Open madhavajay opened 6 years ago
I should add, demo.py works fine.
I found the solution here: https://github.com/chuanqi305/MobileNet-SSD/issues/9
force_color: true
I found this implementation to know where to put it in the file: https://github.com/intel/caffe/blob/master/models/intel_optimized_models/ssd/VGGNet/coco/SSD_300x300/train.prototxt
I noticed they also have:
mean_value: 104
mean_value: 117
mean_value: 123
I assume this is better for COCO than the default also they dont have a scale: value. @chuanqi305 Is this correct for COCO? If so I could add a flag to the script to generate a different training and testing file for coco.
I assume these changes should be made to MobileNetSSD_test.prototxt as well.
@madhavajay It works for me
20h Hi Team, I was trying to train COCO dataset with MobileNet-SSD. I followed the github
https://github.com/chuanqi305/MobileNet-SSD
First, I train the COCO dataset with SSD and I run 400,000 iterations. Then I took the 400,000 caffe model and run Mobilenet on it. I followed the same instructions mentioned at the Github page, but the loss never came down less than 5. I found the same issue for other uses also for the COCO dataset. Which is listed at https://github.com/chuanqi305/MobileNet-SSD/issues/95
I did the fine-tuning according to the instruction like
and 2. mean_value: 104 mean_value: 117 mean_value: 123
But I got no success. Is there anyone who got success training Mobilenet-SSD.
P.S I used my own dataset also and trained the data with SSD first and then MobileNet. The loss never came down below 6. I don’t know if there is a trick I am missing. Any help on this topic is greatly appreciated.
Thanks, Debo
I am trying to finetune the mobilenet_iter_73000.caffemodel with a subset of COCO. I ran the normal Caffe VGG retraining with this same dataset and it runs, but then I realised I want MobileNet so I switched to this model.
I get these errors and have no idea why:
Full output: