initial transforms code. HCP struct wandb configs

transforms.py

@@ -0,0 +1,207 @@
+import numpy as np
+import random
+
+
+class _BaseTransform:
+
+    def __init__(self, p):
+        """
+        p: Probability of applying transform.
+        """
+
+        assert 0 <= p <= 1
+        self.p = p
+
+
+    def apply(self, signals):
+
+        raise NotImplementedError
+
+
+    def __call__(self, signals):
+
+        if random.random() < self.p:
+
+            signals = self.apply(signals)
+
+        return signals
+
+
+class Compose:
+
+    def __init__(self, transforms, p=1):
+        """
+        transforms: List of transforms to apply to signals.
+        p: Probability of applying the list of transforms.
+        """
+
+        assert 0 <= p <= 1
+
+        self.transforms = transforms
+        self.p = p
+
+
+    def __call__(self, signals):
+
+        if random.random() < self.p:
+
+            for transform in self.transforms:
+
+                signals = transform(signals)
+
+        return signals
+
+
+class FlipTime(_BaseTransform):
+    """Randomly flip signals along temporal dimension"""
+
+    def __init__(self, p=0.5):
+        """
+        p: Probability of applying transform.
+        """
+
+        super().__init__(p)
+
+
+    def apply(self, signals):
+
+        if len(signals.shape) > 1:
+
+            signals = np.fliplr(signals)
+
+        else:
+
+            signals = np.flipud(signals)
+
+        return signals
+
+
+class MaskTime(_BaseTransform):
+    """Randomly mask signal"""
+
+    def __init__(self, min_fraction=0.0, max_fraction=0.5, p=0.5):
+        """
+        min_fraction: Minimum length of the mask as a fraction of the total time series length.
+        max_fraction: Maximum length of the mask as a fraction of the total time series length.
+        p: Probability of applying transform.
+        """
+
+        super().__init__(p)
+
+        assert 0 <= min_fraction <= 1
+        assert 0 <= max_fraction <= 1
+        assert max_fraction >= min_fraction
+
+        self.min_fraction = min_fraction
+        self.max_fraction = max_fraction
+
+
+    def apply(self, signals):
+
+        num_samples = signals.shape[-1]
+        length = random.randint(int(num_samples * self.min_fraction), int(num_samples * self.max_fraction))
+        start = random.randint(0, num_samples - length)
+
+        mask = np.zeros(length)
+        masked_signals = signals.copy()
+        masked_signals[..., start : start + length] *= mask
+
+        return masked_signals
+
+
+class Shift(_BaseTransform):
+    """Shift the signals forwards or backwards along the temporal dimension"""
+
+    def __init__(self, min_fraction=-0.5, max_fraction=0.5, rollover=True, p=0.5):
+        """
+        min_fraction: Fraction of total timeseries to shift.
+        max_fraction: Fraction of total timeseries to shift.
+        rollover: Samples that roll beyond the first or last position are re-introduced at the last or first otherwise set to zero.
+        p: Probability of applying this transform.
+        """
+
+        super().__init__(p)
+
+        assert min_fraction >= -1
+        assert max_fraction <= 1
+
+        self.min_fraction = min_fraction
+        self.max_fraction = max_fraction
+        self.rollover = rollover
+
+
+    def apply(self, signals):
+
+        num_samples = signals.shape[-1]
+
+        num_shift = int(round(random.uniform(self.min_fraction, self.max_fraction) * num_samples))
+
+        signals = np.roll(signals, num_shift, axis=-1)
+
+        if not self.rollover:
+
+            if num_shift > 0:
+                signals[..., :num_shift] = 0.0
+
+            elif num_shift < 0:
+                signals[..., num_shift:] = 0.0
+
+        return signals
+
+
+class FlipPolarity(_BaseTransform):
+    """Randomly flip sign of signal"""
+
+    def __init__(self, p=0.5):
+        """
+        p: Probability of applying transform.
+        """
+
+        super().__init__(p)
+
+
+    def apply(self, signals):
+
+        return -signals
+
+
+class GuassianNoise(_BaseTransform):
+    """Add gaussian noise to the signals"""
+
+    def __init__(self, min_amplitude=0.001, max_amplitude=0.015, p=0.5):
+        """
+        min_amplitude: minimum amplitude of noise.
+        max_amplitude: maximum amplitude of noise.
+        p: Probability of applying this transform.
+        """
+
+        super().__init__(p)
+
+        assert min_amplitude > 0.0
+        assert max_amplitude > 0.0
+        assert max_amplitude >= min_amplitude
+
+        self.min_amplitude = min_amplitude
+        self.max_amplitude = max_amplitude
+
+
+    def apply(self, signals):
+
+        amplitude = random.uniform(self.min_amplitude, self.max_amplitude)
+
+        noise = np.random.randn(*signals.shape).astype(np.float32)
+        signals = signals + amplitude * noise
+
+        return signals   
+
+
+if __name__ == "__main__":
+
+    # generate multivariate time series: 250 features, 1000 length
+    signals = np.random.uniform(low=-0.2, high=0.2, size=(250, 1000,)).astype(np.float32)
+    # compose a list of transforms
+    augment = Compose([FlipTime(p=0.5), FlipPolarity(p=0.5), GuassianNoise(p=0.5), Shift(p=0.5), MaskTime(p=0.5)])
+    # augment signals
+    augmented_signals = augment(signals)
+
+    assert not np.allclose(signals, augmented_signals)

wandb_sweeps/final/final_hcp_try1_N_struct_fold1.yaml

@@ -0,0 +1,89 @@
+method: random
+
+name: final_hcp_try1_N_struct_fold1
+program: main_loop.py
+parameters:
+  dataset_type:
+    value: hcp
+  fold_num:
+    value: 1
+  sweep_type:
+    value: node_meta
+  tcn_depth:
+    value: 3
+  tcn_norm_strategy:
+    value: batchnorm
+  tcn_kernel:
+    value: 7
+  tcn_hidden_units:
+    value: 8
+  tcn_final_transform_layers:
+    value: 1
+  temporal_embed_size:
+    value: 16
+  lr_scheduler:
+    value: plateau
+  optimiser:
+    value: rmsprop
+  use_ema:
+    value: false
+  edge_weights:
+    value: true
+  conn_type:
+    value: struct
+  batch_size:
+    value: 100
+  analysis_type:
+    value: st_multimodal
+  num_epochs:
+    value: 150
+  num_nodes:
+    value: 68
+  activation:
+    value: relu
+  channels_conv:
+    value: 8
+  conv_strategy:
+    value: tcn_entire
+  dropout:
+    distribution: uniform
+    max: 0.5
+    min: 0
+  early_stop_steps:
+    value: 33
+  encoding_strategy:
+    value: none
+  num_gnn_layers:
+    value: 1
+  threshold:
+    distribution: categorical
+    values:
+      - 5
+      - 10
+      - 20
+      - 30
+      - 40
+  lr:
+    distribution: log_uniform
+    max: -2.3025850929940455
+    min: -11.512925464970229
+  weight_decay:
+    distribution: log_uniform
+    max: -2.3025850929940455
+    min: -27.631021115928547
+  normalisation:
+    value: subject_norm
+  pooling:
+    value: concat
+  target_var:
+    value: gender
+  time_length:
+    value: 490
+  nodemodel_aggr:
+    value: all
+  nodemodel_scalers:
+    value: none
+  nodemodel_layers:
+    value: 3
+  final_mlp_layers:
+    value: 1

wandb_sweeps/final/final_hcp_try1_N_struct_fold2.yaml

@@ -0,0 +1,89 @@
+method: random
+
+name: final_hcp_try1_N_struct_fold2
+program: main_loop.py
+parameters:
+  dataset_type:
+    value: hcp
+  fold_num:
+    value: 2
+  sweep_type:
+    value: node_meta
+  tcn_depth:
+    value: 3
+  tcn_norm_strategy:
+    value: batchnorm
+  tcn_kernel:
+    value: 7
+  tcn_hidden_units:
+    value: 8
+  tcn_final_transform_layers:
+    value: 1
+  temporal_embed_size:
+    value: 16
+  lr_scheduler:
+    value: plateau
+  optimiser:
+    value: rmsprop
+  use_ema:
+    value: false
+  edge_weights:
+    value: true
+  conn_type:
+    value: struct
+  batch_size:
+    value: 100
+  analysis_type:
+    value: st_multimodal
+  num_epochs:
+    value: 150
+  num_nodes:
+    value: 68
+  activation:
+    value: relu
+  channels_conv:
+    value: 8
+  conv_strategy:
+    value: tcn_entire
+  dropout:
+    distribution: uniform
+    max: 0.5
+    min: 0
+  early_stop_steps:
+    value: 33
+  encoding_strategy:
+    value: none
+  num_gnn_layers:
+    value: 1
+  threshold:
+    distribution: categorical
+    values:
+      - 5
+      - 10
+      - 20
+      - 30
+      - 40
+  lr:
+    distribution: log_uniform
+    max: -2.3025850929940455
+    min: -11.512925464970229
+  weight_decay:
+    distribution: log_uniform
+    max: -2.3025850929940455
+    min: -27.631021115928547
+  normalisation:
+    value: subject_norm
+  pooling:
+    value: concat
+  target_var:
+    value: gender
+  time_length:
+    value: 490
+  nodemodel_aggr:
+    value: all
+  nodemodel_scalers:
+    value: none
+  nodemodel_layers:
+    value: 3
+  final_mlp_layers:
+    value: 1