Deploy Pretrained Vision Model from MxNet on VTA

Author: Thierry Moreau

This tutorial provides an end-to-end demo, on how to run ImageNet classification inference onto the VTA accelerator design to perform ImageNet classification tasks. It showcases Relay as a front end compiler that can perform quantization (VTA only supports int8/32 inference) as well as graph packing (in order to enable tensorization in the core) to massage the compute graph for the hardware target.

Install dependencies

To use the autotvm package in tvm, we need to install some extra dependencies. (change “3” to “2” if you use python2):

pip3 install --user mxnet requests "Pillow<7"

Now return to the python code. Import packages.

from __future__ import absolute_import, print_function

import argparse, json, os, requests, sys, time
from io import BytesIO
from os.path import join, isfile
from PIL import Image

from mxnet.gluon.model_zoo import vision
import numpy as np
from matplotlib import pyplot as plt

import tvm
from tvm import te
from tvm import rpc, autotvm, relay
from tvm.contrib import graph_runtime, util, download
from tvm.contrib.debugger import debug_runtime
from tvm.relay import transform

import vta
from vta.testing import simulator
from import graph_pack

# Make sure that TVM was compiled with RPC=1
assert tvm.runtime.enabled("rpc")

Define the platform and model targets

Execute on CPU vs. VTA, and define the model.

# Load VTA parameters from the 3rdparty/vta-hw/config/vta_config.json file
env = vta.get_env()

# Set ``device=arm_cpu`` to run inference on the CPU
# or ``device=vta`` to run inference on the FPGA.
device = "vta"
target = if device == "vta" else env.target_vta_cpu

# Dictionary lookup for when to start/end bit packing
pack_dict = {
    "resnet18_v1": ["nn.max_pool2d", "nn.global_avg_pool2d"],
    "resnet34_v1": ["nn.max_pool2d", "nn.global_avg_pool2d"],
    "resnet18_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"],
    "resnet34_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"],
    "resnet50_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"],
    "resnet101_v2": ["nn.max_pool2d", "nn.global_avg_pool2d"],

# Name of Gluon model to compile
# The ``start_pack`` and ``stop_pack`` labels indicate where
# to start and end the graph packing relay pass: in other words
# where to start and finish offloading to VTA.
model = "resnet18_v1"
assert model in pack_dict

Obtain an execution remote

When target is ‘pynq’, reconfigure FPGA and runtime. Otherwise, if target is ‘sim’, execute locally.

if env.TARGET not in ["sim", "tsim"]:

    # Get remote from tracker node if environment variable is set.
    # To set up the tracker, you'll need to follow the "Auto-tuning
    # a convolutional network for VTA" tutorial.
    tracker_host = os.environ.get("TVM_TRACKER_HOST", None)
    tracker_port = os.environ.get("TVM_TRACKER_PORT", None)
    # Otherwise if you have a device you want to program directly from
    # the host, make sure you've set the variables below to the IP of
    # your board.
    device_host = os.environ.get("VTA_RPC_HOST", "")
    device_port = os.environ.get("VTA_RPC_PORT", "9091")
    if not tracker_host or not tracker_port:
        remote = rpc.connect(device_host, int(device_port))
        remote = autotvm.measure.request_remote(env.TARGET, tracker_host, int(tracker_port), timeout=10000)

    # Reconfigure the JIT runtime and FPGA.
    # You can program the FPGA with your own custom bitstream
    # by passing the path to the bitstream file instead of None.
    reconfig_start = time.time()
    vta.program_fpga(remote, bitstream=None)
    reconfig_time = time.time() - reconfig_start
    print("Reconfigured FPGA and RPC runtime in {0:.2f}s!".format(reconfig_time))

# In simulation mode, host the RPC server locally.
    remote = rpc.LocalSession()

# Get execution context from remote
ctx = remote.ext_dev(0) if device == "vta" else remote.cpu(0)

Build the inference graph runtime

Grab vision model from Gluon model zoo and compile with Relay. The compilation steps are:

  1. Front end translation from MxNet into Relay module.

  2. Apply 8-bit quantization: here we skip the first conv layer, and dense layer which will both be executed in fp32 on the CPU.

  3. Perform graph packing to alter the data layout for tensorization.

  4. Perform constant folding to reduce number of operators (e.g. eliminate batch norm multiply).

  5. Perform relay build to object file.

  6. Load the object file onto remote (FPGA device).

  7. Generate graph runtime, m.

# Load pre-configured AutoTVM schedules
with autotvm.tophub.context(target):

    # Populate the shape and data type dictionary for ImageNet classifier input
    dtype_dict = {"data": 'float32'}
    shape_dict = {"data": (env.BATCH, 3, 224, 224)}

    # Get off the shelf gluon model, and convert to relay
    gluon_model = vision.get_model(model, pretrained=True)

    # Measure build start time
    build_start = time.time()

    # Start front end compilation
    mod, params = relay.frontend.from_mxnet(gluon_model, shape_dict)

    # Update shape and type dictionary
    shape_dict.update({k: v.shape for k, v in params.items()})
    dtype_dict.update({k: str(v.dtype) for k, v in params.items()})

    if target.device_name == "vta":
        # Perform quantization in Relay
        # Note: We set opt_level to 3 in order to fold batch norm
        with tvm.transform.PassContext(opt_level=3):
            with relay.quantize.qconfig(global_scale=8.0,
                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(
        relay_prog = mod["main"]

    # Compile Relay program with AlterOpLayout disabled
    if target.device_name != "vta":
        with tvm.transform.PassContext(opt_level=3, disabled_pass={"AlterOpLayout"}):
            graph, lib, params =
                relay_prog, target=target,
                params=params, target_host=env.target_host)
        with vta.build_config(opt_level=3, disabled_pass={"AlterOpLayout"}):
            graph, lib, params =
                relay_prog, target=target,
                params=params, target_host=env.target_host)

    # Measure Relay build time
    build_time = time.time() - build_start
    print(model + " inference graph built in {0:.2f}s!".format(build_time))

    # Send the inference library over to the remote RPC server
    temp = util.tempdir()"graphlib.o"))
    lib = remote.load_module("graphlib.o")

    # Graph runtime
    m = graph_runtime.create(graph, lib, ctx)


...12%, 0.01 MB, 48 KB/s, 0 seconds passed
...25%, 0.02 MB, 97 KB/s, 0 seconds passed
...38%, 0.02 MB, 145 KB/s, 0 seconds passed
...51%, 0.03 MB, 194 KB/s, 0 seconds passed
...64%, 0.04 MB, 241 KB/s, 0 seconds passed
...77%, 0.05 MB, 289 KB/s, 0 seconds passed
...89%, 0.05 MB, 336 KB/s, 0 seconds passed
...100%, 0.06 MB, 383 KB/s, 0 seconds passed
resnet18_v1 inference graph built in 5.85s!

Perform image classification inference

We run classification on an image sample from ImageNet We just need to download the categories files, synset.txt and an input test image.

# Download ImageNet categories
categ_url = ""
categ_fn = "synset.txt", categ_fn), categ_fn)
synset = eval(open(categ_fn).read())

# Download test image
image_url = ''
image_fn = 'cat.png', image_fn)

# Prepare test image for inference
image =, 224))
image = np.array(image) - np.array([123., 117., 104.])
image /= np.array([58.395, 57.12, 57.375])
image = image.transpose((2, 0, 1))
image = image[np.newaxis, :]
image = np.repeat(image, env.BATCH, axis=0)

# Set the network parameters and inputs
m.set_input('data', image)

# Perform inference and gather execution statistics
# More on: :py:method:`tvm.runtime.Module.time_evaluator`
num = 4 # number of times we run module for a single measurement
rep = 3 # number of measurements (we derive std dev from this)
timer = m.module.time_evaluator("run", ctx, number=num, repeat=rep)

if env.TARGET in ["sim", "tsim"]:
    sim_stats = simulator.stats()
    print("\nExecution statistics:")
    for k, v in sim_stats.items():
        # Since we execute the workload many times, we need to normalize stats
        # Note that there is always one warm up run
        # Therefore we divide the overall stats by (num * rep + 1)
        print("\t{:<16}: {:>16}".format(k, v // (num * rep + 1)))
    tcost = timer()
    std = np.std(tcost.results) * 1000
    mean = tcost.mean * 1000
    print("\nPerformed inference in %.2fms (std = %.2f) for %d samples" % (mean, std, env.BATCH))
    print("Average per sample inference time: %.2fms" % (mean/env.BATCH))

# Get classification results
tvm_output = m.get_output(0, tvm.nd.empty((env.BATCH, 1000), "float32", remote.cpu(0)))
for b in range(env.BATCH):
    top_categories = np.argsort(tvm_output.asnumpy()[b])
    # Report top-5 classification results
    print("\n{} prediction for sample {}".format(model, b))
    print("\t#1:", synset[top_categories[-1]])
    print("\t#2:", synset[top_categories[-2]])
    print("\t#3:", synset[top_categories[-3]])
    print("\t#4:", synset[top_categories[-4]])
    print("\t#5:", synset[top_categories[-5]])
    # This just checks that one of the 5 top categories
    # is one variety of cat; this is by no means an accurate
    # assessment of how quantization affects classification
    # accuracy but is meant to catch changes to the
    # quantization pass that would accuracy in the CI.
    cat_detected = False
    for k in top_categories[-5:]:
        if "cat" in synset[k]:
            cat_detected = True


File synset.txt exists, skip.
File cat.png exists, skip.

Execution statistics:
        inp_load_nbytes :          5549568
        wgt_load_nbytes :         12763136
        acc_load_nbytes :            30720
        uop_load_nbytes :            22832
        out_store_nbytes:          1680896
        gemm_counter    :          6623232
        alu_counter     :           572320

resnet18_v1 prediction for sample 0
        #1: tiger cat
        #2: Egyptian cat
        #3: tabby, tabby cat
        #4: lynx, catamount
        #5: weasel

Gallery generated by Sphinx-Gallery