Where to get weights?

[dexter2406]

Hi, thanks for your work. But where could I got the weights file, e.g. saved_model_0.19968511 ? I don't find link in your description.

Thanks for your time!

[anshkumar]

Currently, I'm not providing pre-trained checkpoints. Will add coco pre-trained checkpoints in future.

[Ankur-singh]

any updates on it? At least can you tell how to train the model & generate the weights.

[anshkumar]

@Ankur-singh For training you have to just change the path in train.sh file. And for creating tfrecord you have to use create_coco_tfrecord.sh file.

[Ankur-singh]

I think, one will have to download the coco dataset and unzip it before creating tfrecords. Because, I cannot find the code to download the dataset. Please correct me if am wrong or if I am missing something.

[anshkumar]

Yes, you have to download the dataset.

[shashank-sharma333]

@Ankur-singh Were you able to train the model?
I am getting the below error while running train.sh with the required parameters

tensorflow.python.framework.errors_impl.InvalidArgumentError: buffer_size must be greater than zero.

[Ankur-singh]

@shashank-sharma333 yes I was able to train the model. I had to make quite a few changes to make it work. But in the process, I never came across this error. 😅

[shashank-sharma333]

Thank you @Ankur-singh for your comment. Which environment you are using for the training model? Also, Can you list down the changes you made? It would be helpful.

[Ankur-singh]

I have a 3060Ti, so I followed GPU installation instruction from the readme file. Its pretty straight forward. Just make sure your compile the protobufs. Just run this from the root directory of the project

protoc protos/*.proto --python_out=.

Regarding the changes, TBH, I don't remember all of them. It was a couple of weeks ago. One was regarding SGD optimizer, the API has changed, so you should use legacy version. And there were quite a few data type issue, you should be able to fix them with basic type casting. Hope this helps.