Hi, we verified there’s a problem with the latest firmware with CNN stuff. We will do another release with a fix for this and a fix for another issue with set windowing a customer found.
Regarding CNN issues. We talked to Pete Warden recently and tensor flow light may be coming soon. If this happens we will move away from this custom stuff and have tensor flow support.
Hi, thank you for the verification! Do you know how long until the release? I’m asking because I have a demo on Wednesday and would like to keep an eye open to have my openMV updated earlier.
The tensor flow support would be amazing!
Thank you,
Nicholas
Um, well Ibrahim fixed the issue with the file not being able to be opened. As for the accuracy I can’t say how long its going to fix that. Getting you the binary for the latest firmware is pretty simple.
Okay no problem, if you can please let me know when you have the bin file ready!
Alternatively, how is the performance on the H7? Did that fall too? If not, I can use my H7 temporary until the M7 firmware is ready.
Thanks,
Nicholas
Um, well, I’ve only tested the H7 with my cifar10 net and it works great as long as the image is binarized.
I’m sorry about this not being the best right now. However, with tensor lite now available we should be able to get this to some level of spec where these issues go away. The plan was to start main streaming this more with the H7.
Okay great! Thanks for the information!
Definitely looking forward in the updates and new features!
Ibrahim will release v3.3.1 today which fixes the file loading issue.
Great! I just installed it!
Please keep me informed about the fix for the CNN processing
Thank you,
Nicholas
Ibrahim is on it. This is the only thing on his plate right now.
@Eric I’ve tested your network using the first 1000 images in val_data on the camera, and the accuracy is very low indeed. Then I checked the first 100 images in val_data and I don’t think the network will learn anything useful from these images (some images are mislabeled, others contain multiple persons, heads, hands, hidden people etc…). This may explain the low accuracy, I think if you filter the dataset you will get much better results.