Models: object detection: how to train ssd from scratch use my own base network?

[gfphoenix]

Hi, all:
I want to make some model changes from ssd/mobilenet, but when I do this, the training process crashes. The error message tells me that the tensors with same names between the pre-trained model and new model are incompatible. So I can't use the training method as the tutorial recommends.
Now, I want to train the model from scratch. My machine info is:

• CPU : i7-6700
• Memory : 16GB
• GPU : GTX 1060 6GB
• OS : ubuntu 16.04
• TensorFlow : 1.4.0
• Python : 3.5
  any help is appreciated !!!

[pkulzc]

Could you please do the followings:

1. upgrade to tf 1.5 if possible
2. sync to HEAD and retry
3. report the cmd to reproduce your crash, and error message

so I can take a look and help understand the problem.

[v-prgmr]

How to train from scratch?

[pkulzc]

@vishal-noob If you just want to train from scratch, simple removing fine_tune_checkpoint field in config file.

[pkulzc]

I'm closing this issue, feel free to re-open it if you have more questions.