The meaning of the figure/graph when Training

[ALMachineLearning]

Hi Fausto,

I am quite new to this Caffe framework. I have a few questions regarding the training: 1) What does the graph mean when you train the model? At first, I assume it is the cost/loss function, but it is increasing. 2) I saw 2 numbers printed on the terminal on every iteration when you train the model. When it is stable (~30000 iterations), it shows number such as 0.97211 and 0.98123. Is this the training and validation error?

Thank you very much.

[faustomilletari]

It is the dice. Depending on batch size, in this implementation can go up to batch size.

I suggest you to use the tensorflow implementation done in niftynet. Google niftynet. Their code is as correct as this one here and it's tensorflow!

Fausto Milletarì Sent from my iPhone

On 1. Oct 2017, at 18:51, ALMachineLearning notifications@github.com wrote:

Hi Fausto,

I am quite new to this Caffe framework. I have a few questions regarding the training:

What does the graph mean when you train the model? At first, I assume it is the cost/loss function, but it is increasing. I saw 2 numbers printed on the terminal on every iteration when you train the model. When it is stable (~30000 iterations), it shows number such as 0.97211 and 0.98123. Is this the training and validation error? Thank you very much.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[ALMachineLearning]

Hi Fausto,

Thank you so much for the quick response.

I see, that is the dice, it makes much more sense now. Silly me for missing that. One more question though: Is this graph showing a training performance OR validation performance? Because I notice once it is stable, it never get worse. I assume this is training, not validation. Usually validation performance will reach the peak and then slowly getting worse due to overfitting.

Thank you

[faustomilletari]

Training only as far as I recall.

Fausto Milletarì Sent from my iPhone

On 1. Oct 2017, at 22:09, ALMachineLearning notifications@github.com wrote:

Hi Fausto,

Thank you so much for the quick response.

I see, that is the dice, it makes much more sense now. Silly me for missing that. One more question though: Is this graph showing a training performance OR validation performance? Because I notice once it is stable, it never get worse. I assume this is training, not validation. Usually validation performance will reach the peak and then slowly getting worse due to overfitting.

Thank you

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ALMachineLearning]

Hi Fausto,

Ooo, if that is the case, there is a possibility of it being overfitted isn't it? Since there is no validation test to check that.

By the way, thank you for the suggestion on the Tensorflow framework, I will have a look.

Thank you very much.

[faustomilletari]

Yes, overfitting might happen. Hopefully it's not a huge problem due to the augmentation. The prostate dataset is very small and it was decided not to split it further into train and validation.

Fausto Milletarì Sent from my iPhone

On 1. Oct 2017, at 22:49, ALMachineLearning notifications@github.com wrote:

Hi Fausto,

Ooo, if that is the case, there is a possibility of it being overfitted isn't it? Since there is no validation test to check that.

By the way, thank you for the suggestion on the Tensorflow framework, I will have a look.

Thank you very much.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ALMachineLearning]

Hi Fausto,

ooo, that is good to hear.

That is true that MRI data is still very limited currently. I hope we can have more data available in the future.

Anyway, thank you so much for your quick and helpful response. I hope this information can also be useful for the others.

Kind Regards.

[faustomilletari]

It will surely be useful in the future.

Please don't forget to close the issue.

Regards,

Fausto Milletarì Sent from my iPhone

On 1. Oct 2017, at 23:41, ALMachineLearning notifications@github.com wrote:

Hi Fausto,

ooo, that is good to hear.

That is true that MRI data is still very limited currently. I hope we can have more data available in the future.

Anyway, thank you so much for your quick and helpful response. I hope this information can also be useful for the others.

Kind Regards.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[ALMachineLearning]

Hi Fausto,

My apology for forgetting to close the issue. Thank you for your assistance. :)

Kind Regards