编译 OneFlow 模型

本篇文章译自英文文档 Compile OneFlow Models tvm 0.14.dev0 documentation

作者是 BBuf (Xiaoyu Zhang) · GitHub

更多 TVM 中文文档可访问 →Apache TVM 是一个端到端的深度学习编译框架，适用于 CPU、GPU 和各种机器学习加速芯片。 | Apache TVM 中文站

本文介绍如何用 Relay 部署 OneFlow 模型。

首先安装 OneFlow 包，可通过 pip 快速安装：

pip install flowvision==0.1.0
python3 -m pip install -f https://release.oneflow.info oneflow==0.7.0+cpu

或参考官网：
https://github.com/Oneflow-Inc/oneflow

目前 TVM 支持 OneFlow 0.7.0，其他版本可能不稳定。

import os, math
from matplotlib import pyplot as plt
import numpy as np
from PIL import Image

# OneFlow 导入
import flowvision
import oneflow as flow
import oneflow.nn as nn

import tvm
from tvm import relay
from tvm.contrib.download import download_testdata

输出结果：

/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional_pil.py:193: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
  def resize(img, size, interpolation=Image.BILINEAR):
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:65: DeprecationWarning: NEAREST is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.NEAREST or Dither.NONE instead.
  Image.NEAREST: "nearest",
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:66: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
  Image.BILINEAR: "bilinear",
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:67: DeprecationWarning: BICUBIC is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BICUBIC instead.
  Image.BICUBIC: "bicubic",
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:68: DeprecationWarning: BOX is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BOX instead.
  Image.BOX: "box",
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:69: DeprecationWarning: HAMMING is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.HAMMING instead.
  Image.HAMMING: "hamming",
/usr/local/lib/python3.7/dist-packages/flowvision/transforms/functional.py:70: DeprecationWarning: LANCZOS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
  Image.LANCZOS: "lanczos",
/usr/local/lib/python3.7/dist-packages/flowvision/data/auto_augment.py:28: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
  _RANDOM_INTERPOLATION = (Image.BILINEAR, Image.BICUBIC)
/usr/local/lib/python3.7/dist-packages/flowvision/data/auto_augment.py:28: DeprecationWarning: BICUBIC is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BICUBIC instead.
  _RANDOM_INTERPOLATION = (Image.BILINEAR, Image.BICUBIC)

加载和保存 OneFlow 的预训练模型

model_name = "resnet18"
model = getattr(flowvision.models, model_name)(pretrained=True)
model = model.eval()

model_dir = "resnet18_model"
if not os.path.exists(model_dir):
    flow.save(model.state_dict(), model_dir)

输出结果：

Downloading: "https://oneflow-public.oss-cn-beijing.aliyuncs.com/model_zoo/flowvision/classification/ResNet/resnet18.zip" to /workspace/.oneflow/flowvision_cache/resnet18.zip

  0%|          | 0.00/41.5M [00:00

加载测试图像​

还是用猫的图像：

from PIL import Image

img_url = "https://github.com/dmlc/mxnet.js/blob/main/data/cat.png?raw=true"
img_path = download_testdata(img_url, "cat.png", module="data")
img = Image.open(img_path).resize((224, 224))

# 预处理图像，并转换为张量
from flowvision import transforms

my_preprocess = transforms.Compose(
    [
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)
img = my_preprocess(img)
img = np.expand_dims(img.numpy(), 0)

将计算图导入到 Relay 中​

将 OneFlow 计算图转换为 Relay 计算图，输入任意名称。

class Graph(flow.nn.Graph):
    def __init__(self, module):
        super().__init__()
        self.m = module

    def build(self, x):
        out = self.m(x)
        return out

graph = Graph(model)
_ = graph._compile(flow.randn(1, 3, 224, 224))

mod, params = relay.frontend.from_oneflow(graph, model_dir)

使用 Relay 构建​

用给定的输入规范，将计算图编译为 llvm target。

target = tvm.target.Target("llvm", host="llvm")
dev = tvm.cpu(0)
with tvm.transform.PassContext(opt_level=3):
    lib = relay.build(mod, target=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​

接下来在 target 上部署编译好的模型：

target = "cuda"
with tvm.transform.PassContext(opt_level=10):
    intrp = relay.build_module.create_executor("graph", mod, tvm.cuda(0), target)

print(type(img))
print(img.shape)
tvm_output = intrp.evaluate()(tvm.nd.array(img.astype("float32")), **params)

输出结果：

(1, 3, 224, 224)

查找分类集名称​

在 1000 个类的分类集中，查找分数最高的第一个：

synset_url = "".join(
    [
        "https://raw.githubusercontent.com/Cadene/",
        "pretrained-models.pytorch/master/data/",
        "imagenet_synsets.txt",
    ]
)
synset_name = "imagenet_synsets.txt"
synset_path = download_testdata(synset_url, synset_name, module="data")
with open(synset_path) as f:
    synsets = f.readlines()

synsets = [x.strip() for x in synsets]
splits = [line.split(" ") for line in synsets]
key_to_classname = {spl[0]: " ".join(spl[1:]) for spl in splits}

class_url = "".join(
    [
        "https://raw.githubusercontent.com/Cadene/",
        "pretrained-models.pytorch/master/data/",
        "imagenet_classes.txt",
    ]
)
class_name = "imagenet_classes.txt"
class_path = download_testdata(class_url, class_name, module="data")
with open(class_path) as f:
    class_id_to_key = f.readlines()

class_id_to_key = [x.strip() for x in class_id_to_key]

# 获得 TVM 分数最高的第一个结果
top1_tvm = np.argmax(tvm_output.numpy()[0])
tvm_class_key = class_id_to_key[top1_tvm]

# 将输入转换为 OneFlow 变量，并获取 OneFlow 结果进行比较
with flow.no_grad():
    torch_img = flow.from_numpy(img)
    output = model(torch_img)

    # 获取 OneFlow 分数最高的第一个结果
    top_oneflow = np.argmax(output.numpy())
    oneflow_class_key = class_id_to_key[top_oneflow]

print("Relay top-1 id: {}, class name: {}".format(top1_tvm, key_to_classname[tvm_class_key]))
print(
    "OneFlow top-1 id: {}, class name: {}".format(top_oneflow, key_to_classname[oneflow_class_key])
)

输出结果：

Relay top-1 id: 281, class name: tabby, tabby cat
OneFlow top-1 id: 281, class name: tabby, tabby cat

下载 Python 源代码：「链接」

下载 Jupyter Notebook：「链接」