Changes in the toolbox with lasted caffe

[x1155665]

@arikpoz Thank you very much for making the toolbox functional with the new version of caffe 👍

I've seen that you have made many changes in terms of usages, which you didn't mentioned in the README.md.
The two main changes I‘ve found:

1. The local settings are removed to the folder 'model_settings"
2. The locations of the pre-computed jpgs are moved to the caffevis_outputs_dir. And the storage structure is also changed. This results in a necessary preprocess of the jpgs that fetch.sh downloads.

Are there any other changes that are worth noticing in terms of using the toolbox?

I hope you can mention the usage changes in the README, and provide a solution for the second change I mentioned above, so that the toolbox can be fully functional out-of-box.

[arikpoz]

There are plenty of new visualization and many other improvements to the tool, I haven't had the time to update the README, but you can see a "short" summary in the pull request I've created for pushing it all to the original branch: #128

Regarding the tool to relocate the jpgs, I'm not sure I follow. My version of the tool knows how to load both the previous type of outputs and the updated outputs. Admittedly, I haven't tested the old outputs for a while and it may require an update. In any case, your fix should probably only change the parts around caffevis_outputs_dir_folder_format == 'original_combined_single_image', and not caffevis_outputs_dir_folder_format == 'max_tracker_output' which is the new format.

[x1155665]

Thanks for pointing me to the summary.

Sorry, I didn't know the existence of caffevis_outputs_dir_folder_format when opening this issue. But since

patterns view with maximal input images is not implemented when settings.caffevis_outputs_dir_folder_format == 'original_combined_single_image'"

, my fix can be a workaround for the user to load the caffenet-yos without having much trouble.

[arikpoz]

I've updated the README file with some examples on what's new

[arikpoz]

Regarding your fix, it seems it will break the loading of new outputs

[x1155665]

Thanks for the updates!

Sorry, I did not concerned about the new outputs. my intent was simply to load the images that fetch.sh downloads.
How does my fix break the loading of new output? By changing 'maxim*.png' to 'maxim*.jpg' and 'deconv*.png' to 'deconv*.jpg'? I did that just because the format of the downloaded caffenet-yos images is jpg. I can write a new fix later, in which the format will be changed back to png.

[x1155665]

The fixed fix is now in the pull request

[ZhuweiQin]

Hi, @x1155665 I encountered some troubles when I try to show the per unit optimize visualization by using the new toolbox. For example, I want to see some unit in the conv1 of CaffeNet. The terminal says "loading opt*.jpg images for layer conv1 channel 90 out of 96", and I can see nothing on the top right corner of this toolbox. I wonder how to solve this problem, and any suggestions will be greatly appreciated. By the way, can I use this toolbox to see the per unit optimize visualization of the ResNet?

[x1155665]

Hi, @ZhuweiQin The per unit optimize visualization of CaffeNet-yos is downloaded by fetch.sh . Since this version of DeepVis toolbox has a lot of changes, you may need to relocate the jpgs so that the toolbox can find them. This tool can help you do that.
You can use it on ResNet, this will help you.

[ZhuweiQin]

Thanks for your kind reminder. @x1155665 So, you mean, I should synthesize the per unit optimize visualization of ResNet by using this tool and then change the code to load the pre-synthesized jpgs, Right?

[x1155665]

@ZhuweiQin Yes and no. The produced images will be placed in the right order, you don't need to relocate them. The relocating tool is just for the downloaded images (CaffeNet-yos).

[ZhuweiQin]

@x1155665 Thanks a lot. Let me try.