Data augmentation #7
Conversation
benchmark_utils/transformation.py
Outdated
| @@ -6,15 +6,14 @@ | |||
| # - getting requirements info when all dependencies are not installed. | |||
| with safe_import_context() as import_ctx: | |||
| import numpy as np | |||
|
|
|||
| from numpy import concatenate | |||
There was a problem hiding this comment.
prefer using np.concatenate instead of importing the function, as this makes it easier to read the code.
benchmark_utils/transformation.py
Outdated
| X_augm = to_numpy(X) | ||
| y_augm = y | ||
| for i in range(n_augmentation): | ||
| transform = ChannelsDropout(probability=probability) |
There was a problem hiding this comment.
why not instantiate this object outside the loop?
Also, a seed should be given otherwise the benchmark is not reproducible.
So not sure why you made this change, can you comment?
There was a problem hiding this comment.
I put it inside the loop because when it was outside the augmentation was always giving the same X_tr
There was a problem hiding this comment.
and I have the same issue when I fix the seed
There was a problem hiding this comment.
Maybe we could predefine a sequence a seed, to have a different seed for each transformation, do you have an other idea ?
There was a problem hiding this comment.
yes, having a sequence of seeds is indeed a nice idea.
But I would look at the Transform object and try to understand what is happening, because the standard for data augmentation would be that repeated calls to the transform should give different augmented data.
Can you do simple example and check what is the behavior of the object?
There was a problem hiding this comment.
Okay I think I know what is the exact issue, what defines the transformation is mask_len_samples and mask_start_per_sample. They are generated "randomly" in transform.get_augmentation_params, the issue is that to choose randomly, the transformation uses rng.uniform with a seed that is fixed. So we are always having the same parameters for the augmentation and we get at each iteration the same augmented data.
benchmark_utils/transformation.py
Outdated
| probability=probability, | ||
| mask_len_samples=mask_len_samples, | ||
| ) |
| def channels_dropout( | ||
| X, y, n_augmentation, seed=0, probability=0.5, p_drop=0.2 | ||
| X, y, n_augmentation, probability=0.5, p_drop=0.2 |
There was a problem hiding this comment.
| X, y, n_augmentation, probability=0.5, p_drop=0.2 | |
| X, y, n_augmentation, probability=0.5, p_drop=0.2, seed=None |
| @@ -43,51 +47,58 @@ def channels_dropout( | |||
| The labels. | |||
|
|
|||
| """ | |||
| transform = ChannelsDropout(probability=probability, random_state=seed) | |||
|
|
|||
| seed = gen_seed() | |||
There was a problem hiding this comment.
| seed = gen_seed() | |
| rng = np.random.RandomState(seed) |
There was a problem hiding this comment.
For a reproducible sequence of random numbers, the best is to use a random number generator, that will be used to sample different number for each draw, but with a reproducible order.
| transform = ChannelsDropout( | ||
| probability=probability, | ||
| random_state=next(seed) | ||
| ) |
There was a problem hiding this comment.
| transform = ChannelsDropout( | |
| probability=probability, | |
| random_state=next(seed) | |
| ) | |
| transform = ChannelsDropout( | |
| probability=probability, | |
| random_state=rng | |
| ) |
|
|
||
| return X_augm, y_augm | ||
|
|
||
|
|
||
| def smooth_timemask( | ||
| X, y, n_augmentation, sfreq, seed=0, probability=0.5, second=0.1 | ||
| X, y, n_augmentation, sfreq, probability=0.8, second=0.2 |
There was a problem hiding this comment.
| X, y, n_augmentation, sfreq, probability=0.8, second=0.2 | |
| X, y, n_augmentation, sfreq, probability=0.8, second=0.2, seed=None |
| @@ -13,7 +13,7 @@ | |||
| SmoothTimeMask, | |||
| ) | |||
| from braindecode.models import ShallowFBCSPNet | |||
| from numpy import linspace, pi | |||
| from numpy import linspace | |||
There was a problem hiding this comment.
| from numpy import linspace | |
| import numpy as np |
| @@ -96,25 +96,25 @@ def set_objective(self, X, y, sfreq): | |||
| mask_len_samples=int(sfreq * second), | |||
| random_state=seed, | |||
| ) | |||
| for second in linspace(0.1, 2, 10) | |||
| for second in linspace(0.1, 2, 3) | |||
There was a problem hiding this comment.
| for second in linspace(0.1, 2, 3) | |
| for second in np.linspace(0.1, 2, 3) |
There was a problem hiding this comment.
Why did you decrease the number of samples?
There was a problem hiding this comment.
It was only to have faster results
| "augmentation": [ | ||
| "SmoothTimeMask", | ||
| ], | ||
| "covariances_estimator": ["oas"], |
There was a problem hiding this comment.
| "augmentation": [ | |
| "SmoothTimeMask", | |
| ], | |
| "covariances_estimator": ["oas"], | |
| "covariances_estimator": ["oas"], |
| ] | ||
|
|
||
| elif self.augmentation == "ChannelDropout": | ||
| transforms = [ | ||
| ChannelsDropout( | ||
| probability=self.proba, p_drop=prob, random_state=seed | ||
| ) | ||
| for prob in linspace(0, 1, 10) | ||
| for prob in linspace(0, 1, 3) |
There was a problem hiding this comment.
| for prob in linspace(0, 1, 3) | |
| for prob in np.linspace(0, 1, 3) |
| random_state=seed, | ||
| ) | ||
| for prob in linspace(0, 2 * pi, 10) | ||
| for phase_freq in linspace(0, 1, 3) |
There was a problem hiding this comment.
| for phase_freq in linspace(0, 1, 3) | |
| for phase_freq in np.linspace(0, 1, 3) |
Co-authored-by: Thomas Moreau <[email protected]>
The augmentation process is now working as we were expecting and we improved the way we compute the scores