search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.53k stars 11.68k forks source link

Why all the tutorial of model training will use trained model? #2723

Open leekwunfung817 opened 2 years ago

leekwunfung817 commented 2 years ago

Why do all the tutorials of model training will use trained models? Theoretically, it will be affected by previous training data to cause it to be less accurate.

1: How could I train a model without using a trained model? 2: Why not many people ask to train a pure new model for more accuracy? 3: Is that better?

tbnxyzfz1 commented 2 years ago

You can comment out the load weight codes to train a new model. I tried this... after 80 epochs, the result is still not good. With a trained model, I can easily get a good model within 40 epochs.

leekwunfung817 commented 2 years ago

You can comment out the load weight codes to train a new model. I tried this... after 80 epochs, the result is still not good. With a trained model, I can easily get a good model within 40 epochs.

I want to inference new objects which totally do not exist in the trained model, did you train with the same set of objects?

tbnxyzfz1 commented 2 years ago

You can comment out the load weight codes to train a new model. I tried this... after 80 epochs, the result is still not good. With a trained model, I can easily get a good model within 40 epochs.

I want to inference new objects which totally do not exist in the trained model, did you train with the same set of objects?

I used my own dataset. But I only have about 300 images.

leekwunfung817 commented 2 years ago

You can comment out the load weight codes to train a new model. I tried this... after 80 epochs, the result is still not good. With a trained model, I can easily get a good model within 40 epochs.

I want to inference new objects which totally do not exist in the trained model, did you train with the same set of objects?

I used my own dataset. But I only have about 300 images.

Do those data contain the previous existing classes in the trained model?

nataliameira commented 2 years ago

@leekwunfung817

A good deep learning training uses thousands of images to learn its characteristics. Therefore, the pre-trained model has already learned many features, such as: borders, textures, colors, among others. So when you train your model in the last layers you won't need thousands of images. Hundreds of images will be enough.

tbnxyzfz1 commented 2 years ago

You can comment out the load weight codes to train a new model. I tried this... after 80 epochs, the result is still not good. With a trained model, I can easily get a good model within 40 epochs.

I want to inference new objects which totally do not exist in the trained model, did you train with the same set of objects?

I used my own dataset. But I only have about 300 images.

Do those data contain the previous existing classes in the trained model?

Not at all