Feat/loss soft hkr #86
Conversation
deel/lip/losses.py
Outdated
| Note that `y_true` should be one-hot encoded or pre-processed with the | ||
| `deel.lip.utils.process_labels_for_multi_gpu()` function. | ||
|
|
||
| Using a multi-GPU/TPU strategy requires to set `multi_gpu` to True and to | ||
| pre-process the labels `y_true` with the | ||
| `deel.lip.utils.process_labels_for_multi_gpu()` function. |
There was a problem hiding this comment.
If multi-GPU is not supported, we might remove/update these lines.
There was a problem hiding this comment.
comment on multi-gpu removed
deel/lip/losses.py
Outdated
| (self.min_margin_v,), | ||
| dtype=tf.float32, | ||
| constraint=lambda x: tf.clip_by_value(x, 0.005, 1000), | ||
| name="moving_mean", |
There was a problem hiding this comment.
Maybe change the name of this variable to avoid confusion with self.moving_mean which has the same name
deel/lip/losses.py
Outdated
| if self.one_hot_ytrue: | ||
| y_true = tf.where(y_true > 0, 1, -1) # switch to +/-1 |
There was a problem hiding this comment.
Is it possible to use the same line as hinge_preproc? sign = tf.where(y_true > 0, 1, -1) is done whenever the y_true is 0/1 or -1/1.
If yes, it is then possible to remove self.one_hot_ytrue from the arguments of the function.
It is also possible to preprocess the y_true labels only once in hkr() function, instead of twice in preproc_kr and preproc_hinge. Like for F_soft_KR computed once at the beginning of hkr()
Co-authored-by: cofri <[email protected]>
…inibatches is not possible with this loss
| temperature (float): factor for softmax temperature | ||
| (higher value increases the weight of the highest non y_true logits) | ||
| alpha_mean (float): geometric mean factor |
There was a problem hiding this comment.
Docstrings of temperature and alpha_mean are not in the same order than in the __init__() signature
| name="current_mean", | ||
| ) | ||
|
|
||
| self.temperature = temperature * self.min_margin_v |
There was a problem hiding this comment.
I'm wondering if it's a problem when saving with get_config() if the temperature is not the same as initialization of the loss (i.e. keeping temperature * self.min_margin_v instead of temperature alone in get_config)
There was a problem hiding this comment.
self.temperature has to be divided by self.min_margin_v before saving. The only variable is self.current_mean but the question of serialization of the value is good
remove useless docstring Co-authored-by: cofri <[email protected]>
Introduce the multiclass loss : MulticlassSoftHKR defined in the paper "On the explainable properties of 1-Lipschitz Neural Networks: An Optimal Transport Perspective" Serrurier et al (Neurips'23) that combine an optimal transport loss with a softmax temperature coefficient (useful when the numer of classes is high like in Imagenet dataset)