Tuning of Parameters

Hi, I am currently trying to train both yolov4 and ssd tlt model (ver 3.0) for my own use cases. I just wanted to double confirm that for the output_height and output_width parameter under augmentation_config, we have to specify those parameter with the image resolution that were used to train the model am i right? Much thanks in advance :)

It is the Input size of the model you want to train.

So if i were to train a yolov4 model, the output_height and output_width parameters should always remain the same at 384 and 1248 respectively?

No, any value you want to train.
But make sure meet the requirement of https://docs.nvidia.com/tao/tao-toolkit/text/object_detection/yolo_v4.html#input-requirement

oh i see, so the bigger the value for those two parameters, means the bigger size of the model that i am using to train?

Also, does this mean that for frcnn, the height and width parameters under the size_height_width segment is different from the output_image_width and output_image_height parameters under the augmentation config segment or can they be the same?

Yes.

Refer to FasterRCNN — TAO Toolkit 3.22.05 documentation

sorry but what i was trying to ask is that must both sets of parameters (height, width parameters under the size_height_width segment and the output_image_width, output_image_height parameters under the augmentation config segment) must be the same? i did take a look at the documentation and i do note that in the example they gave is that both sets of parameters are the same

Please see the details in FasterRCNN — TAO Toolkit 3.22.05 documentation

Oh, so if i specify the values for output_height and output_width parameters to be 384 and 1248 respectively (for example), the image resolution will be resize to those values as well. May i ask does this process happen for all three models (frcnn, yolov4 and ssd) or only for frcnn?

Yolov4 and SSD can also support resizing images during training.

Just to be sure, do i have to specify a parameter like how it does for frcnn ( By setting enable_auto_resize in augmentation_config to True we will enable automatic resize during training.) or does the models (yolov4 and ssd) help to auto resize the image to the values specify in output_height and output_width parameters respectively?

Yes, yolov4 and ssd will resize images during training.
The augmented images will be resized to the output shape defined in augmentation_config .

1 Like

ok thank you for your help

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.