Is there any documentation list of the downloadable pre-trained weight file name with description for PRETRAIN_WEIGHTS_FILE?


I found the link that describes the PRETRAIN_WEIGHTS_FILE environment variable in environment.json here

I’m not sure there is any documentation for pre-train weights file names with the description of the model weights that could download via PRETRAIN_WEIGHTS_FILE setting or not?

Thank you.


Could you clarify your ask ? it seems you are trying to finetune a model, either starting from our NVIDIA models or you trained your network and would like to continue training. For this you should run the where you point it to the model to load by modifying the MMAR_CKPT=<your/path/model.ckpt

Hope that helps

No, I am not trying to finetune a model. I am sorry that I gave the wrong section in the URL.
I want to train the model from scratch with pre-trained weight like ImageNet weight.
I found that there is an environment variable named PRETRAIN_WEIGHTS_FILE in environment.json.
This variable described in the document as

Location of the pre-trained weights file. NOTE: if the file does not exist and is needed, the training program will download it from predefined URL from the web.

, So I would like to know if there is any list of model weights that I can use as the initial weight of the model.

Thank you.

Oh I got it now. So the training from scratch can be done by starting from random weights --> simple enough you just remove this variable usage in the model section of the train_config.json
starting with image net weights. --> we only have support for image net weights used by the chest Xray model or AHnet segmentation model used for spleen and liver segmentation tasks. We are not providing other models as this requires changing the network architecture

Hope that answers your question


If I understand correctly, there are only 2 pre-trained ImageNet weights

  1. chest x-ray using cxr_pretrain_weights.h5
  2. spleen and liver segmentation using resnet50_weights_tf_dim_ordering_tf_kernels.h5


You are correct