EclipseAidge/ds_cnn

DS-CNN

DS-CNN model from MLCommons repository https://github.com/mlcommons/tiny/tree/master/benchmark/training/keyword_spotting

ONNX version exported from .pb model doing

# setup environment
python3.10 -m venv pytf
source pytf/bin/activate

# install latest officially compatible versions
pip install tensorflow==2.15.1 tf2onnx==1.16.1

# use most recent opset officially supported
python -m tf2onnx.convert \
  --saved-model <path/to/dir> \
  --output converted_ds_cnn.onnx --opset 18

This version input format is NHWC. The following Python code fuses MatMul+Add to Gemm and folds the first Reshape operator.

import onnx
import onnxruntime

import aidge_core as ai
import aidge_onnx

model_onnx = onnx.load_model("converted_ds_cnn.onnx")
model_onnx_clean_nhwc = aidge_onnx.onnx_cleaner.clean_onnx(
    model_onnx, {"input_1": [[1, 49, 10, 1]]}, "test_clean", opset_version=18
)
model = aidge_onnx.convert_onnx_to_aidge(model_onnx_clean_nhwc)

to_replace: set[ai.Node] = set(
    [
        model.get_node("StatefulPartitionedCall_functional_1_conv2d_BiasAdd__6"),
        model.get_node("new_shape__103_out0"),
    ]
)

model.replace(to_replace, set())
model.set_mandatory_inputs_first()

model.forward_dims(dims=[[1, 1, 49, 10]], allow_data_dependency=True)
model_onnx_clean_nchw = aidge_onnx.convert_aidge_to_onnx(model, "ds_cnn", opset=18)
onnx.save_model(model_onnx_clean_nchw, "ds_cnn.onnx")

Aidge support

Note: We tested this network for the following features. If you encounter any error please open an issue. Features not tested in CI may not be functional.

Feature	Tested in CI
ONNX import	✔️
Runtime CPU	✔️
Runtime CUDA	✔️
Export CPU	✔️

Model

• operators: 43 (9 types)
  • AvgPooling2D: 1
  • Conv2D: 4
  • FC: 1
  • PaddedConv2D: 1
  • PaddedConvDepthWise2D: 4
  • Producer: 21
  • ReLU: 9
  • Reshape: 1
  • Softmax: 1

Google Speech Commands v2

• Opset: 18
• Source: Google
• Input
  • size: [N, 1, 49, 10]
  • format: [N, C, H, W]
  • preprocessing:
    • ?
• Output
  • size: [N, 12]