Log Softmax as False (#624)

* Softmax as False * Updating the examples * Updating the tests * updating benchmark * changing models again * Updating the whats_new file * reverting the modification
braindecode · May 31, 2024 · ffff87c · ffff87c
1 parent f96265b
commit ffff87c
Show file tree

Hide file tree

Showing 24 changed files with 33 additions and 29 deletions.
diff --git a/braindecode/models/atcnet.py b/braindecode/models/atcnet.py
@@ -127,7 +127,7 @@ def __init__(
         n_channels=None,
         n_classes=None,
         input_size_s=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, input_window_seconds = deprecated_args(
             self,

diff --git a/braindecode/models/deep4.py b/braindecode/models/deep4.py
@@ -120,7 +120,7 @@ def __init__(
         in_chans=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, n_times = deprecated_args(
             self,

diff --git a/braindecode/models/eegconformer.py b/braindecode/models/eegconformer.py
@@ -106,7 +106,7 @@ def __init__(
         n_classes=None,
         n_channels=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_outputs, n_chans, n_times = deprecated_args(
             self,
@@ -395,7 +395,11 @@ def forward(self, x):
 
 class _FinalLayer(nn.Module):
     def __init__(
-        self, n_classes, hidden_channels=32, return_features=False, add_log_softmax=True
+        self,
+        n_classes,
+        hidden_channels=32,
+        return_features=False,
+        add_log_softmax=False,
     ):
         """Classification head for the transformer encoder.
 

diff --git a/braindecode/models/eeginception.py b/braindecode/models/eeginception.py
@@ -108,7 +108,7 @@ def __init__(
         in_channels=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         (
             n_chans,

diff --git a/braindecode/models/eeginception_erp.py b/braindecode/models/eeginception_erp.py
@@ -111,7 +111,7 @@ def __init__(
         in_channels=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         (
             n_chans,

diff --git a/braindecode/models/eeginception_mi.py b/braindecode/models/eeginception_mi.py
@@ -76,7 +76,7 @@ def __init__(
         in_channels=None,
         n_classes=None,
         input_window_s=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         (
             n_chans,

diff --git a/braindecode/models/eegitnet.py b/braindecode/models/eegitnet.py
@@ -130,7 +130,7 @@ def __init__(
         n_classes=None,
         in_channels=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         (
             n_outputs,

diff --git a/braindecode/models/eegnet.py b/braindecode/models/eegnet.py
@@ -264,7 +264,7 @@ def __init__(
         in_chans=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, n_times = deprecated_args(
             self,

diff --git a/braindecode/models/eegresnet.py b/braindecode/models/eegresnet.py
@@ -61,7 +61,7 @@ def __init__(
         in_chans=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, n_times = deprecated_args(
             self,

diff --git a/braindecode/models/hybrid.py b/braindecode/models/hybrid.py
@@ -36,7 +36,7 @@ def __init__(
         in_chans=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
         input_window_seconds=None,
         sfreq=None,
         chs_info=None,

diff --git a/braindecode/models/shallow_fbcsp.py b/braindecode/models/shallow_fbcsp.py
@@ -88,7 +88,7 @@ def __init__(
         in_chans=None,
         n_classes=None,
         input_window_samples=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, n_times = deprecated_args(
             self,

diff --git a/braindecode/models/sleep_stager_blanco_2020.py b/braindecode/models/sleep_stager_blanco_2020.py
@@ -64,7 +64,7 @@ def __init__(
         n_channels=None,
         n_classes=None,
         input_size_s=None,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         (
             n_chans,

diff --git a/braindecode/models/tidnet.py b/braindecode/models/tidnet.py
@@ -315,7 +315,7 @@ def __init__(
         temp_span=0.05,
         bottleneck=3,
         summary=-1,
-        add_log_softmax=True,
+        add_log_softmax=False,
     ):
         n_chans, n_outputs, n_times = deprecated_args(
             self,

diff --git a/docs/whats_new.rst b/docs/whats_new.rst
@@ -60,7 +60,7 @@ Bugs
 API changes
 ~~~~~~~~~~~
 - Expose the ``use_mne_epochs parameter`` of :function:`braindecode.preprocessing.create_windows_from_events` (:gh:`607` by `Pierre Guetschel`_)
-
+- Parameter ``use_log_softmax`` is default as `False` for all the models in (:gh:`624` by `Bruno Aristimunha`_)
 
 .. _changes_0_8_0:
 Current 0.8 (11-2022)

diff --git a/examples/advanced_training/plot_data_augmentation.py b/examples/advanced_training/plot_data_augmentation.py
@@ -244,7 +244,7 @@ def plot_psd(data, axis, label, color):
     model,
     iterator_train=AugmentedDataLoader,  # This tells EEGClassifier to use a custom DataLoader
     iterator_train__transforms=transforms,  # This sets the augmentations to use
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=predefined_split(valid_set),  # using valid_set for validation
     optimizer__lr=lr,

diff --git a/examples/advanced_training/plot_data_augmentation_search.py b/examples/advanced_training/plot_data_augmentation_search.py
@@ -253,7 +253,7 @@
     model,
     iterator_train=AugmentedDataLoader,  # This tells EEGClassifier to use a custom DataLoader
     iterator_train__transforms=[],  # This sets is handled by GridSearchCV
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=None,  # GridSearchCV will control the split and train/validation over the dataset
     optimizer__lr=lr,

diff --git a/examples/datasets_io/benchmark_lazy_eager_loading.py b/examples/datasets_io/benchmark_lazy_eager_loading.py
@@ -201,7 +201,7 @@ def create_example_model(
         model.cuda()
 
     optimizer = optim.Adam(model.parameters())
-    loss = nn.NLLLoss()
+    loss = nn.CrossEntropyLoss()
 
     return model, loss, optimizer
 

diff --git a/examples/model_building/plot_bcic_iv_2a_moabb_cropped.py b/examples/model_building/plot_bcic_iv_2a_moabb_cropped.py
@@ -277,7 +277,7 @@
     model,
     cropped=True,
     criterion=CroppedLoss,
-    criterion__loss_function=torch.nn.functional.nll_loss,
+    criterion__loss_function=torch.nn.functional.cross_entropy,
     optimizer=torch.optim.AdamW,
     train_split=predefined_split(valid_set),
     optimizer__lr=lr,

diff --git a/examples/model_building/plot_bcic_iv_2a_moabb_trial.py b/examples/model_building/plot_bcic_iv_2a_moabb_trial.py
@@ -238,7 +238,7 @@
 
 clf = EEGClassifier(
     model,
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=predefined_split(valid_set),  # using valid_set for validation
     optimizer__lr=lr,

diff --git a/examples/model_building/plot_how_train_test_and_tune.py b/examples/model_building/plot_how_train_test_and_tune.py
@@ -271,7 +271,7 @@
 
 clf = EEGClassifier(
     model,
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=None,
     optimizer__lr=lr,
@@ -387,7 +387,7 @@ def plot_simple_train_test(ax, all_dataset, train_set, test_set):
 
 clf = EEGClassifier(
     model,
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=predefined_split(val_subset),
     optimizer__lr=lr,
@@ -503,7 +503,7 @@ def plot_train_valid_test(ax, all_dataset, train_subset, val_subset, test_set):
 
 clf = EEGClassifier(
     model,
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     train_split=None,
     optimizer__lr=lr,

diff --git a/examples/model_building/plot_hyperparameter_tuning_with_scikit-learn.py b/examples/model_building/plot_hyperparameter_tuning_with_scikit-learn.py
@@ -263,7 +263,7 @@
 
 clf = EEGClassifier(
     model,
-    criterion=torch.nn.NLLLoss,
+    criterion=torch.nn.CrossEntropyLoss,
     optimizer=torch.optim.AdamW,
     optimizer__lr=[],  # This will be handled by GridSearchCV
     batch_size=batch_size,

diff --git a/examples/model_building/plot_train_in_pure_pytorch_and_pytorch_lightning.py b/examples/model_building/plot_train_in_pure_pytorch_and_pytorch_lightning.py
@@ -332,7 +332,7 @@ def test_model(dataloader: DataLoader, model: Module, loss_fn, print_batch_stats
 # Define the loss function
 # We used the NNLoss function, which expects log probabilities as input
 # (which is the case for our model output)
-loss_fn = torch.nn.NLLLoss()
+loss_fn = torch.nn.CrossEntropyLoss()
 
 # train_set and test_set are instances of torch Datasets, and can seamlessly be
 # wrapped in data loaders.
@@ -380,7 +380,7 @@ class LitModule(L.LightningModule):
     def __init__(self, module):
         super().__init__()
         self.module = module
-        self.loss = torch.nn.NLLLoss()
+        self.loss = torch.nn.CrossEntropyLoss()
 
     def training_step(self, batch, batch_idx):
         x, y, _ = batch

diff --git a/test/acceptance_tests/test_trialwise_decoding.py b/test/acceptance_tests/test_trialwise_decoding.py
@@ -103,7 +103,7 @@ def test_trialwise_decoding():
     clf = EEGClassifier(
         model,
         cropped=False,
-        criterion=torch.nn.NLLLoss,
+        criterion=torch.nn.CrossEntropyLoss,
         optimizer=torch.optim.Adam,
         train_split=train_valid_split,
         optimizer__lr=0.001,

diff --git a/test/unit_tests/training/test_scoring.py b/test/unit_tests/training/test_scoring.py
@@ -318,7 +318,7 @@ def on_epoch_end(self, net, *args, **kwargs):
 
     clf = EEGClassifier(
         model,
-        criterion=torch.nn.NLLLoss,
+        criterion=torch.nn.CrossEntropyLoss,
         optimizer=optim.AdamW,
         train_split=None,
         optimizer__lr=0.0625 * 0.01,
@@ -465,7 +465,7 @@ def test_predict_trials():
     # cropped EEGClassifier, cropped data
     clf = EEGClassifier(
         model,
-        criterion=torch.nn.NLLLoss,
+        criterion=torch.nn.CrossEntropyLoss,
         optimizer=optim.AdamW,
         train_split=None,
         optimizer__lr=0.0625 * 0.01,