Hi, I have this simple Relay graph:
def @main(%input: Tensor[(1, 3, 224, 224), float32], %model/Conv1/Conv2D_weights_fused_bn: Tensor[(32, 3, 3, 3), float32], %model/Conv1/Conv2D_bias_fused_bn: Tensor[(32), float32], %model/expanded_conv_depthwise/depthwise_weights_fused_bn: Tensor[(32, 1, 3, 3), float32], %model/expanded_conv_depthwise/depthwise_bias_fused_bn: Tensor[(32), float32], %model/expanded_conv_project/Conv2D_weights_fused_bn: Tensor[(16, 32, 1, 1), float32], %model/expanded_conv_project/Conv2D_bias_fused_bn: Tensor[(16), float32], %v307: Tensor[(96, 16, 1, 1), float32], %v309: Tensor[(96), float32]) {
%0 = nn.conv2d(%input, %model/Conv1/Conv2D_weights_fused_bn, strides=[2, 2], padding=[0, 0, 1, 1], kernel_size=[3, 3]);
%1 = nn.bias_add(%0, %model/Conv1/Conv2D_bias_fused_bn);
%2 = nn.relu(%1);
%3 = nn.conv2d(%2, %model/expanded_conv_depthwise/depthwise_weights_fused_bn, padding=[1, 1, 1, 1], groups=32, kernel_size=[3, 3]);
%4 = nn.bias_add(%3, %model/expanded_conv_depthwise/depthwise_bias_fused_bn);
%5 = nn.relu(%4);
%6 = nn.conv2d(%5, %model/expanded_conv_project/Conv2D_weights_fused_bn, padding=[0, 0, 0, 0], kernel_size=[1, 1]);
%7 = nn.bias_add(%6, %model/expanded_conv_project/Conv2D_bias_fused_bn);
%8 = nn.conv2d(%7, %v307, padding=[0, 0, 0, 0], kernel_size=[1, 1]);
%9 = nn.bias_add(%8, %v309);
nn.relu(%9)
}
I want to run it using the interpreter and have all %1 through %9 tensors saved for later inspection. I tried using the debug runtme like this:
mod, params = relay.frontend.from_onnx(onnx_model, shape_dict, freeze_params=False)
with tvm.transform.PassContext(opt_level=0):
graph_json, libmod, params = relay.build_module.build(mod, target)
dbg_ext = debug_executor.create(graph_json, libmod, tvm.cpu(0),"./test2")
etc… This works and produces the tensors. However, this is not quite what I need, because some optimizations are applied and the output tensors do not match tensors %1 through %9 from the Relay graph above.
Is there a way to run a Relay graph on an input and save exactly those intermediate tensors that are defined in the Relay graph, instead of some transformation of them?
Thank you.