部署 Single Shot Multibox Detector（SSD）模型
备注
单击 此处 下载完整的示例代码
本文介绍如何用 TVM 部署 SSD 模型。这里将使用 GluonCV 预训练的 SSD 模型，并将其转换为 Relay IR。
import tvm
from tvm import te
from matplotlib import pyplot as plt
from tvm import relay
from tvm.contrib import graph_executor
from tvm.contrib.download import download_testdata
from gluoncv import model_zoo, data, utils
输出结果：
/usr/local/lib/python3.7/dist-packages/gluoncv/__init__.py:40: UserWarning: Both `mxnet==1.6.0` and `torch==1.11.0+cpu` are installed. You might encounter increased GPU memory footprint if both framework are used at the same time.
warnings.warn(f'Both `mxnet=={mx.__version__}` and `torch=={torch.__version__}` are installed. '
初步参数设置
备注
现在支持在 CPU 和 GPU 上编译 SSD。
为取得 CPU 上的最佳推理性能，需要根据设备修改 target 参数——对于 x86 CPU：参考 为 x86 CPU 自动调整卷积网络 来调整；对于 arm CPU：参考 为 ARM CPU 自动调整卷积网络 来调整。
为在 Intel 显卡上取得最佳推理性能，将 target 参数修改为 opencl -device=intel_graphics 。注意：在 Mac 上使用 Intel 显卡时，target 要设置为 opencl ，因为 Mac 上不支持 Intel 子组扩展。
为取得基于 CUDA 的 GPU 上的最佳推理性能，将 target 参数修改为 cuda；对于基于 OPENCL 的 GPU，将 target 参数修改为 opencl，然后根据设备来修改设备参数。
supported_model = [
"ssd_512_resnet50_v1_voc",
"ssd_512_resnet50_v1_coco",
"ssd_512_resnet101_v2_voc",
"ssd_512_mobilenet1.0_voc",
"ssd_512_mobilenet1.0_coco",
"ssd_300_vgg16_atrous_voc" "ssd_512_vgg16_atrous_coco",
]
model_name = supported_model[0]
dshape = (1, 3, 512, 512)
下载并预处理 demo 图像：
im_fname = download_testdata(
"https://github.com/dmlc/web-data/blob/main/" + "gluoncv/detection/street_small.jpg?raw=true",
"street_small.jpg",
module="data",
)
x, img = data.transforms.presets.ssd.load_test(im_fname, short=512)
为 CPU 转换和编译模 型：
block = model_zoo.get_model(model_name, pretrained=True)
def build(target):
mod, params = relay.frontend.from_mxnet(block, {"data": dshape})
with tvm.transform.PassContext(opt_level=3):
lib = relay.build(mod, target, params=params)
return lib
输出结果：
/usr/local/lib/python3.7/dist-packages/mxnet/gluon/block.py:1389: UserWarning: Cannot decide type for the following arguments. Consider providing them as input:
data: None
input_sym_arg_type = in_param.infer_type()[0]
Downloading /workspace/.mxnet/models/ssd_512_resnet50_v1_voc-9c8b225a.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/ssd_512_resnet50_v1_voc-9c8b225a.zip...
0%| | 0/132723 [00:00<?, ?KB/s]
2%|1 | 2429/132723 [00:00<00:05, 24288.69KB/s]
8%|8 | 10888/132723 [00:00<00:02, 59757.20KB/s]
14%|#4 | 18798/132723 [00:00<00:01, 68586.60KB/s]
21%|## | 27307/132723 [00:00<00:01, 75099.17KB/s]
27%|##7 | 35836/132723 [00:00<00:01, 78765.91KB/s]
33%|###3 | 44460/132723 [00:00<00:01, 81298.98KB/s]
40%|###9 | 53075/132723 [00:00<00:00, 82882.32KB/s]
46%|####6 | 61612/132723 [00:00<00:00, 83671.87KB/s]
53%|#####2 | 69980/132723 [00:00<00:00, 82355.51KB/s]
59%|#####9 | 78462/132723 [00:01<00:00, 83105.52KB/s]
65%|######5 | 86777/132723 [00:01<00:00, 79179.66KB/s]
72%|#######1 | 95291/132723 [00:01<00:00, 80915.06KB/s]
78%|#######7 | 103417/132723 [00:01<00:00, 62776.56KB/s]
84%|########4 | 111967/132723 [00:01<00:00, 68364.35KB/s]
90%|########9 | 119368/132723 [00:01<00:00, 44237.04KB/s]
96%|#########6| 127829/132723 [00:01<00:00, 51926.12KB/s]
100%|##########| 132723/132723 [00:02<00:00, 64946.94KB/s]
创建 TVM runtime，并进行推理，注意：
Use target = "cuda -libs" to enable thrust based sort, if you
enabled thrust during cmake by -DUSE_THRUST=ON.
def run(lib, dev):
# 构建 TVM runtime
m = graph_executor.GraphModule(lib["default"](dev))
tvm_input = tvm.nd.array(x.asnumpy(), device=dev)
m.set_input("data", tvm_input)
# 执行
m.run()
# 得到输出
class_IDs, scores, bounding_boxs = m.get_output(0), m.get_output(1), m.get_output(2)
return class_IDs, scores, bounding_boxs
for target in ["llvm", "cuda"]:
dev = tvm.device(target, 0)
if dev.exist:
lib = build(target)
class_IDs, scores, bounding_boxs = run(lib, dev)
输出结果：
/workspace/python/tvm/driver/build_module.py:268: UserWarning: target_host parameter is going to be deprecated. Please pass in tvm.target.Target(target, host=target_host) instead.
"target_host parameter is going to be deprecated. "
显示结果：
ax = utils.viz.plot_bbox(
img,
bounding_boxs.numpy()[0],
scores.numpy()[0],
class_IDs.numpy()[0],
class_names=block.classes,
)
plt.show()
脚本总运行时长：（ 2 分 32.231 秒）