NNモデルの中間層出力テンソルをC#で取得する

備忘録のためのメモ
NNの中間層を特徴量抽出器として扱うことがあったりします

pytorchでforward関数の書き換え

モデル本来の推論をしつつ中間層を取得するのは(恐らく)容易ではないため，forward関数を書き換えて中間層の出力を無理やりreturnします．

python

import torch
model :torch.nn.Module = torch.hub.load('pytorch/vision:v0.10.0', 'wide_resnet50_2', pretrained=True)
model.eval()

def _forward_impl(self, x):
    # See note [TorchScript super()]
    # xs = []
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)

    x = self.layer1(x) # <- 欲しい中間層でreturn
    # x = self.layer2(x)
    # x = self.layer3(x)
    # x = self.layer4(x)

    # x = self.avgpool(x)
    # x = torch.flatten(x, 1)
    # x = self.fc(x)

    return x

# %% モデルのforward関数を書き換える
model._forward_impl = type(model._forward_impl)(_forward_impl, model)

ONNX形式でモデル出力

python

#入力テンソルのダミー
x = torch.randn(1, 3, 224, 224, requires_grad=True)
 
# モデル出力
torch.onnx.export(
    model, x, "WideResNet50Layer1.onnx", export_params=True, 
    opset_version=10, do_constant_folding=True,
    input_names = ['input'], output_names = ['output']
)

C#で推論しList型で取得

ONNX RuntimeとImageSharpをNuGetでインストールします．

C#

using System.IO;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

//追加:
using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;
using SixLabors.ImageSharp;
using SixLabors.ImageSharp.Processing;
using SixLabors.ImageSharp.PixelFormats;
using SixLabors.ImageSharp.Formats;


namespace ResNetIntermediateLayerWithONNX
{

    class Program
    {
        static void Main(string[] args)
        {
            string imageFilePath = @"入力画像パス";
            Image<Rgb24> image = Image.Load<Rgb24>(imageFilePath, out IImageFormat format);

            using (Stream imageStream = new MemoryStream())
            {
                image.Mutate(x =>
                {
                    x.Resize(new ResizeOptions
                    {
                        Size = new Size(224, 224),
                        Mode = ResizeMode.Crop
                    });
                });
                image.Save(imageStream, format);
            }

            Tensor<float> input = new DenseTensor<float>(new[] { 1, 3, 224, 224 });
            var mean = new[] { 0.485f, 0.456f, 0.406f };
            var stddev = new[] { 0.229f, 0.224f, 0.225f };
            image.ProcessPixelRows(accessor =>
            {
                for (int y = 0; y < accessor.Height; y++)
                {
                    Span<Rgb24> pixelSpan = accessor.GetRowSpan(y);
                    for (int x = 0; x < accessor.Width; x++)
                    {
                        input[0, 0, y, x] = ((pixelSpan[x].R / 255f) - mean[0]) / stddev[0];
                        input[0, 1, y, x] = ((pixelSpan[x].G / 255f) - mean[1]) / stddev[1];
                        input[0, 2, y, x] = ((pixelSpan[x].B / 255f) - mean[2]) / stddev[2];
                    }
                }
            });

            var inputs = new List<NamedOnnxValue>
            {
                //ここのstring "input" はONNXモデル出力時に決定したinputレイヤの名前
                NamedOnnxValue.CreateFromTensor("input", input) 
            };

            string modelFilePath = "WideResNet50_Layer1_1.onnx";

            var session = new InferenceSession(modelFilePath);
            IDisposableReadOnlyCollection<DisposableNamedOnnxValue> results = session.Run(inputs);

            IEnumerable<float> output = results.First().AsEnumerable<float>();

            List<float> outputAsList = output.ToList();
            Console.WriteLine(outputAsList.Count); 
        }
    }
}

参考