[VTA] ResNet-18 VTA offloading

jwlee · January 18, 2020, 8:36am

Hello, everyone.

I have a question regarding deploy_vision_on_vta.py.

The code states that it offloads ResNet-18 part from [nn.max_pool2d] to [nn.global_avg_pool2d].

I want to how the 1st convolution before nn.max_pool2d is processed in VTA.

Thanks,

Jake

aca88 · January 20, 2020, 7:04am

It isn’t.

AFAIK the 1st convolution will be performed on the Arm core which ~~is in the FPGA fabric~~ part of the Pynq target but not part of the FPGA fabric which is where the VTA is actually implemented.

jwlee · January 21, 2020, 7:45am

Thanks for responding.

Yes, you are right.

What I meant was how 1st convolution was done on the pynq board.

I wrote “VTA” instead of “pynq”, my bad.

Can you tell me the source code folder that helps to understand offloading part?

Thanks,

Jake

aca88 · January 21, 2020, 10:40am

No problem.

I am not really up-to-date with the latest VTA codebase but taking an educated guess (and leaving some of the details for you to find) you can start here

github.com

apache/incubator-tvm/blob/master/vta/tutorials/frontend/deploy_vision_on_vta.py#L178




if target.device_name == "vta":
    # Perform quantization in Relay
    # Note: We set opt_level to 3 in order to fold batch norm
    with relay.build_config(opt_level=3):
        with relay.quantize.qconfig(global_scale=8.0,
                                    skip_conv_layers=[0]):
            mod = relay.quantize.quantize(mod, params=params)
        # Perform graph packing and constant folding for VTA target
        assert env.BLOCK_IN == env.BLOCK_OUT
        relay_prog = graph_pack(
            mod["main"],
            env.BATCH,
            env.BLOCK_OUT,
            env.WGT_WIDTH,
            start_name=pack_dict[model][0],
            stop_name=pack_dict[model][1])
else:
    relay_prog = mod["main"]


# Compile Relay program with AlterOpLayout disabled