ONNX Support in Caten

[hikettei]

• caten/vision (providing utils for cv stuff)
• caten/onnx (from onnx -> caten translator, almost copied from cl-waffe2/frontend)
• Let's start by trying more simple models. MobileNetV2 is a good target.
• Download models from torchvision
• MobileNetV2
• Dynamic Shape Helper
• oops !argmax has a problem? write tests
• Running Classification in ImageNet
• MobileNetV3 or ResNet is doable
• EfficientNet
• TODO: Creating a fuzzer using ONNX
• Write a test for expt
• Running MobileNetv2 for ci and merge this tomorrow
• Or: create a ci named ONNX Test
• GPT2 Test (or GGUF) In Benchmark
• update cl-onnx to not to have cl-easel dependencies

# include them in ci, create a dir scripts/export_mobilenetv2.py
import torchvision
import torch
import torch.onnx

model = torchvision.models.mobilenet_v2(weights='IMAGENET1K_V1')
input_tensor = torch.rand((1, 3, 224, 224), dtype=torch.float32)
torch.onnx.export(model, (input_tensor,), f"./dummy_graph/mobilenetv2.onnx", input_names=["input"])

$ python ./scripts/export_onnx.py --task classification --model mobilenet_v2 --weights IMAGENET1K_V1 --output_dir ./models

Missing Ops

• Clip-13
• Flatten
• Update ConvND (Permute the weight)

TODO

• Documentation for caten/gguf,
• caten/onnx
• caten/llm
• caten/aaps.gpt2

[hikettei]

wip: creating a schedule from onnx