Compile PyTorch Object Detection Models

This article is an introductory tutorial to deploy PyTorch object detection models with Relay VM.

For us to begin with, PyTorch should be installed. TorchVision is also required since we will be using it as our model zoo.

A quick solution is to install via pip

pip install torch==1.4.0
pip install torchvision==0.5.0

or please refer to official site

PyTorch versions should be backwards compatible but should be used with the proper TorchVision version.

Currently, TVM supports PyTorch 1.4 and 1.3. Other versions may be unstable.

import tvm
from tvm import relay
from tvm import relay
from tvm.runtime.vm import VirtualMachine
from import download

import numpy as np
import cv2

# PyTorch imports
import torch
import torchvision

Load pre-trained maskrcnn from torchvision and do tracing

in_size = 300

input_shape = (1, 3, in_size, in_size)

def do_trace(model, inp):
    model_trace = torch.jit.trace(model, inp)
    return model_trace

def dict_to_tuple(out_dict):
    if "masks" in out_dict.keys():
        return out_dict["boxes"], out_dict["scores"], out_dict["labels"], out_dict["masks"]
    return out_dict["boxes"], out_dict["scores"], out_dict["labels"]

class TraceWrapper(torch.nn.Module):
    def __init__(self, model):
        self.model = model

    def forward(self, inp):
        out = self.model(inp)
        return dict_to_tuple(out[0])

model_func = torchvision.models.detection.maskrcnn_resnet50_fpn
model = TraceWrapper(model_func(pretrained=True))

inp = torch.Tensor(np.random.uniform(0.0, 250.0, size=(1, 3, in_size, in_size)))

with torch.no_grad():
    out = model(inp)
    script_module = do_trace(model, inp)


/usr/local/lib/python3.6/dist-packages/torch/ RuntimeWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
  'incorrect results).', category=RuntimeWarning)
/usr/local/lib/python3.6/dist-packages/torchvision/models/detection/ UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  return torch.tensor(M + 2 * padding).to(torch.float32) / torch.tensor(M).to(torch.float32)

Download a test image and pre-process

img_path = "test_street_small.jpg"
img_url = (
    "" "master/gluoncv/detection/street_small.jpg"
download(img_url, img_path)

img = cv2.imread(img_path).astype("float32")
img = cv2.resize(img, (in_size, in_size))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = np.transpose(img / 255.0, [2, 0, 1])
img = np.expand_dims(img, axis=0)


File test_street_small.jpg exists, skip.

Import the graph to Relay

input_name = "input0"
shape_list = [(input_name, input_shape)]
mod, params = relay.frontend.from_pytorch(script_module, shape_list)

Compile with Relay VM

Note: Currently only CPU target is supported. For x86 target, it is highly recommended to build TVM with Intel MKL and Intel OpenMP to get best performance, due to the existence of large dense operator in torchvision rcnn models.

# Add "-libs=mkl" to get best performance on x86 target.
# For x86 machine supports AVX512, the complete target is
# "llvm -mcpu=skylake-avx512 -libs=mkl"
target = "llvm"

with tvm.transform.PassContext(opt_level=3, disabled_pass=["FoldScaleAxis"]):
    vm_exec = relay.vm.compile(mod, target=target, params=params)

Inference with Relay VM

ctx = tvm.cpu()
vm = VirtualMachine(vm_exec, ctx)
vm.set_input("main", **{input_name: img})
tvm_res =

Get boxes with score larger than 0.9

score_threshold = 0.9
boxes = tvm_res[0].asnumpy().tolist()
valid_boxes = []
for i, score in enumerate(tvm_res[1].asnumpy().tolist()):
    if score > score_threshold:

print("Get {} valid boxes".format(len(valid_boxes)))


Get 9 valid boxes

Total running time of the script: ( 2 minutes 2.379 seconds)

Gallery generated by Sphinx-Gallery