search

intel / webml-polyfill

Deprecated, the Web Neural Network Polyfill project has been moved to https://github.com/webmachinelearning/webnn-polyfill
Apache License 2.0
160 stars 42 forks source link

[BNNS]Support Deeplab model #445

Open lisa0314 opened 5 years ago

lisa0314 commented 5 years ago

Since the workaround doesn't work well for deeplab since the depth_out is to large. We will need the support from Apple for depthwise_conv.

huningxin commented 5 years ago

@lisa0314 , could you please specify the original issue, your workaround and the result? Thanks!

lisa0314 commented 5 years ago

Since bnns.h doesn't support the depthwise_conv operation, I try to translate the depthwise_conv into the common conv operation. It works for mobilenet at high cost. However, for deeplab, it crashed . When the input for a depthwise_conv layer is [1, 65, 65, 960], I will need to create conv_weight which has a size of 65 65 960 * 960. And When I pass the conv_weight to the function BNNSFilterCreateConvolutionLayer, it will crash since it’s too large.

lisa0314 commented 5 years ago