train/valid probs

[danpovey]

Guys,
We should have a mechanism to compute either accuracies on a subset of training data, or objective function values on validation data. (Are we doing this already?)
This will show whether our model is underfitting or overfitting, which right now I have no idea about.

[YiwenShaoStephen]

train.py will give you the BCE loss on train and val data. If you want to further see the segmentation results on train and val, you need to run segment.py on them on using the scoring.py to get the MAP (mean average precision)

[hhadian]

But, the segmentation algorithm also produces a final logprob for each image. I guess it would be helpful to write that to disk too (and maybe also its average on a whole test set).

…

On Thu, May 31, 2018 at 1:13 AM, Yiwen Shao ***@***.***> wrote: train.py will give you the BCE loss on train and val data. If you want to further see the segmentation results on train and val, you need to run segment.py on them on using the scoring.py to get the MAP (mean average precision) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#63 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AOW_Dbdi-Y_cr-Vm47LL9WtkK5MfHxjSks5t3wR-gaJpZM4UT-V4> .

[danpovey]

OK. How different are the BCE losses for train and val, in the DBS2018 setup?

…

On Wed, May 30, 2018 at 4:43 PM, Yiwen Shao ***@***.***> wrote: train.py will give you the BCE loss on train and val data. If you want to further see the segmentation results on train and val, you need to run segment.py on them on using the scoring.py to get the MAP (mean average precision) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#63 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu5XIU0VkVwOg0GXryVC-QhfIKi59ks5t3wR9gaJpZM4UT-V4> .

[YiwenShaoStephen]

Almost the same. So there is no overfitting yet. I think heavier image augmentation will help since the dataset is relatively small.

[danpovey]

Hossein, that's a good point. Perhaps we could produce a log line per file, summarizing various stats relating to the segmentation. I assume the stderr (and probably stdout too) of the segmenter code gets put in a log file. Yiwen, regarding the train/valid objective values: I think you can safely increase the number of parameters in the model until you see overfitting, and at that point start to worry about image augmentation. (I see image augmentation as primarily a way to reduce overfitting by artificially expanding the amount of training data). On Wed, May 30, 2018 at 4:47 PM, Hossein Hadian <[email protected]> wrote:

…

But, the segmentation algorithm also produces a final logprob for each image. I guess it would be helpful to write that to disk too (and maybe also its average on a whole test set). On Thu, May 31, 2018 at 1:13 AM, Yiwen Shao ***@***.***> wrote: > train.py will give you the BCE loss on train and val data. If you want to > further see the segmentation results on train and val, you need to run > segment.py on them on using the scoring.py to get the MAP (mean average > precision) > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#63 (comment)>, or mute > the thread > <https://github.com/notifications/unsubscribe-auth/AOW_Dbdi-Y_cr- Vm47LL9WtkK5MfHxjSks5t3wR-gaJpZM4UT-V4> > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#63 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu5oR8Iiz68AIcf4jPMfOxKF9ReIPks5t3wVNgaJpZM4UT-V4> .

[YiwenShaoStephen]

OK, will try it.

[hhadian]

Will do it. On Thu, May 31, 2018 at 1:32 AM, Daniel Povey <[email protected]> wrote:

…

Hossein, that's a good point. Perhaps we could produce a log line per file, summarizing various stats relating to the segmentation. I assume the stderr (and probably stdout too) of the segmenter code gets put in a log file. Yiwen, regarding the train/valid objective values: I think you can safely increase the number of parameters in the model until you see overfitting, and at that point start to worry about image augmentation. (I see image augmentation as primarily a way to reduce overfitting by artificially expanding the amount of training data). On Wed, May 30, 2018 at 4:47 PM, Hossein Hadian ***@***.***> wrote: > But, the segmentation algorithm also produces a final logprob for each > image. I guess it would be helpful to write that to disk too (and maybe > also its average on a whole test set). > > > On Thu, May 31, 2018 at 1:13 AM, Yiwen Shao ***@***.***> > wrote: > > > train.py will give you the BCE loss on train and val data. If you want to > > further see the segmentation results on train and val, you need to run > > segment.py on them on using the scoring.py to get the MAP (mean average > > precision) > > > > — > > You are receiving this because you are subscribed to this thread. > > Reply to this email directly, view it on GitHub > > <#63 (comment)>, > or mute > > the thread > > <https://github.com/notifications/unsubscribe-auth/AOW_Dbdi-Y_cr- > Vm47LL9WtkK5MfHxjSks5t3wR-gaJpZM4UT-V4> > > . > > > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#63 (comment)>, or mute > the thread > <https://github.com/notifications/unsubscribe-auth/ ADJVu5oR8Iiz68AIcf4jPMfOxKF9ReIPks5t3wVNgaJpZM4UT-V4> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#63 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AOW_Dcnq6lpxkGFVByLE2p1hwBfLdKxtks5t3wjmgaJpZM4UT-V4> .