I'd like to create a resnet18 model with 1w1a. I use the network from https://github.com/Xilinx/brevitas/blob/dev/src/brevitas_examples/bnn_pynq/models/resnet.py
Use the following code for export

from models.resnet import QuantBasicBlock, QuantResNet

model = QuantResNet(
        block_impl=QuantBasicBlock,
        num_blocks=[2, 2, 2, 2],
        num_classes=10,
        weight_bit_width=1,
        act_bit_width=1,
        round_average_pool=False)

model.eval()

input_tensor = torch.randn(1, 3, 32, 32) 

export_brevitas_onnx(model, input_tensor, export_path='resnet18_1w1a.onnx')

an error occured when converting:

Running step: step_qonnx_to_finn [1/19]                                                                                               
Running step: step_resnet18_attach_preproc [2/19]                                                                                     
Running step: step_tidy_up [3/19]                                                                                                     
Running step: step_resnet18_streamline [4/19]                                                                                         
Running step: step_resnet18_lower [5/19]                                                                                              
Running step: step_resnet18_to_hw [6/19]  
AssertionError: FMPadding_Batch DataType must support zero

It seems that FMPadding_Batch must accept an input of datatype contains zero, not +1, -1

When I tried to fix it by adding input quant params input_quant=Int8ActPerTensorFloat for qnn.QuantConv2d, another error occured:

Running step: step_qonnx_to_finn [1/19]                                                                                               
Running step: step_resnet18_attach_preproc [2/19]                                                                                     
Running step: step_tidy_up [3/19]                                                                                                     
Running step: step_resnet18_streamline [4/19]                                                                                         
Running step: step_resnet18_lower [5/19]                                                                                              
Running step: step_resnet18_to_hw [6/19]                                                                                              
Running step: step_create_dataflow_partition [7/19]                                                                                   
Running step: step_specialize_layers [8/19]                                                                                           
Running step: step_target_fps_parallelization [9/19]                                                                                  
Running step: step_apply_folding_config [10/19]                                                                                       
Running step: step_minimize_bit_width [11/19]                                                                                         
Running step: step_generate_estimate_reports [12/19]                                                                                  
Running step: step_hw_codegen [13/19]
Exception: True binary (non-bipolar) inputs not yet supported

It seems that I can only use +1/-1 in binary network, FMPadding and this feature made conflicts.

for convert steps, I use https://github.com/Xilinx/finn-examples/blob/feature/rn-18/build/resnet18/resnet18_custom_steps.py.

how to modify for creating an network that support padding in resnet-18?

I tried to modify the padding mode, such as set it from padding_mode='zeros' to padding_mode='circular', but it will lead to additional F.pad node in graphs, hard to convert for following steps.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

solution for padding on binary (+1, -1) in resnet-18 #1332

Uh oh!

{{title}}

Uh oh!

Replies: 0 comments

Select a reply

Uh oh!

solution for padding on binary (+1, -1) in resnet-18 #1332

Uh oh!

aininot260 May 10, 2025

Replies: 0 comments

aininot260
May 10, 2025