Example Models

Overview

The models/ directory contains example QONNX models, including all models from the original finn-examples repository.

These serve as input to our regression test CI infrastructure, which is documented here.

To download these models, run dvc pull within the cloned finn-plus repository.

How to build these examples locally

Build flows for these models are defined and configured using the "benchmarking" system in FINN+. Users can run this locally via the finn bench command:

finn bench --bench_config PATH_TO_CONFIG.YML

This will launch the build(s) defined in the YAML file within the bench_work/ directory. Output artifacts will be saved under a timestamp in the bench_output/ directory, with additional debug artifacts saved in bench_save/.

To see how each build flow is configured, start by looking at the CI regression test configs, for example: regression_basic.yml. The dut entry defines which included flow to execute. These are defined in src/finn/benchmarking/dut either as a simple YAML config or python script for more complex flows. The experiment-specific config passed to finn bench extends/overwrites this model-/topology-specific build config.

Structure and list of available models

The models themselves (.onnx) and golden I/O pairs for verification (.npy) are tracked via DVC, while corresponding build config files (folding_config.json and specialize_layers.json) are tracked directly via Git.

Models are tracked either

• as a single .onnx file (passed as model_path to the benchmarking system)
• as a directory containing a model.onnx file & the inp.npy/out.npy files (passed as model_dir)

Model	Dataset	File names
VGG-10*	RadioML 2018	vgg10/radioml_w4a4_small_tidy.onnx
MNv1*	ImageNet 2012	mobilenetv1/mobilenetv1-w4a4_pre_post_tidy_opset-11.onnx
Cybersecurity (MLP)*	UNSW-NB15	cybsec/unsw_nb15-mlp-w2a2.onnx
BNN-PYNQ tfc* (3 variants)	MNIST	bnn-pynq/cnv-w1a1_qonnx.onnx bnn-pynq/cnv-w1a2_qonnx.onnx bnn-pynq/cnv-w2a2_qonnx.onnx
BNN-PYNQ cnv* (3 variants)	CIFAR-10	bnn-pynq/tfc-w1a1_qonnx.onnx bnn-pynq/tfc-w1a2_qonnx.onnx bnn-pynq/tfc-w2a2_qonnx.onnx
GTSRB (cnv-w1a1)*	GTSRB	gtsrb/cnv_1w1a_gtsrb.onnx
KWS (MLP)*	Google Speech Commands v2	kws/MLP_W3A3_python_speech_features_pre-processing_QONNX_opset-11.onnx
RN-50*	ImageNet 2012	resnet50/resnet50_w1a2_exported.onnx
Transformer GPT (4 variants)	TinyStories	transformer/gpt_a_6b_gpt2-s256-t2048-l2-h4-e256 transformer/gpt_b_4b_gpt2-s256-t2048-l2-h4-e256 transformer/gpt_c_gpt2-s512-t2048-l2-h4-e512 transformer/gpt_d_gpt2-s256-t2048-l1-h2-e256
Transformer RadioML (5 variants)	RadioML 2018	transformer/rml_transformer_0 transformer/rml_transformer_a transformer/rml_transformer_b transformer/rml_transformer_c transformer/rml_transformer_d
Conformer RadioML (1 variant)	RadioML 2018	transformer/rml_conformer
Models imported from finn-transformers repo	N/A	DVC cross-repo import at transformer/finn-transformers

* based on finn-examples