索尼神经网络控制台Windows版本进行了更新，新增了用于后处理评估结果的插件功能。此次更新增加了四个插件：Grad-CAM用于可视化影响识别结果的输入数据；Cross Tabulation用于执行交叉制表分析，例如比较分类任务中每个类别的准确性；Parameter Stats用于显示模型参数的各种统计信息，例如最小值、最大值、均值、标准差等；Tile Images用于将评估数据中包含的图像拼接成单个图像。此外，更新还增加了GELU、ReLU6等多个层和求解器。索尼鼓励用户反馈，以进一步改进神经网络控制台。

💡 **插件功能新增：**此次更新的核心是引入了插件功能，允许用户对评估结果进行后处理，例如可视化、统计分析和图像拼接等操作，提升了神经网络控制台的灵活性和实用性。

📊 **Grad-CAM插件：**该插件可以可视化输入数据中对识别结果影响最大的部分，帮助用户理解模型的决策过程，例如在图像识别任务中，可以显示出模型关注图像的哪些区域从而做出判断。

🧮 **Cross Tabulation插件：**此插件可以对评估结果进行交叉制表分析，例如在分类任务中，可以比较不同类别的准确率，或者分析不同特征之间的关系，帮助用户更深入地理解模型的性能。

📈 **Parameter Stats插件：**此插件用于显示模型参数的各种统计信息，例如最小值、最大值、均值、标准差等，这对于模型的分析和调试非常有用，例如在参数量化时，可以检查参数值是否在可量化的范围内。

🖼️ **Tile Images插件：**此插件可以将多个图像拼接成一个图像，例如在使用GAN生成图像时，可以将生成的图像拼接成一个整体，方便用户观察生成结果。

We have updated Neural Network Console Windows today.
In this version, plug-ins are now available for post-processing evaluation results.

To use the plug-in, right-click the evaluation result of the EVALUATION tab to open a shortcut menu, and click Plugin.

The four plug-ins added this time are introduced below.

 

Grad-CAM

Grad-CAM (*1) is one of the popular methods for visualizing input data that greatly affects recognition results.
To use Grad-CAM,

Run training and evaluation in image recognition project using Convolutional Neural NetworksSelect the evaluation image displayed on the EVALUATION tabRight-click the evaluation result of the EVALUATION tab to open a shortcut menu, and select Plugin and Grad-CAM.Specify the index of the class to be visualized in class_index (for example, 0 to 999 for 1000 class classification)

To display a larger view of the Grad-CAM result image, double-click the result image displayed on the EVALUATION tab.

 

Cross Tabulation

Perform cross-tabulation analysis the evaluation results.
For example, cross tabulation can be used to compare the accuracy of each class in a classification tasks.

Specify the variable name used for the row of the tabulation result in variable1 and the variable name used for the column in variable2.
To use correct / incorrect between the label and the estimation for the column, specify the variable name to be compared with variable2 in valuable2_eval.

 

Parameter Stats

Displays various statistics (minimum, maximum, mean, standard deviation, etc.) for model parameters.
The following is an example of using statistics of parameters.

Check whether the parameter value is within the quantifiable range when performing parameter quantization at inferenceSet the minimum value (Delta) for quantization when performing parameter quantization at training

 

Tile Images

The images included in the evaluation data are tiled as a single image.

For example, it can be used to tile the images generated by using a method such as GAN.

Specify the number of images to be arranged in row in num_column, and the index of the first and last images in start_index and end_index.

To display a larger image of the result, Double-click the result image displayed on the EVALUATION tab.

 

You can easily add your own plug-ins by creating an executable Python script from the command line.
Please refer to the existing plug-in files in the libs/plugins folder for how to make plug-ins.
The created plug-in can be called from Neural Network Console by copying it to the libs/plugins folder.

 

In addition, several other layers and solvers have been added.

GELU
ReLU6
HardSigmoid
HardTanh
LogSigmoid
SoftPlus
SoftSign
TanhShrink
Tile
ATan2
FusedBatchNormalization
AdaBound
AMSBound

 

We will continue to improve Neural Network Console.
We look forward to hearing feedbacks from the users for further improvements!

 

Neural Network Console Windows
https://dl.sony.com/app/

※1
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra
https://arxiv.org/abs/1610.02391