首先安装 mxnet 模块，可通过 pip 快速安装：
pip install mxnet --user
或参考官方安装指南：https://mxnet.apache.org/versions/master/install/index.html
# 一些标准的导包 import mxnet as mx import tvm import tvm.relay as relay import numpy as np
从 Gluon Model Zoo 下载 Resnet18 模型
本节会下载预训练的 imagenet 模型，并对图像进行分类。
from tvm.contrib.download import download_testdata from mxnet.gluon.model_zoo.vision import get_model from PIL import Image from matplotlib import pyplot as plt block = get_model("resnet18_v1", pretrained=True) img_url = "https://github.com/dmlc/mxnet.js/blob/main/data/cat.png?raw=true" img_name = "cat.png" synset_url = "".join( [ "https://gist.githubusercontent.com/zhreshold/", "4d0b62f3d01426887599d4f7ede23ee5/raw/", "596b27d23537e5a1b5751d2b0481ef172f58b539/", "imagenet1000_clsid_to_human.txt", ] ) synset_name = "imagenet1000_clsid_to_human.txt" img_path = download_testdata(img_url, "cat.png", module="data") synset_path = download_testdata(synset_url, synset_name, module="data") with open(synset_path) as f: synset = eval(f.read()) image = Image.open(img_path).resize((224, 224)) plt.imshow(image) plt.show() def transform_image(image): image = np.array(image) - np.array([123.0, 117.0, 104.0]) image /= np.array([58.395, 57.12, 57.375]) image = image.transpose((2, 0, 1)) image = image[np.newaxis, :] return image x = transform_image(image) print("x", x.shape)
输出结果：
Downloading /workspace/.mxnet/models/resnet18_v1-a0666292.zip08d19deb-ddbf-4120-9643-fcfab19e7541 from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet18_v1-a0666292.zip... x (1, 3, 224, 224)
编译计算图
只需几行代码，即可将 Gluon 模型移植到可移植计算图上。mxnet.gluon 支持 MXNet 静态图（符号）和 HybridBlock。
shape_dict = {"data": x.shape} mod, params = relay.frontend.from_mxnet(block, shape_dict) ## 添加 softmax 算子来提高概率 func = mod["main"] func = relay.Function(func.params, relay.nn.softmax(func.body), None, func.type_params, func.attrs)
接下来编译计算图：
target = "cuda" with tvm.transform.PassContext(opt_level=3): lib = relay.build(func, target, params=params)
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. "
在 TVM 上执行可移植计算图
接下来用 TVM 重现相同的前向计算：
from tvm.contrib import graph_executor dev = tvm.cuda(0) dtype = "float32" m = graph_executor.GraphModule(lib["default"](dev)) # 设置输入 m.set_input("data", tvm.nd.array(x.astype(dtype))) # 执行 m.run() # 得到输出 tvm_output = m.get_output(0) top1 = np.argmax(tvm_output.numpy()[0]) print("TVM prediction top-1:", top1, synset[top1])
输出结果：
TVM prediction top-1: 282 tiger cat
使用带有预训练权重的 MXNet 符号
MXNet 常用 arg_params 和 aux_params 分别存储网络参数，下面将展示如何在现有 API 中使用这些权重：
def block2symbol(block): data = mx.sym.Variable("data") sym = block(data) args = {} auxs = {} for k, v in block.collect_params().items(): args[k] = mx.nd.array(v.data().asnumpy()) return sym, args, auxs mx_sym, args, auxs = block2symbol(block) # 通常将其保存/加载为检查点 mx.model.save_checkpoint("resnet18_v1", 0, mx_sym, args, auxs) # 磁盘上有 "resnet18_v1-0000.params" 和 "resnet18_v1-symbol.json"
对于一般性 MXNet 模型：
mx_sym, args, auxs = mx.model.load_checkpoint("resnet18_v1", 0) # 用相同的 API 来获取 Relay 计算图 mod, relay_params = relay.frontend.from_mxnet(mx_sym, shape_dict, arg_params=args, aux_params=auxs) # 重复相同的步骤，用 TVM 运行这个模型
下载 Python 源代码：from_mxnet.py
下载 Jupyter Notebook：from_mxnet.ipynb
