Choose convolution algorithm from relay


#1

Is there a way to configure the relay.nn.conv2d to perform convolution differently? Is there a way to say layer = conv2d(3,64,(3,3), method='direct') or use im2col -> gemm or winograd?


#2

You can use tvm.relay.op.nn.nn.contrib_conv2d_winograd_nnpack_weight_transform and its variants for example.