[models] Create specific model's heads for binary classification #19
Assignees
Labels
ext: references
Related to references
help wanted
Extra attention is needed
module: models
Related to models
Milestone
Comments
frgfm
added
help wanted
frgfm
added a commit
that referenced
this issue
Oct 16, 2019
* feat: Added fastai training script Model's head still has to be switched to binary classification and ImageDataBunch needs to be built while respecting the existing split of train/validation. See #3 * feat: Updated fastai training script Fixed the split of train and valid set so that the ImageDataBunch takes the existing split. See #3 * feat: Added training script without fastai See #3 * feat: Added deterministic mode to torch_train See #3 * style: Added endfile lines * fix: Fixed training script train/valid split For users that did not have the validation split, the training was failing when instantiating OpenFire in non training mode. * refactor: Reorganized reference scripts Separated fastai and non-fastai scripts to be able to have different requirements.txt for cleaner usage. * docs: Updated readme Added instructions for training using references scripts. * refactor: Removed unused imports and variables * fix: Fixed fastai training script Missing pandas dependency * fix: Fixed scheduler step on training scripts * fix: Reflected script argument in scheduler * refactor: Removed unused variable * feat: Updated torch training script Changed evaluation and logging to have similar metrics to fastai script. See #3 * chore: Updated version Switched from alpha to beta. * refactor: Harmonize default data path between scripts * chore: Updated training script requirements OneCycleLR is only available since torch 1.3.0. * style: Renamed learner * feat: Added device setting to fastai training * refactor: Unified device resolution for training scripts Closes #3 See #19, #24
blenzi
pushed a commit
to blenzi/PyroNear
that referenced
this issue
Oct 21, 2019
* feat: Added fastai training script Model's head still has to be switched to binary classification and ImageDataBunch needs to be built while respecting the existing split of train/validation. See pyronear#3 * feat: Updated fastai training script Fixed the split of train and valid set so that the ImageDataBunch takes the existing split. See pyronear#3 * feat: Added training script without fastai See pyronear#3 * feat: Added deterministic mode to torch_train See pyronear#3 * style: Added endfile lines * fix: Fixed training script train/valid split For users that did not have the validation split, the training was failing when instantiating OpenFire in non training mode. * refactor: Reorganized reference scripts Separated fastai and non-fastai scripts to be able to have different requirements.txt for cleaner usage. * docs: Updated readme Added instructions for training using references scripts. * refactor: Removed unused imports and variables * fix: Fixed fastai training script Missing pandas dependency * fix: Fixed scheduler step on training scripts * fix: Reflected script argument in scheduler * refactor: Removed unused variable * feat: Updated torch training script Changed evaluation and logging to have similar metrics to fastai script. See pyronear#3 * chore: Updated version Switched from alpha to beta. * refactor: Harmonize default data path between scripts * chore: Updated training script requirements OneCycleLR is only available since torch 1.3.0. * style: Renamed learner * feat: Added device setting to fastai training * refactor: Unified device resolution for training scripts Closes pyronear#3 See pyronear#19, pyronear#24
frgfm
added a commit
that referenced
this issue
Oct 22, 2019
frgfm
added a commit
that referenced
this issue
Oct 23, 2019
* fix: Fixed layer freezing Pure pytorch reference script had the model fully unfrozen by default which would not be comparable to the one using fastai. See #3 * feat: Added options for model unfreezing * style: Corrected whitespaces and indentation * refactor: Switched to TTA Test Time Augmentation is now similar to training augmentation. * feat: Added resume option on references * style: Clarified argument description * fix: Moved saving message display Saving message should only be displayed if the checkpoint is actually being created. * docs: Updated docstrings * feat: Added binary classification option See #19 * style: Fixed pep8 compliance * fix: Fixed argument description * fix: Fixed training script saving message
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
🚀 Feature
Create a
models.utilssubmodule with a function to easily create model heads depending on the type of task, namely binary classification here.Motivation
Current training scripts in
referencesassume that the target is categorical (multi-label) rather than binary. A proper task setting would allow the model to better optimize during training.Pitch
The function would take as arguments:
and return a
torch.nn.Sequentialobject.The text was updated successfully, but these errors were encountered: