caffe：即使损失收敛，回归多维输出也会产生奇怪的结果

Question

我有一个带有EculideanLoss图层的回归网络，经过培训后，loss的值大约为3.

layer {
    name: "conv2"
    type: "Convolution"
    bottom: "conv"
    top: "conv2"
    convolution_param {
        num_output: 1 # <-- this is correct right??
        kernel_size: 5
        pad: 2
        stride: 1
        ...
    }
}

layer {
    name: "relu"
    type: "ReLU"
    bottom: "conv2"
    top: "result"
    relu_param{
        negative_slope: 0.01
    }
}

我的数据是尺寸为1 x 128 x 128的图像，而我的.g ground_truth也是具有相同尺寸的图像。我缩放[0,255]的所有值 - ＆gt; [0,1]。

当我尝试检索predicted image我的结果时，会发送一个奇怪的图像，甚至看起来不像ground_truth

我检索输出的python脚本如下所示：

net.blobs['data'].data[...] = transformer.preprocess('data', img)
pred = net.forward()
output_blob = pred['result']

predicated_image_array = np.array(output_blob)
predicated_image_array = predicated_image_array.squeeze()

range_value = np.ptp(predicated_image_array)
min_value = predicated_image_array.min()
max_value = predicated_image_array.max()

# make positive
predicated_image_array[:] -= min_value

if not range_value == 0:
    predicated_image_array /= range_value

predicated_image_array *= 255

predicated_image_array = predicated_image_array.astype(np.int64)
cv2.imwrite('predicted_output.jpg', predicated_image_array)

价值损失3是否太高或者我的python脚本出了什么问题？

提示：当我使用SoftmaxWithLoss图层而不是EuclideanLayer时，并且不会缩放[0,255]中的值 - ＆gt; [0,1]但是留下它们让我的标签来自[0,255]我得到了一个非常不错的结果！