deploy: braindecode/braindecode@725f67b

Author: bruAristimunha

dev/.buildinfo

@@ -1,4 +1,4 @@
 # Sphinx build info version 1
 # This file records the configuration used when building these files. When it is not found, a full rebuild will be done.
-config: fb11c39ac11c25e5dc07edcd12266e18
+config: 3dc9db1f0f951d39f882ba727f0a7312
 tags: 645f666f9bcd5a90fca523b33c5a78b7

dev/CNAME

@@ -0,0 +1 @@
+braindecode.org

dev/_downloads/07fcc19ba03226cd3d83d4e40ec44385/auto_examples_python.zip

@@ -67,7 +67,7 @@
 # ~~~~~~~~~~~~~~~
 #
 # We extract 30-s windows to be used in the classification task.
-# The :class:`braindecode.models.SleepStagerEldele2021` model takes a
+# The :class:`braindecode.models.AttnSleep` model takes a
 # single channel as input. Here, the Fpz-Cz channel is used as it
 # was found to give better performance than using the Pz-Oz channel
 
@@ -190,7 +190,7 @@ def get_center_label(x):
 import torch
 from torch import nn
 
-from braindecode.models import SleepStagerEldele2021
+from braindecode.models import AttnSleep
 from braindecode.modules import TimeDistributed
 from braindecode.util import set_random_seeds
 
@@ -205,8 +205,8 @@ def get_center_label(x):
 # Extract number of channels and time steps from dataset
 n_channels, input_size_samples = train_set[0][0].shape
 
-feat_extractor = SleepStagerEldele2021(
-    sfreq,
+feat_extractor = AttnSleep(
+    sfreq=sfreq,
     n_outputs=n_classes,
     n_times=input_size_samples,
     return_feats=True,

dev/_downloads/1f2407fc0459906c808c847bdcd84ccf/plot_sleep_staging_eldele2021.py

@@ -11,7 +11,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "Building on the `Trialwise Decoding <bcic-iv-2a-moabb-trial>`,\nwe now do more data-efficient cropped decoding!\n\nIn Braindecode, there are two supported configurations created for\ntraining models: trialwise decoding and cropped decoding. We will\nexplain this visually by comparing trialwise to cropped decoding.\n\n<img src=\"file://_static/trialwise_explanation.png\" alt=\"Trialwise decoding\">\n<img src=\"file://_static/cropped_explanation.png\" alt=\"Cropped decoding\">\n\nOn the left, you see trialwise decoding:\n\n1. A complete trial is pushed through the network.\n2. The network produces a prediction.\n3. The prediction is compared to the target (label) for that trial to\n   compute the loss.\n\nOn the right, you see cropped decoding:\n\n1. Instead of a complete trial, crops are pushed through the network.\n2. For computational efficiency, multiple neighbouring crops are pushed\n   through the network simultaneously (these neighbouring crops are\n   called compute windows)\n3. Therefore, the network produces multiple predictions (one per crop in\n   the window)\n4. The individual crop predictions are *AVERAGED* before computing the\n   loss function\n\nThis averaging of predictions of small sub-windows is the key difference\nbetween trialwise and cropped decoding. It was introduced in [1]_ and it impact\non the parameters of the network.\n\nIt is important to note that the averaging of predictions is only done\nduring training. During testing, the network is still applied to crops\nand the predictions are averaged afterwards.\n\n\n<div class=\"alert alert-info\"><h4>Note</h4><p>-  The network architecture implicitly defines the crop size (it is the\n       receptive field size, i.e., the number of timesteps the network uses\n       to make a single prediction)\n    -  The window size is a user-defined hyperparameter, called\n       ``n_times`` in Braindecode. It mostly affects runtime\n       (larger window sizes should be faster). As a rule of thumb, you can\n       set it to two times the crop size.\n    -  Crop size and window size together define how many predictions the\n       network makes per window: ``#window \u2212 #crop + 1 = #predictions``</p></div>\n\n<div class=\"alert alert-info\"><h4>Note</h4><p>For cropped decoding, the above training setup is mathematically\n    similar to sampling crops in your dataset, pushing them through the\n    network and training directly on the individual crops. However, the\n    if their position would be randomly selected, the crops would be less\n    correlated in contrast to the neighbourhood crops selected from a window.\n    At the same time, the above training setup is much faster as it avoids\n    redundant computations by using dilated convolutions, see [2]_.\n    However, the two setups are only mathematically related in case (1)\n    your network does not use any padding or only left padding and\n    (2) your loss function leads\n    to the same gradients when using the averaged output. The first is true\n    for our shallow and deep ConvNet models and the second is true for the\n    log-softmax outputs and negative log likelihood loss that is typically\n    used for classification in PyTorch.</p></div>\n\n"
+        "Building on the `Trialwise Decoding <bcic-iv-2a-moabb-trial>`,\nwe now do more data-efficient cropped decoding!\n\nIn Braindecode, there are two supported configurations created for\ntraining models: trialwise decoding and cropped decoding. We will\nexplain this visually by comparing trialwise to cropped decoding.\n\n<img src=\"file://_static/trialwise_explanation.png\" alt=\"Trialwise decoding\">\n<img src=\"file://_static/cropped_explanation.png\" alt=\"Cropped decoding\">\n\nOn the left, you see trialwise decoding:\n\n1. A complete trial is pushed through the network.\n2. The network produces a prediction.\n3. The prediction is compared to the target (label) for that trial to\n   compute the loss.\n\nOn the right, you see cropped decoding:\n\n1. Instead of a complete trial, crops are pushed through the network.\n2. For computational efficiency, multiple neighbouring crops are pushed\n   through the network simultaneously (these neighbouring crops are\n   called compute windows)\n3. Therefore, the network produces multiple predictions (one per crop in\n   the window)\n4. The individual crop predictions are *AVERAGED* before computing the\n   loss function\n\nThis averaging of predictions of small sub-windows is the key difference\nbetween trialwise and cropped decoding. It was introduced in [1]_ and it impact\non the parameters of the network.\n\nIt is important to note that the averaging of predictions is only done\nduring training. During testing, the network is still applied to crops\nand the predictions are averaged afterwards.\n\n\n<div class=\"alert alert-info\"><h4>Note</h4><p>-  The network architecture implicitly defines the crop size (it is the\n       receptive field size, i.e., the number of timesteps the network uses\n       to make a single prediction)\n    -  The window size is a user-defined hyperparameter, called\n       ``n_times`` in Braindecode. It mostly affects runtime\n       (larger window sizes should be faster). As a rule of thumb, you can\n       set it to two times the crop size.\n    -  Crop size and window size together define how many predictions the\n       network makes per window: ``#window - #crop + 1 = #predictions``</p></div>\n\n<div class=\"alert alert-info\"><h4>Note</h4><p>For cropped decoding, the above training setup is mathematically\n    similar to sampling crops in your dataset, pushing them through the\n    network and training directly on the individual crops. However, the\n    if their position would be randomly selected, the crops would be less\n    correlated in contrast to the neighbourhood crops selected from a window.\n    At the same time, the above training setup is much faster as it avoids\n    redundant computations by using dilated convolutions, see [2]_.\n    However, the two setups are only mathematically related in case (1)\n    your network does not use any padding or only left padding and\n    (2) your loss function leads\n    to the same gradients when using the averaged output. The first is true\n    for our shallow and deep ConvNet models and the second is true for the\n    log-softmax outputs and negative log likelihood loss that is typically\n    used for classification in PyTorch.</p></div>\n\n"
       ]
     },
     {