Fine-tuning a network trained on ImageNet about ssd-tensorflow

[0rland0]

Did you already unzipped ssd_300_vgg.ckpt.zip which is located in the checkpoint folder?

[chenweiqian]

@0rland0 yes,I did that.

[0rland0]

@chenweiqian Can you maybe post the output of

ls ./checkpoints/

[chenweiqian]

@0rland0 yes, I see ssd_300_vgg.ckpt.index in it. And I also tried to change vgg_16.ckpt to ssd_300_vgg.ckpt and keep rest commands frozen, it can run successfully. But when I delete the "checkpoint_exclude_scopes " command , it gets error. So my successful case is the combining of "Fine-tuning existing SSD checkpoints" and "Fine-tuning a network trained on ImageNet". Is this a success?

[balancap]

Sorry for the problems, the script is buggy after some changes I have made :/
I'll fix this as soon as possible.

[chenweiqian]

@balancap thank you！^_^

[yyjFish]

I have met the same problem

[balancap]

I finally found some time to fix the bug. It should work well now!
I added to the training script a parameter checkpoint_model_scope where you can specify the main scope of the checkpoint file.
I also fixed the list of ignored scope for VGG16. It's a bit long! I could change the variable scopes in the net definition to make that simpler, but I don't want to introduce any modification not backward compatible with existing checkpoints of SSD-VGG.

[chenweiqian]

@balancap thank you for your work!