forked from webmachinelearning/webnn
-
Notifications
You must be signed in to change notification settings - Fork 0
WebNN first wave models and ops
Ningxin Hu edited this page Apr 9, 2020
·
2 revisions
| Op | SqueezeNetV1.1 ONNX TFLite | MobileNetV2 ONNX TFLite | ResNetV2 ONNX TFLite | TinyYOLOV2 ONNX TFLite | Remarks |
|---|---|---|---|---|---|
| Add | ✔️ | ✔️ | ✔️ | Need to complete the spec. | |
| AveragePool | ✔️ | ||||
| BatchNormalization | ✔️ | ✔️ | ✔️ | TFLite models either fuse it into preceding conv2d or replace it by mul and add. Can we support it by mul and add? |
|
| Concat | ✔️ | ||||
| Conv | ✔️ | ✔️ | ✔️ | ✔️ | Need to support groups for MobileNetV2's depthwise conv2d. Will we add groups into conv2d or add groupedConv2d or depthwiseConv2d? |
| Gemm | ✔️ | Can it be covered by matmul? |
|||
| GlobalAveragePool | ✔️ | ✔️ | Can it be covered by AveragePool? |
||
| LeakyRelu | ✔️ | TFLite models substitute it by mul and maximum as y = mul(maximum(x, 0), alpha). As we have mul, will we define maximum instead? |
|||
| MaxPool | ✔️ | ✔️ | ✔️ | ||
| Mul | ✔️ | Need to complete the spec. | |||
| Relu | ✔️ | ✔️ | ✔️ | As relu can be represented by y = maximum(x, 0), can we define maximum instead? |
|
| Reshape | ✔️ | ✔️ | ✔️ | ||
| Softmax | ✔️ | ✔️ | ✔️ | It is for post-processing and not included in ONNX models. |