Can we pass different set of image to teacher and student.
Opened this issue · 6 comments
Hello,
Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Changing config file can't achieve it, you need to modify code
Teacher inference code is here:
LD/mmdet/models/detectors/kd_one_stage.py
Line 71 in bc60bbc
Hello, Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Hello, I also want to change the input of teacher network and student network into two datasets, may I ask you to achieve it?
Hello, Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Hello, I also want to change the input of teacher network and student network into two datasets, may I ask you to achieve it?
Do you mean training teacher and training student using different dataset or in the student training stage, pass different picture to teacher?
Hello, Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Hello, I also want to change the input of teacher network and student network into two datasets, may I ask you to achieve it?
Do you mean training teacher and training student using different dataset or in the student training stage, pass different picture to teacher?
In the distillation phase, teachers and students make predictions with different datasets, and then the predicted results are distilled, aligned between the two data sets.
Hello, Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Hello, I also want to change the input of teacher network and student network into two datasets, may I ask you to achieve it?
Do you mean training teacher and training student using different dataset or in the student training stage, pass different picture to teacher?
In the distillation phase, teachers and students make predictions with different datasets, and then the predicted results are distilled, aligned between the two data sets.
Same to my previous comment, teacher inference is here:
LD/mmdet/models/detectors/kd_one_stage.py
Line 71 in 0d575bf
you need to modify input of the forward_train function to the teacher dataset, which means you need add image preprocessing somewhere else.
Hello, Is it is possible to pass High resolution RGB image to Teacher network and low resolution RGB image to student network, so that student can learn to detect object from low resolution image. If possible what changes do I need to make in the config file. Thank you.
Hello, I also want to change the input of teacher network and student network into two datasets, may I ask you to achieve it?
Do you mean training teacher and training student using different dataset or in the student training stage, pass different picture to teacher?
In the distillation phase, teachers and students make predictions with different datasets, and then the predicted results are distilled, aligned between the two data sets.
Same to my previous comment, teacher inference is here:
LD/mmdet/models/detectors/kd_one_stage.py
Line 71 in 0d575bf
you need to modify input of the forward_train function to the teacher dataset, which means you need add image preprocessing somewhere else.
I get it, but the framework of mmdet is complicated, the underlying calls are difficult to trace, so I have not changed this code. I see that the questioner's problem is similar to mine, so I want to see if he/she has achieved it. In addition, if the high resolution image is distilled with the low resolution image, the teacher's and student's predictions are not aligned, what method should be used to align the predictions of the high resolution image to the low resolution image?