first_wave_models.md

The first-wave models

Model	Use Case
SqueezeNet	Image Classification
MobileNet	Image Classification
ResNet	Image Classification
TinyYOLO	Object Detection
RNNoise	Noise Suppression
NSNet	Noise Suppression

The ops required by the first-wave models

Op	SqueezeNetV1.1 ONNX TFLite	MobileNetV2 ONNX TFLite	ResNetV2 ONNX TFLite	TinyYOLOV2 ONNX TFLite	RNNoise C	NSNet ONNX	ONNX	XLA-HLO
add		✔️	✔️	✔️	✔️	✔️	Add	Add
averagePool2d	✔️						AveragePool	Lowering to ReduceWindow, Add and Div.
batchNormalization (1)		✔️	✔️	✔️			BatchNormalization	BatchNormInference
clamp		✔️ (6)				✔️	Clip	Clamp
concat	✔️				✔️		Concat	Concatenate
conv2d	✔️	✔️ (2)	✔️	✔️			Conv	ConvGeneralDilated
gemm			✔️				Gemm	Lowering to Broadcast, Transpose, Mul, Dot and Add
globalAveragePool (a variant of averagePool2d)		✔️	✔️				GlobalAveragePool	Lowering to ReduceWindow, Add and Div.
gru					✔️ (5)	✔️	GRU	Lowering to Add, Mul, and Tanh.
leakyRelu (3)				✔️			LeakyRelu	Lowering to Mul, Gt and Select.
matmul					✔️	✔️	MatMul	Dot
maxPool2d	✔️		✔️	✔️			MaxPool	Lowering to ReduceWindow and Max.
mul				✔️			Mul	Mul
relu	✔️	✔️	✔️		✔️		Relu	Lowering to Max
reshape	✔️	✔️	✔️				Reshape	Reshape
sigmoid					✔️	✔️	Sigmoid	Lowering to Tanh
softmax (4)	✔️	✔️	✔️				Softmax	Lowering to Add, Div, Exp, Max, Sub and Reduce.
split					✔️		Split	Lowering to Slice
squeeze						✔️	Squeeze	Lowering to Reshape
tanh					✔️		Tanh	Tanh

Remarks:

1. TFLite models don't contain batchNormalization. It is fused with other operations during the model conversion.
2. A depthwise conv2d operation is a variant of grouped conv2d, used in MobileNet, where the groups = input_channels = output_channels.
3. TFLite models don't contain leakyRelu. It is replaced with other operations during the model conversion.
4. ONNX models don't contain softmax. It is implemented in the post-processing script.
5. RNNoise implements gru as a combination of sigmoid, tanh, relu, add, split, matmul and mul.
6. TFLite model uses Relu6 which can be lowered to clamp.