MPavg and MPmin

[huanmofazong]

Hello, I would like to know which part of the code I should modify if I want to reproduce MPavg and MPmin?

[Openning07]

Primarily, you should change the data directory of AVA and make sure the environment is prepared to reproduce the claimed results in my paper. Please read the Note and one example code for details. I also upload my checkpoint files in ** MyCheckpoint** folder.

[huanmofazong]

@Openning07 Thank you for your reply. I configured the environment and changed the path of the dataset to run your code. But it should be MPada. I would like to ask which part of the code should be modified if I want to reproduce MPavg and MPmin.

[Openning07]

The core difference is the weights_JensenFactor in function compute_loss_and_error. If you want to change MPada to other two modes, please change the implementation of weights_JensenFactor: for MPavg, weights_JensenFactor is a vector [1, 1, ..., 1]; for MPmin, weights_JensenFactor is an one-hot vector where 1 indicates the bit of minimum confidence.

[huanmofazong]

@Openning07 Thank you for your reply. I would like to ask if the value of --JensenFactor should be set to 0.5 during training?

[Openning07]

Yes. I used 0.5 for JensenFactor when MPada mode is selected.

[huanmofazong]

@Openning07 For MPmin,you said weights_JensenFactor is an one-hot vector where 1 indicates the bit of minimum confidence.May I ask which Tensor I should use, I try to use target_p，But that doesn't seem right. `weights_JensenFactor_old = tf.zeros_like(target_p, dtype='float32')

weight_index = tf.argmin(target_p, axis=0)

index = tf.cast(weight_index, dtype='int32')

weight_a = weights_JensenFactor_old[:index]

weight_b = tf.constant([1.], dtype='float32')

weight_c = weights_JensenFactor_old[index+1:]

weights_JensenFactor = tf.concat([weight_a, weight_b, weight_c], axis=0)`

Is this the right weights_JensenFactor?Thank you for your reply.