ML model artifacts for learning

This repository contains model artifacts used by Arm Learning Paths that teach model analysis, graph inspection, deployment-artifact inspection, and target-aware ML workflow concepts.

These artifacts are companion material for learning-path exercises and explanations. Different Learning Paths may use different subsets of the repository depending on whether the focus is MLIA analysis, Model Explorer inspection, ExecuTorch runtime traces, TOSA lowering, TensorFlow Lite models, or Vulkan ML artifacts.

These artifacts are for learning purposes only, in the context of the learning paths that reference them.

Git LFS

This repository uses Git LFS for model artifacts. Install Git LFS before cloning, or run git lfs pull after cloning to download the actual .pt2, .pte, .tflite, .tosa, .vgf, .etdp, and .etrecord files.

Repository layout

ml-model-artifacts/
├── LICENSE.md
├── README.md
├── etdump/
│   ├── mobilenetv2_fp32_ethosu.etdp
│   ├── mobilenetv2_int8_ethosu.etdp
│   ├── mobilenetv2_lrn_int8_ethosu.etdp
│   ├── opt125m_portable.etdp
│   └── opt125m_xnnpack.etdp
├── etrecord/
│   ├── mobilenetv2_fp32_ethosu.etrecord
│   ├── mobilenetv2_int8_ethosu.etrecord
│   ├── mobilenetv2_lrn_int8_ethosu.etrecord
│   ├── opt125m_portable.etrecord
│   └── opt125m_xnnpack.etrecord
├── pte/
│   ├── add_sigmoid_vgf.pte
│   ├── mv2_cortex_m.pte
│   ├── mv2_fp32_ethos_u85.pte
│   ├── mv2_int8_ethos_u85.pte
│   ├── mv2_lrn_int8_ethos_u85.pte
│   ├── opt125m_cortex_a_portable.pte
│   ├── opt125m_cortex_a_xnnpack.pte
│   ├── small_upscaler_ptq_vgf.pte
│   ├── small_upscaler_qat_vgf.pte
│   ├── toy_conditional_select_int8_ethos_u55_256.pte
│   └── toy_conditional_select_int8_ethos_u85_256.pte
├── pt2/
│   ├── mv2_fp32.pt2
│   └── toy_conditional_select_fp32.pt2
├── tflite/
│   ├── mv2_fp32.tflite
│   ├── mv2_int8.tflite
│   └── mv2_lrn_int8.tflite
├── tosa/
│   ├── mv2_fp32.tosa
│   ├── mv2_int8.tosa
│   ├── mv2_lrn_int8_1.tosa
│   ├── mv2_lrn_int8_2.tosa
│   ├── small_upscaler_ptq.tosa
│   └── small_upscaler_qat.tosa
└── vgf/
    ├── add_sigmoid.vgf
    ├── small_upscaler_ptq.vgf
    └── small_upscaler_qat.vgf

Artifact groups

ETDump artifacts

The etdump/ directory contains ExecuTorch debug data files. Open these with the ExecuTorch Model Explorer extension to inspect runtime events and relate execution behavior back to exported program artifacts.

File	Purpose
mobilenetv2_fp32_ethosu.etdp	Debug data for the MobileNetV2 floating-point Ethos-U85 run.
mobilenetv2_int8_ethosu.etdp	Debug data for the MobileNetV2 int8 Ethos-U85 run.
mobilenetv2_lrn_int8_ethosu.etdp	Debug data for the fragmented MobileNetV2 int8 Ethos-U85 run.
opt125m_portable.etdp	Debug data for the OPT-125M portable-kernel run.
opt125m_xnnpack.etdp	Debug data for the OPT-125M XNNPACK-delegated run.

ETRecord artifacts

The etrecord/ directory contains ExecuTorch record files. Open these with the ExecuTorch Model Explorer extension to map runtime trace data to exported programs, delegate regions, and operator-level execution details.

File	Purpose
mobilenetv2_fp32_ethosu.etrecord	Record file for the MobileNetV2 floating-point Ethos-U85 export.
mobilenetv2_int8_ethosu.etrecord	Record file for the MobileNetV2 int8 Ethos-U85 export.
mobilenetv2_lrn_int8_ethosu.etrecord	Record file for the fragmented MobileNetV2 int8 Ethos-U85 export.
opt125m_portable.etrecord	Record file for the OPT-125M portable-kernel export.
opt125m_xnnpack.etrecord	Record file for the OPT-125M XNNPACK-delegated export.

PTE artifacts

The pte/ directory contains ExecuTorch program files. Use these with the PTE adapter (part of the Model Explorer ExecuTorch extension) to inspect deployment graphs, delegate regions, backend partitioning, and work outside accelerator delegates. You can also use these with MLIA.

Non-delegated work is not guaranteed to run on every CPU runtime. ExecuTorch runtime behavior depends on the kernels linked into the target build and on the operators, dtypes, layouts, and shapes those kernels support. Cortex-M bare-metal runtimes often include a narrower kernel set than Cortex-A runtimes, and both can use selective builds that include only the kernels needed by a product.

File	Purpose
add_sigmoid_vgf.pte	Small ExecuTorch program using the Arm VGF backend path.
mv2_cortex_m.pte	MobileNetV2 Cortex-M ExecuTorch artifact.
mv2_fp32_ethos_u85.pte	MobileNetV2 floating-point Ethos-U85 artifact.
mv2_int8_ethos_u85.pte	MobileNetV2 int8 Ethos-U85 artifact.
mv2_lrn_int8_ethos_u85.pte	MobileNetV2 int8 Ethos-U85 artifact with fragmented lowering.
opt125m_cortex_a_portable.pte	OPT-125M Cortex-A artifact using portable kernels.
opt125m_cortex_a_xnnpack.pte	OPT-125M Cortex-A artifact with XNNPACK delegation.
small_upscaler_ptq_vgf.pte	Small upscaler post-training quantized artifact using the Arm VGF backend path.
small_upscaler_qat_vgf.pte	Small upscaler quantization-aware trained artifact using the Arm VGF backend path.
toy_conditional_select_int8_ethos_u55_256.pte	Synthetic int8 ExecuTorch artifact for Ethos-U55-256, used to demonstrate target-dependent delegation and MLIA Corstone analysis.
toy_conditional_select_int8_ethos_u85_256.pte	Synthetic int8 ExecuTorch artifact for Ethos-U85-256, used to demonstrate target-dependent delegation and MLIA Corstone analysis.

PT2 artifacts

The pt2/ directory contains PyTorch exported programs for use with MLIA converter plugins or PyTorch/ExecuTorch-oriented workflows.

File	Purpose
mv2_fp32.pt2	MobileNetV2 floating-point PyTorch exported program.
toy_conditional_select_fp32.pt2	Synthetic floating-point PyTorch exported program for a small convolution plus conditional selection model.

TFLite artifacts

The tflite/ directory contains TensorFlow Lite MobileNetV2 artifacts for use with Model Explorer or MLIA

File	Purpose
mv2_fp32.tflite	MobileNetV2 floating-point TensorFlow Lite model.
mv2_int8.tflite	MobileNetV2 full-integer TensorFlow Lite model with int8 input and output tensors.
mv2_lrn_int8.tflite	MobileNetV2 full-integer TensorFlow Lite model with an inserted local response normalization operation.

TOSA artifacts

The tosa/ directory contains TOSA intermediate representations. Use these with the TOSA adapter to inspect lowered operators, tensor shapes, quantized types, graph splits, and optimization opportunities. You can also use these with MLIA.

File	Purpose
mv2_fp32.tosa	MobileNetV2 floating-point TOSA graph.
mv2_int8.tosa	MobileNetV2 int8 TOSA graph.
mv2_lrn_int8_1.tosa	First TOSA graph partition for the fragmented MobileNetV2 int8 lowering example.
mv2_lrn_int8_2.tosa	Second TOSA graph partition for the fragmented MobileNetV2 int8 lowering example.
small_upscaler_ptq.tosa	Small upscaler TOSA graph produced from post-training quantization.
small_upscaler_qat.tosa	Small upscaler TOSA graph produced from quantization-aware training.

VGF artifacts

The vgf/ directory contains Vulkan Graph Format artifacts. Use these with the VGF adapter to inspect graph connectivity, tensor metadata, constants, and SPIR-V graph modules used by Vulkan ML workflows.

File	Purpose
add_sigmoid.vgf	Small add/sigmoid VGF graph.
small_upscaler_ptq.vgf	Small upscaler VGF artifact produced from post-training quantization.
small_upscaler_qat.vgf	Small upscaler VGF artifact produced from quantization-aware training.

Note

Some artifacts are large, especially the OPT-125M .pte and .etrecord files. They may take longer to download, load, and render than the smaller examples.

License

This repository uses the Arm Education End User License Agreement for teaching and learning content. See LICENSE.md.