A Filter Pruning Method of CNN Models Based on Feature Maps Clustering

Round 1

Reviewer 1 Report

This is a good topic, quite popular right now.

 

There is work to do on the presentation side.

English needs to be revised  (like in line #12).

It is not clear what the results are. Defining research questions (and then answers) would help.

The conclusions are basically a paragraph long recap.

Overall, the paper needs a structure.

Author Response

We gratefully thank you for your time spend on making your constructive remarks and useful suggestions which have significantly raised the quality of the manuscript and have enable us to improve the manuscript. According to the remarks we made the following revision.

• We add some relate network compression and acceleration algorithms in the related part the added parts are tilted.
• We refine the figure format and add some other references.

 

 

Reviewer 2 Report

This paper proposes a filter pruning method based on feature maps clustering to accelerate and compress CNN. K-means and HCA methods were selected and compared regarding accuracy and speed. Silhouette Coefficient Method and Elbow were used to determine the number of clusters. My major concerns are:

1. This paper looks like a survey rather than new research, as the authors spend many sections to explain the basic ideas like weight pruning, quantization, k-means, PCA, etc., which is already known to the society. 
2. This work targeting limited-resource computation like mobile devices and embedded platforms, but the experiment using Titan X GPU and i7 CPU is not aligned with that. In addition, the survey has not covered recent research regarding distributed computing. References [1] and [2] are encouraged to add to the introduction part. 
3. The abstract part mentioned "self driving cars", but none of the following sections touched that. The experiment on  WIDER FACE dataset has no relation to self driving cars. Suggest to add reference [3] as an example of object tracking in self driving cars. If possible, there could be one experiment on self driving car dataset. 

Minor concerns are:

1. Figure and its caption need to be centered
2. line 113, page 3, what is the meaning of n?
3. line 381, "Hence, P5 is considered as the most important layer and selected to be pruned by our method firstly." Reader may get confused why P3/P4 are not pruned first as they have low impact on the accuracy. 

[1] Distributed mean-field-type filters for Big Data assimilation, in the second IEEE International Conference on Data Science and Systems (HPCC-SmartCity-DSS), Sydney, Australia, Dec, 2016, pp. 1446-1453.

[2] Correlative Mean-Field Filter for Sequential and Spatial Data Processing, in the Proceedings of IEEE International Conference on Computer as a Tool (EUROCON), Ohrid, Macedonia, July 2017

[3] "Distributed Mean-Field-Type Filters for Traffic Networks," in IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 2, pp. 507-521, Feb. 2019. doi:10.1109/TITS.2018.2816811

Author Response

We gratefully thank you for your time spend on making your constructive remarks and useful suggestions which have significantly raised the quality of the manuscript and have enable us to improve the manuscript. According to the remarks we made the following revision.

• Reference [1][2][3] are added in the introduction part as [1][8][9]
• Figures and captain are adjusted to the center.
• In the revised version, add the meaning of n. It refers to the size of the filter.
• In the revised version in order to make it clear that we want to the find the most important layer related to the output so that we can test the proposed algorithm on this layer, we add “In order to measure their effects on the performance of neural networks this paper prunes all filters namely reduced related layer in three convolutional layers one by one and evaluate them on the validation set.” And delete the firstly in the “"Hence, P5 is considered as the most important layer and selected to be pruned by our method firstly."”

Reviewer 3 Report

The paper proposes a filter pruning method of CNN neural networks based on the feature maps clustering to compress the neural models. K-Means and Hierarchical Clustering algorithms were used and a comparison between them was made. Conducted experiments showed that the hierarchical clustering algorithm can be the effective method for filter pruning and silhouette coefficient method can be used to determine the number of pruned filters.

In my opinion, the article is interesting and is suitable for publication in the journal Applied Sciences. However, I would ask the Authors to correct and supplement the following issues:

• In the “Related Work” section please provide appropriate references for the given network compression and acceleration algorithms. Please also clearly indicate in the text their advantages and disadvantages.
• The same applies to 4.1 section (K–means method). Please provide an appropriate reference.
• In line 230 it should be “…A, B, C are clustered to one group, D, E are…” instead of “…A, B, C are clustered to one group, C, D are…”. Please correct it.
• What is the example in Figure 6 about? The Cost Function J takes different values depending on the issue under consideration. Please add an appropriate comment to this example in the text.
• The same applies to the caption of Figure 6. It is no “general relationship” between Cost Function and Cluster Number, but it is only an example for use of the Elbow method (see the Figure 7, for use of SC Method). Please correct it.
• In section 4.2 – HCA Method, please provide relevant literature references to the bottom-up and top-down hierarchical clustering algorithms.
• The captions in Figure 10 (“scale”, “pose” and so on) are illegible. Please correct them.
• In Figure 15, please add a description of the x axis (cluster number). Please also correct the Figure caption. It is better to write “The Relationship between silhouette coefficient and cluster number” instead of “Silhouette coefficient-clustering number curve”.
• In line 442 it should be: “As is shown in Figure 17 (a) and Figure 17 (c)” instead of “As is shown in Figure 17 (a) and Figure 17 (b)”. The same applies to the line 445. It should be: “…HCA and K-Means methods is shown in Figure 17 (b) and Figure 17 (d)” instead of “…HCA and K-Means methods is shown in Figure 17 (c) and Figure 17 (d). Alternatively, Authors may change the order of the graphs in Figure 17 as described in the text.
• There is no description of the x and y axes in Figure 17. What do the individual bars of the chart refer to? Please comment in the text.
• Some of the graphs are in the rough format, probable copied from excel. To improve the readability of the paper, the authors need to further work on their graphical details.

After responding to the above comments, I can recommend the article for publication.

Author Response

We gratefully thank you for your time spend on making your constructive remarks and useful suggestions which have significantly raised the quality of the manuscript and have enable us to improve the manuscript. According to the remarks we made the following revision.

• We provide appropriate references for the given network compression and acceleration algorithms in the related part the added parts are tilted.
• We add relative reference to the k-means method
• We changed it to “A, B, C are clustered to one group, D, E” in the revised version as titled show.
• We add some comment on figure 6 and the caption is changed to Cost Function vs. Cluster Numbers
• We add relative reference to the HCA method.
• We delete the above captions (“scale”, “pose” and so on) in Figure 10.
• In Figure 15, we add a description of the x axis (cluster number). Figure caption is changed to “The Relationship between silhouette coefficient and cluster number”
• We changed as “As is shown in Figure 17 (a) and Figure 17 (c)” && “…HCA and K-Means methods is shown in Figure 17 (b) and Figure 17 (d)”
• We add caption in the figure 17 as the comment. 

Reviewer 4 Report

In my opinion, the article entitled "A Filter Pruning Method of CNN Models Based on Feature Maps Clustering" is interesting. However, the manuscript might need some revisions to enhance the quality of the article. Some points I have indicated throughout the manuscript need to be upgraded. Details of my comments are in the attached file. Good luck to the authors.

Comments for author File: Comments.pdf

Author Response

Thank you for your comments on our manuscript. Those comments are very helpful for revising and improving our paper as well as the important guiding significance to other research. We have studied the comments carefully and made corrections which we hope meet with approval. The modification can be seen in the attached cover letter.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

Thank you for the excellent work. Although you have increased the quality of some graphs, I think you need to improve the quality of the rest.  And if you can provide some numeric results and future needs of your work in the abstract, it would be more attractive to the readers. 

Author Response

Thank you for your time.

• We improve the quality of all the graphs.
• We add some future needs of our work in the abstract.