What is the difference between using a pre-train-weight file and not using a pre-train-weight file?


I evaluate the accuracy of models generated with and without pretrain weights.

The results tend to be more versatile with pretrain weights.
I understand the complex structure of the model, but I can’t explain why it’s better.
Could you tell me the difference?

Thank you and best regards,

Thanks for your interest in Clara train SDK
Could you provide more info regarding which model you are referring to. Also did you use NVIDIA’s model to fine tune your for your data

I created segmentation models using non-contrast CT images and built from scratch without using NVIDIA models.

To understand AH-NET, I studied model accuracy differences with the same parameters except for setting pre-training weight, but when I re-evaluated model accuracy after I wrote this topic, the model with the pre-training weights did not tend to be more versatile. I’m sorry I jumped to conclusions.

I might think it’s a basic question,Are there any indicators for using pretrain weights? For example, it is effective for organs with low density, such as the lungs. etc.
Also , The pre-train I am using now is “resnet50_weights_tf_dim_ordering_tf_kernels.h5”, but do you have a pre-train-weight-file suitable for organ segmentation?

Best regards,