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
pip install torchvision

or please refer to official site https://pytorch.org/get-started/locally/

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

Currently, TVM supports PyTorch 1.7 and 1.4. Other versions may be unstable.

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

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)
    model_trace.eval()
    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):
        super().__init__()
        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))

model.eval()
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)

/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.
  warnings.warn(
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=MaskRCNN_ResNet50_FPN_Weights.COCO_V1`. You can also use `weights=MaskRCNN_ResNet50_FPN_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth" to /workspace/.cache/torch/hub/checkpoints/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth

  0%|          | 0.00/170M [00:00<?, ?B/s]
  4%|3         | 6.30M/170M [00:00<00:03, 45.3MB/s]
  6%|6         | 10.6M/170M [00:00<00:06, 26.8MB/s]
  9%|9         | 16.0M/170M [00:00<00:05, 32.0MB/s]
 14%|#4        | 24.0M/170M [00:00<00:04, 36.4MB/s]
 18%|#7        | 30.3M/170M [00:00<00:04, 31.5MB/s]
 20%|#9        | 33.5M/170M [00:01<00:04, 31.1MB/s]
 24%|##3       | 40.0M/170M [00:01<00:03, 35.9MB/s]
 27%|##7       | 46.3M/170M [00:01<00:03, 42.0MB/s]
 30%|##9       | 50.6M/170M [00:01<00:03, 38.1MB/s]
 33%|###2      | 56.0M/170M [00:01<00:02, 40.3MB/s]
 37%|###6      | 62.3M/170M [00:01<00:02, 43.0MB/s]
 39%|###9      | 66.5M/170M [00:01<00:02, 39.9MB/s]
 42%|####2     | 72.0M/170M [00:02<00:02, 40.0MB/s]
 46%|####6     | 78.3M/170M [00:02<00:02, 45.2MB/s]
 49%|####8     | 82.8M/170M [00:02<00:02, 41.6MB/s]
 52%|#####1    | 88.0M/170M [00:02<00:01, 43.4MB/s]
 56%|#####5    | 94.3M/170M [00:02<00:01, 46.8MB/s]
 58%|#####8    | 98.9M/170M [00:02<00:01, 41.3MB/s]
 61%|######    | 103M/170M [00:02<00:01, 38.9MB/s]
 66%|######5   | 112M/170M [00:02<00:01, 42.2MB/s]
 71%|#######   | 120M/170M [00:03<00:01, 45.7MB/s]
 75%|#######5  | 128M/170M [00:03<00:00, 44.6MB/s]
 80%|#######9  | 136M/170M [00:03<00:00, 51.8MB/s]
 83%|########2 | 141M/170M [00:03<00:00, 50.1MB/s]
 86%|########5 | 146M/170M [00:03<00:00, 42.6MB/s]
 88%|########8 | 150M/170M [00:03<00:00, 35.1MB/s]
 91%|######### | 154M/170M [00:04<00:00, 34.0MB/s]
 94%|#########4| 160M/170M [00:04<00:00, 36.1MB/s]
 98%|#########7| 166M/170M [00:04<00:00, 36.3MB/s]
100%|##########| 170M/170M [00:04<00:00, 39.8MB/s]
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torch/nn/functional.py:3912: 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).
  (torch.floor((input.size(i + 2).float() * torch.tensor(scale_factors[i], dtype=torch.float32)).float()))
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/ops/boxes.py:157: 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).
  boxes_x = torch.min(boxes_x, torch.tensor(width, dtype=boxes.dtype, device=boxes.device))
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/ops/boxes.py:159: 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).
  boxes_y = torch.min(boxes_y, torch.tensor(height, dtype=boxes.dtype, device=boxes.device))
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torch/__init__.py:1209: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert condition, message
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/models/detection/transform.py:298: 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).
  torch.tensor(s, dtype=torch.float32, device=boxes.device)
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/models/detection/transform.py:299: 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).
  / torch.tensor(s_orig, dtype=torch.float32, device=boxes.device)
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/torchvision/models/detection/roi_heads.py:389: 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_url = (
    "https://raw.githubusercontent.com/dmlc/web-data/master/gluoncv/detection/street_small.jpg"
)
img_path = download_testdata(img_url, "test_street_small.jpg", module="data")

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)

Import the graph to Relay

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

/workspace/python/tvm/relay/frontend/pytorch_utils.py:47: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
  return LooseVersion(torch_ver) > ver
/venv/apache-tvm-py3.8/lib/python3.8/site-packages/setuptools/_distutils/version.py:346: DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
  other = LooseVersion(other)
/workspace/python/tvm/relay/build_module.py:345: DeprecationWarning: Please use input parameter mod (tvm.IRModule) instead of deprecated parameter mod (tvm.relay.function.Function)
  warnings.warn(

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

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

Get boxes with score larger than 0.9

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

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

Get 9 valid boxes