Torch - 加载保存的模型并测试单个图像进行分类

Question

我正在使用来自此存储库的代码 - gtsrb。我已经成功地以超过 95% 的准确率训练了模型。但是，我不确定如何针对已经训练好的模型测试单个图像。我有一个训练模型保存在saved_models 目录下，名称为egg_net_best.h5。

这是我到目前为止所做的：

input = image.load(path_to_single_image, 3)
model = torch.load('saved_models/vgg_net_best.h5')
output = model:forward(input)

这给了我一个输出，但我不明白格式。它给了我一个 9x11 张量，看起来像这样

-3.1302 -2.7557 -3.6514 -5.1083 -0.4339 -3.6722 -6.7406 -4.1163 -3.2751 -3.0417 -2.4740
-4.6895 -4.5554 -1.1019 -4.1954 -1.0187 -3.2853 -4.7917 -5.1205 -2.6658  -2.3842 -2.8388
-3.2914 -2.7516 -2.1224 -4.6314 -0.7596 -3.0647 -5.5749 -3.9037 -3.4147 -2.9719 -1.9181
-3.7006 -3.0220 -2.3870 -5.2183 -0.5729 -3.1286 -6.1271 -4.4617 -3.6552 -3.1698 -1.9662
-2.4560 -2.7865 -2.0729 -2.7589 -2.4757 -2.6934 -1.3172 -2.5542 -3.0839 -3.2817 -2.4930
-3.5149 -3.2504 -2.4154 -4.9333 -0.6504 -3.3700 -5.3838 -4.4408 -2.5233 -2.3310 -2.4659
-3.5452 -4.1833 -3.3995 -2.3409 -2.9932 -2.7852 -0.5702 -3.0504 -2.8158 -3.8218 -3.9147
-3.1225 -2.7842 -2.9371 -4.6879 -0.6197 -3.1266 -5.6755 -4.0995 -3.7424 -3.1942 -1.8000
-2.9004 -2.4582 -2.5652 -4.6114 -0.6186 -3.3890 -4.1276 -3.7448 -3.1053 -3.0396 -2.6929

问题：

1. 针对经过训练的模型测试单个图像的方法是否正确？
2. 为什么我看到的是 9x11 张量，而它应该是 1x11？

有人能指出正确的方向吗？

PS：这个问题SO- training single image不是我要找的。​​p>

编辑 1：

根据建议我做了以下但结果是一样的。

Lua 5.1 版权所有 (C) 1994-2008 Lua.org, PUC-Rio

path = '/Users/pamio/Workspace/practices/torch/test_model/test/00010/IMG_2522.JPG'

input = image.load(path, 3)

model = torch.load('saved_models/vgg_net_best.h5')

模型：评估（）

输出 = 模型：前向（输入）：exp（）

=输出

第 1 到 10 列

   0.0257  0.0621  0.0027  0.0169  0.0132  0.0530  0.6320  0.0798  0.0022  0.0138 0.0985
   0.1033  0.0838  0.0705  0.0609  0.1639  0.0678  0.1156  0.0971  0.0653  0.1042 0.0676
   0.0101  0.0262  0.0010  0.0053  0.0014  0.0312  0.8307  0.0349  0.0002  0.0050 0.0541
   0.0666  0.1656  0.0054  0.0337  0.0112  0.0537  0.3638  0.1694  0.0062  0.0134 0.1111
   0.0589  0.0663  0.0298  0.0148  0.1565  0.0976  0.2328  0.1159  0.0324  0.0978 0.0972
   0.0511  0.0786  0.0153  0.0549  0.0341  0.1121  0.3010  0.1052  0.0314  0.0994 0.1169
   0.0619  0.0777  0.0395  0.0521  0.0358  0.1125  0.3184  0.0785  0.0150  0.0588 0.1498
   0.0153  0.0319  0.0009  0.0080  0.0015  0.0386  0.7779  0.0449  0.0003  0.0079 0.0726
   0.0123  0.0285  0.0059  0.0062  0.0048  0.0455  0.7893  0.0269  0.0013  0.0177 0.0615

   [torch.DoubleTensor of size 9x11]

编辑2：

> print(model)
nn.Sequential {
  [input -> (1) -> (2) -> output]
    (1): nn.Sequential {
      [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> (13) -> (14) -> (15) -> (16) -> (17) -> (18) -> (19) -> (20) -> (21) -> output]
    (1): nn.SpatialConvolution(3 -> 32, 3x3, 1,1, 1,1)
    (2): nn.SpatialBatchNormalization (4D) (32)
    (3): nn.ReLU
    (4): nn.SpatialConvolution(32 -> 32, 3x3, 1,1, 1,1)
    (5): nn.SpatialBatchNormalization (4D) (32)
    (6): nn.ReLU
    (7): nn.SpatialMaxPooling(2x2, 2,2)
    (8): nn.SpatialConvolution(32 -> 64, 3x3, 1,1, 1,1)
    (9): nn.SpatialBatchNormalization (4D) (64)
    (10): nn.ReLU
    (11): nn.SpatialConvolution(64 -> 64, 3x3, 1,1, 1,1)
    (12): nn.SpatialBatchNormalization (4D) (64)
    (13): nn.ReLU
    (14): nn.SpatialMaxPooling(2x2, 2,2)
    (15): nn.SpatialConvolution(64 -> 128, 3x3, 1,1, 1,1)
    (16): nn.SpatialBatchNormalization (4D) (128)
    (17): nn.ReLU
    (18): nn.SpatialConvolution(128 -> 128, 3x3, 1,1, 1,1)
    (19): nn.SpatialBatchNormalization (4D) (128)
    (20): nn.ReLU
    (21): nn.SpatialMaxPooling(2x2, 2,2)
  }
 (2): nn.Sequential {
    [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> output]
    (1): nn.View(4608)
    (2): nn.Linear(4608 -> 512)
    (3): nn.ReLU
    (4): nn.BatchNormalization (2D) (512)
    (5): nn.Dropout(0.500000)
    (6): nn.Linear(512 -> 11)
    (7): nn.LogSoftMax
  }
 }

Answer 1

您正在加载的图片似乎尺寸不正确。该网络使用 3x48x48 的图像进行训练，而您为它们提供更大的高度和宽度会使模型内部层混乱，并将它们视为批次而不是单个图像。

以下代码对我有用：

model:evaluate();

-- We assume that the image has a wrong size, place here your image loading code
img = torch.Tensor(3, 200, 200)
-- First we scale it to 3x48x48
img = image.scale(img, 48, 48, 'bilinear')
-- Then to avoid errors with the batchnorm layer we make it 4D 
-- (batchNorm expects a 4D tensor)
image = image:reshape(1, 3, 48, 48)
-- Feed the image to the network, add the :exp() if we want to retrieve the probs.
pred = model:forward(img):exp()


-- output
Columns 1 to 10
0.01 *
9.4125  9.0546  8.9142  9.0716  9.0065  9.0865  9.4289  8.7353  9.2937  9.2811

Columns 11 to 11
0.01 *
8.7152
[torch.DoubleTensor of size 1x11]