Plugin reference

Tuesday, July 30, 2019

Uncategorized

Posted by Masayoshi Kimura

This chapter describes the plugins and their properties that can be used with the Neural Network Console.

 

1 Plugins

1.1 Cross Tabulation

This plugin performs a cross tabulation on a dataset CSV file. It can be used to tabulate the number of data samples per label, calculate accuracies, and the like.

input Specify the dataset CSV file to be processed.
To perform cross tabulation on a CSV file containing evaluation results from Output Result of the Evaluation tab, use the default output_result.csv file.
variable1 Specify the variable name to use for the rows of the cross tabulation result table.
variable2 Specify the variable name to use for the columns of the cross tabulation result table.
variable2_eval To assign the results (correct/incorrect) of an accuracy evaluation performed on a variable specified by Variable2 to the columns of the cross tabulation result table, specify the variable name that will be used to compare to variable2.
For example, to assign the results of comparing the correct label in image classification (y) and the results estimated by the neural network (y’) to the columns of the cross tabulation result table, set variable2 to “y” and variable2_eval to “y’.”
If you specify blank, the value specified by varable2 is assigned to the column of the cross tabulation result table.
output_in_ratio Specify whether to output ratios in a way that each row is 1.
If this is not checked, the number of data samples is output as-is as a value of each cell.
output Specify the name of the CSV file to output the cross tabulation results to.
If cross tabulation is executed from Output Result of the Evaluation tab, the cross tabulation results are saved to the specified file in the training result folder.

 

1.2 Grad-CAM

Using a method called Grad-CAM, the areas of the input image that affect the classification result are made visible in Convolutional Neural Networks, which performs image classification.

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

model Specify the Convolutional Neural Networks model file (*.nnp) that will be used in the Grad-CAM computation.
To perform Grad-CAM based on the training result selected in the Evaluation tab, use the default results.nnp.
image Specify the image file to analyze.
To perform Grad-CAM on a specific image shown in the evaluation results of the Evaluation tab, start the plugin with the cell containing the image file name selected. The image file name will be automatically input in image.
class_index Specify the index of the class to perform visualization on.
By default, visualization is performed on class number 0.
output Specify the name of the image file to output the visualization results to.
If Grad-CAM is executed from the evaluation results of the Evaluation tab, the visualization results are saved to the specified file in the training result folder.

 

1.3 Tile Images

This plugin creates an image by tiling the images included in the input CSV file. This can be used, for example, to list incorrectly classified images.

input Specify the dataset CSV file containing the image file names.
To create an image by tiling the images contained in Output Result of the Evaluation tab, use the default output_result.csv.
To perform processing based on results obtained by filtering Output Result (e.g., incorrectly classified images), save the filtered results to a CSV file using the shortcut menu, and specify the name of the saved file in input.
variable Of the variables included in the CSV file specified by input, specify the variable name that includes the images to be tiled.
If this is not specified, all images for which file names are included in the CSV file specified by input will be processed.
image_width Specify the width of a single input image. If this is not specified, the width of the first image is used.
If the size of an input image is different from the size specified by image_width, the width of the image will be resized to the value specified by image_width.
image_height Specify the height of a single input image. If this is not specified, the height of the first image is used.
If the size of an input image is different from the size specified by image_height, the width of the image will be resized to the value specified by image_height.
num_column Specify the number of columns to tile the images.
The maximum horizontal width of the output image is image_width num_column.
start_index Specify the index of the first data sample to display in tiles.
end_index Specify the index of the last data sample to display in tiles.
end_index start_index + 1 pages of images are displayed in tiles.
If end_index is not specified, all images contained in all the data in the CSV file specified by input will be displayed.
output Specify the name of the image file to output the tiled images to.
If Tile Images is executed from the evaluation results of the Evaluation tab, the tiled images are saved to the specified file in the training result folder.

 

1.4 Parameter Stats

This plugin calculates various statistics (size, maximum value, minimum value, absolute maximum value, absolute minimum value, absolute value, average, standard deviation) of the parameters included in the train model.

model Specify the model file (*.nnp) to calculate statistics.
To calculate statistics based on the training result selected in the Evaluation tab, use the default results.nnp.
output Specify the name of the CSV file to output the statistics to.
If Parameter Stats is executed from the evaluation results of the Evaluation tab, a table summarizing the statistics is saved to the specified file in the training result folder.