[Triplet Margin Loss] Issue 1118 #1120
Conversation
|
@vroulet May I know if there's anything that needs to be changed? |
optax/losses/_self_supervised.py
Outdated
| anchor: The anchor embeddings. Shape: [batch_size, feature_dim]. | ||
| positive: The positive embeddings. Shape: [batch_size, feature_dim]. | ||
| negative: The negative embeddings. Shape: [batch_size, feature_dim]. | ||
| margin: The margin value. Default: 1.0. |
There was a problem hiding this comment.
No need to put the default values since they are given in the signature.
optax/losses/_self_supervised.py
Outdated
| by V. Balntas et al. Default: False. | ||
| reduction: Specifies the reduction to apply to the output: | ||
| 'none' | 'mean' | 'sum'. Default: 'mean'. | ||
|
|
optax/losses/_self_supervised.py
Outdated
| margin: The margin value. Default: 1.0. | ||
| p: The norm degree for pairwise distance. Default: 2. | ||
| eps: Small epsilon value to avoid numerical issues. Default: 1e-6. | ||
| swap: Use the distance swap optimization from "Learning shallow |
There was a problem hiding this comment.
Use rst formatting for references (see e.g. the docstring of Adam)
optax/losses/_self_supervised.py
Outdated
| swap: bool = False, | ||
| reduction: str = 'mean', | ||
| ) -> chex.Array: | ||
| """Triplet margin loss function. |
| @@ -53,5 +53,41 @@ def test_batched(self): | |||
| ) | |||
|
|
|||
|
|
|||
| class TripletMarginLossTest(chex.TestCase): | |||
|
|
|||
| def setUp(self): | |||
There was a problem hiding this comment.
Avoid using numerical values as expected returns.
They may fail depending on the backend for example.
You may consider simple test cases with a "handmade" function (see e.g. the lbfgs tests). You can check for specific inputs (like zeros or ones).
You may also add a test for some specific behaviors (like using swap here).
Also you should test this function under jit/vmap etc... (see the chex.all_variant utility in some other tests).
|
@vroulet we have worked on your suggestion and all the tests are passing. I think the code is ready to be merged. |
|
@vroulet |
|
@vroulet we tried multiple things to solve the error in pipeline, the tests for triplet_loss are passing locally. The errors we are getting here seems to be not from the function we implemented. Can you guide us on this? |
|
Hello @cvnad1 , @Saanidhyavats , |
|
@vroulet We have modified the code based on your review. Could you please verify if everything's correct? |
| anchors: chex.Array, | ||
| positives: chex.Array, | ||
| negatives: chex.Array, | ||
| axis: chex.Numeric = -1, |
| positives: chex.Array, | ||
| negatives: chex.Array, | ||
| axis: chex.Numeric = -1, | ||
| p: chex.Numeric = 2, |
There was a problem hiding this comment.
norm_degree rather than p.
Moreover we have
||x||_p = (sum_i |x_i|^p)**(1/p)
not
||x||_p = sqrt(sum_i x_i^p)
You may want to include the case ||x||_inf in a separate PR?
| >>> Array([0.14142442, 0.14142442], dtype=float32) | ||
|
|
||
| Args: | ||
| anchors: An array of anchor embeddings, with shape [batch, feature_dim]. |
There was a problem hiding this comment.
Add indents appropriately like that:
anchors: An array of anchor embeddings, with shape [batch, feature_dim].
positives: An array of positive embeddings
(similar to anchors), with shape [batch, feature_dim].
negatives: An array of negative embeddings
(dissimilar to anchors), with shape [batch, feature_dim].
axis: The axis along which to compute the distances
(default is -1).
p: The norm degree for distance calculation
(default is 2 for Euclidean distance).
margin: The minimum margin by which the positive distance
should be smaller than the negative distance.
eps: A small epsilon value to ensure numerical stability
in the distance calculation.
reduction: Specifies the reduction to apply to the
output: 'none' | 'mean' | 'sum'.
| If reduction is 'mean' or 'sum', returns a scalar. | ||
|
|
||
| References: | ||
| Learning shallow convolutional feature descriptors with triplet losses |
There was a problem hiding this comment.
Use the following formatting for references
References:
V. Balntas et al, `Learning shallow convolutional feature descriptors with triplet losses
<https://bmva-archive.org.uk/bmvc/2016/papers/paper119/abstract119.pdf>`_, 2016
| by V. Balntas, E. Riba et al. | ||
| <https://bmva-archive.org.uk/bmvc/2016/papers/paper119/abstract119.pdf> | ||
| """ | ||
| chex.assert_type([anchors], float) |
There was a problem hiding this comment.
Remove the three chex.assert_type(...).
| p: chex.Numeric = 2, | ||
| margin: chex.Numeric = 1.0, | ||
| eps: chex.Numeric = 1e-6, | ||
| reduction: str = 'none', |
There was a problem hiding this comment.
Remove the reduction option. No losses in optax reduce the losses after computation.
So better to keep this loss follow the same principle.
The user can take care of the reduction easily after computing the losses.
| negative_distance = jnp.sqrt(jnp.power(anchors - negatives, p).sum(axis) + eps | ||
| ) | ||
| loss = jnp.maximum(positive_distance - negative_distance + margin, 0) | ||
| if reduction == 'mean': |
There was a problem hiding this comment.
As said above, remove the reduction options.
@vroulet Hi Vincent, Added code and tests for the Triplet Margin Loss Function #1118 . Kindly review the code and please do comment in case of any changes.