simple adding maxpool lead to AssertionError: cycle-free graph violated: partition depends on itself

[aininot260]

I use the network from https://github.com/Xilinx/brevitas/blob/dev/src/brevitas_examples/bnn_pynq/models/resnet.py
My export-to-onnx code are listed as follows:

from models import resnet18_1w1a
import torch
from brevitas.export import export_onnx_qcdq
from brevitas.export import export_brevitas_onnx
import torch.nn as nn
# from models.resnet_imagenet import QuantBasicBlock, QuantResNet
from models.resnet import QuantBasicBlock, QuantResNet

model = QuantResNet(
        block_impl=QuantBasicBlock,
        num_blocks=[2, 2, 2, 2],
        num_classes=1000,
        weight_bit_width=4,
        act_bit_width=4,
        round_average_pool=False)

model.eval()

input_tensor = torch.randn(1, 3, 32, 32) 
# input_tensor = torch.randn(1, 3, 224, 224) 

# export_onnx_qcdq(model, input_tensor, export_path='quant_model.onnx')
export_brevitas_onnx(model, input_tensor, export_path='resnet18_4w4a.onnx')

then, I use the build step from https://github.com/Xilinx/finn-examples/blob/feature/rn-18/build/resnet18/build.py

for simple, I made these tiny changes:

resnet18_build_steps = [
    "step_qonnx_to_finn",
    step_resnet18_attach_preproc,
    "step_tidy_up",
    step_resnet18_streamline,
    step_resnet18_lower,
    step_resnet18_to_hw,
    "step_create_dataflow_partition",
    "step_specialize_layers",
    "step_target_fps_parallelization",
    "step_apply_folding_config",
    "step_minimize_bit_width",
    "step_generate_estimate_reports",
    "step_hw_codegen",
    "step_hw_ipgen",
    "step_set_fifo_depths",
    # "step_create_stitched_ip",
    # Uncomment if you want RTL simulation reports! 
    # "step_measure_rtlsim_performance",
    step_resnet18_slr_floorplan,
    "step_synthesize_bitfile",
    "step_make_pynq_driver",
    "step_deployment_package",
]

    cfg = DataflowBuildConfig(
        steps= resnet18_build_steps,
        output_dir= local_output_dir,
        synth_clk_period_ns= SYNTH_CLK_PERIOD_NS,
        board= board,
        shell_flow_type= shell_flow_type,
        vitis_platform= vitis_platform,
        # folding_config_file= FOLDING_CONFIG_FILE,
        target_fps= 200,
        # verify_steps= [
            # VerificationStepType.STREAMLINED_PYTHON,
            # VerificationStepType.FOLDED_HLS_CPPSIM,
            # VerificationStepType.STITCHED_IP_RTLSIM
        # ],
        # stitched_ip_gen_dcp = True,
        # verify_input_npy= VERIFICATION_IN_OUT_PAIR[0],
        # verify_expected_output_npy= VERIFICATION_IN_OUT_PAIR[1],
        split_large_fifos= True,
        # save_intermediate_models=True,
        vitis_opt_strategy=build_cfg.VitisOptStrategyCfg.PERFORMANCE_BEST,
        generate_outputs= [
            # DataflowOutputType.STITCHED_IP,
            DataflowOutputType.ESTIMATE_REPORTS,
            # DataflowOutputType.OOC_SYNTH,
            DataflowOutputType.BITFILE,
            DataflowOutputType.PYNQ_DRIVER,
            DataflowOutputType.DEPLOYMENT_PACKAGE
        ],
        # vitis_floorplan_file="folding_config/floorplan.json",
        # start_step="step_create_dataflow_partition",
        # stop_step="step_deployment_package"
    )

all steps works well, bitfile generated successfully

But this model is designed for cifar10, only accept input with 32 x 32

to adapt imagenet 224 x 224, I made some tiny change in https://github.com/Xilinx/brevitas/blob/dev/src/brevitas_examples/bnn_pynq/models/resnet.py, just
set first_maxpool=True

Then, error occured

AssertionError: cycle-free graph violated: partition depends on itself

the adding maxpool leading this error, It only made this change

        # MaxPool is typically present for ImageNet but not for CIFAR10
        if first_maxpool:
            self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=0)
        else:
            self.maxpool = nn.Identity()

the step_resnet18_to_hw.onnx shows quite different at begining
the following are before:


the following are after:



I tried some fix steps, such as insert maxpool related steps to custom_steps but failed
To show clearly, I will use https://github.com/Xilinx/finn-examples/blob/feature/rn-18/build/resnet18/resnet18_custom_steps.py as a new start, wondering adjust it to adapt the added maxpool node. all above pics are generated from this code's steps