Skip to content

Latest commit

 

History

History
86 lines (58 loc) · 4.32 KB

SmoothGrad.rst

File metadata and controls

86 lines (58 loc) · 4.32 KB
Raw

XAI/SmoothGrad

Using a method called SmoothGrad, the areas of the input image that affect the classification result are made visible in the model, which performs image classification.

Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, Martin Wattenberg. "Smoothgrad: removing noise by adding noise". Workshop on Visualization for Deep Learning, ICML, 2017.

Input Information

model Specify the model file (*.nnp) that will be used in the SmoothGrad computation. To perform SmoothGrad based on the training result selected in the Evaluation tab, use the default results.nnp.
noise_level Specify the noise level (0.0 to 1.0) to calculate standard deviation used for generating gausian noise.
num_samples Specify the number of times of gradient calculation to get the final averaged sensitivity map. Gausian noise is repeatedly generated and put onto the input image to calculate gradient of each time.
image Specify the image file to analyze. To perform SmoothGrad 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 SmoothGrad is executed from the evaluation results of the Evaluation tab, the visualization results are saved to the specified file in the training result folder.

Output Information

The result of this plugin is saved in the designated 'output' path as PNG file. It is shown in grayscale image as sensitivity map.

XAI/SmoothGrad (batch)

Using a method called SmoothGrad, the areas of the input image that affect the classification result are made visible in the model, which performs image classification.

Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, Martin Wattenberg. "Smoothgrad: removing noise by adding noise". Workshop on Visualization for Deep Learning, ICML, 2017.

Input Information

model Specify the model file (*.nnp) that will be used in the SmoothGrad computation. To perform SmoothGrad based on the training result selected in the Evaluation tab, use the default results.nnp.
noise_level Specify the noise level (0.0 to 1.0) to calculate standard deviation used for generating gausian noise.
num_samples Specify the number of times of gradient calculation to get the final averaged sensitivity map. Gausian noise is repeatedly generated and put onto the input image to calculate gradient of each time.
input Specify the dataset CSV file containing image files to be processed by SmoothGrad. To process SmoothGrad on images in the Output Result shown in the Evaluation tab, use the default output_result.csv.
output Specify the name of the image file to output the visualization results to. If SmoothGrad is executed from the evaluation results of the Evaluation tab, the visualization results are saved to the specified file in the training result folder.
input_variable Of the variables included in the dataset CSV file specified in input, specify the name of the variable to be used in the SmoothGrad computation.
label_variable Of the variables included in the dataset CSV file specified in input, specify the variable name of the class index to be visualized.

Output Information

The result of this plugin is saved in the designated 'output' path as CSV file. The information on the columns of CSV file is as follows. The other columns than listed below are the same meaning as those in output_result.csv file that is generated as a result of evaluation.

SmoothGrad The result path of this plugin for the target instance, the image of which is displayed in NNC window. It is shown in grayscale image as sensitivity map.