Skip to content

noise2noise

Self-supervised denoiser implementation, based on Noise2Noise.

@author: Nicola VIGANÒ, CEA-MEM, Grenoble, France

Classes:

  • N2N –

    Self-supervised denoising from pairs of images.

N2N 

N2N(
    model: int | str | NetworkParams | Module | Mapping,
    data_scale_bias: DataScaleBias | None = None,
    reg_val: float | LossRegularizer | None = None,
    device: str = "cuda" if is_available() else "cpu",
    batch_size: int | None = None,
    augmentation: (
        str
        | Augmentation
        | Sequence[str | Augmentation]
        | None
    ) = None,
    save_epochs_dir: str | None = None,
    verbose: bool = True,
)

Bases: Denoiser

Self-supervised denoising from pairs of images.

Parameters:

  • model (str | NetworkParams | Module | Mapping | None) –

    Type of neural network to use or a specific network (or state) to use

  • data_scale_bias (DataScaleBias | None, default: None ) –

    Scale and bias of the input data, by default None

  • reg_val (float | None, default: None ) –

    Regularization value, by default 1e-5

  • device (str, default: 'cuda' if is_available() else 'cpu' ) –

    Device to use, by default "cuda" if cuda is available, otherwise "cpu"

  • save_epochs_dir (str | None, default: None ) –

    Directory where to save network states at each epoch. If None disabled, by default None

  • verbose (bool, default: True ) –

    Whether to produce verbose output, by default True

Methods:

  • infer –

    Perform inference on the input data.

  • prepare_data –

    Prepare input data for training.

  • train –

    Train the denoiser using the Noise2Noise self-supervised approach.

Attributes:

  • n_channels_in (int) –

    Returns the number of input channels of the model.

  • n_channels_out (int) –

    Returns the number of output channels of the model.

  • n_dims (int) –

    Returns the expected signal dimensions.

Source code in src/autoden/algorithms/denoiser.py 
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
def __init__(
    self,
    model: int | str | NetworkParams | pt.nn.Module | Mapping,
    data_scale_bias: DataScaleBias | None = None,
    reg_val: float | LossRegularizer | None = None,
    device: str = "cuda" if pt.cuda.is_available() else "cpu",
    batch_size: int | None = None,
    augmentation: str | Augmentation | Sequence[str | Augmentation] | None = None,
    save_epochs_dir: str | None = None,
    verbose: bool = True,
) -> None:
    """Initialize the noise2noise method.

    Parameters
    ----------
    model : str | NetworkParams | pt.nn.Module | Mapping | None
        Type of neural network to use or a specific network (or state) to use
    data_scale_bias : DataScaleBias | None, optional
        Scale and bias of the input data, by default None
    reg_val : float | None, optional
        Regularization value, by default 1e-5
    device : str, optional
        Device to use, by default "cuda" if cuda is available, otherwise "cpu"
    save_epochs_dir : str | None, optional
        Directory where to save network states at each epoch.
        If None disabled, by default None
    verbose : bool, optional
        Whether to produce verbose output, by default True
    """
    if isinstance(model, int):
        if self.save_epochs_dir is None:
            raise ValueError("Directory for saving epochs not specified")

        model = load_model_state(self.save_epochs_dir, epoch_num=model)

    if isinstance(model, (str, NetworkParams, Mapping, pt.nn.Module)):
        self.model = create_network(model, device=device)
    else:
        raise ValueError(f"Invalid model {type(model)}")
    if verbose:
        get_num_parameters(self.model, verbose=True)

    # if augmentation is None:
    #     augmentation = []
    # elif isinstance(augmentation, str):
    #     augmentation = [augmentation.lower()]
    # elif isinstance(augmentation, Sequence):
    #     augmentation = [str(a).lower() for a in augmentation]

    self.data_sb = data_scale_bias

    self.reg_val = reg_val
    self.device = device
    self.batch_size = batch_size
    self.augmentation = augmentation
    self.save_epochs_dir = save_epochs_dir
    self.verbose = verbose

n_channels_in property 

n_channels_in: int

Returns the number of input channels of the model.

If the model is an instance of SerializableModel and has an init_params attribute containing the key "n_channels_in", this property returns its value. Otherwise, it defaults to 1.

Returns:

  • int –

    The number of input channels.

n_channels_out property 

n_channels_out: int

Returns the number of output channels of the model.

If the model is an instance of SerializableModel and has an init_params attribute containing the key "n_channels_out", this property returns its value. Otherwise, it defaults to 1.

Returns:

  • int –

    The number of output channels.

n_dims property 

n_dims: int

Returns the expected signal dimensions.

If the model is an instance of SerializableModel and has an init_params attribute containing the key "n_dims", this property returns its value. Otherwise, it defaults to 2.

Returns:

  • int –

    The expected signal dimensions.

infer 

infer(
    inp: NDArray,
    average_splits: bool = True,
    channel_axis_dst: int | None = None,
) -> NDArray

Perform inference on the input data.

Parameters:

  • inp (NDArray) –

    The input data to perform inference on. It is expected to have an extra dimension including the different splits.

  • average_splits (bool, default: True ) –

    If True, the splits are averaged. Default is True.

  • channel_axis_dst (int | None, default: None ) –

    The desired channel axis for the output. If None, the output will have the same channel axis as the input.

Returns:

  • NDArray –

    The inferred output data. If average_splits is True, the splits are averaged.

Notes

If self.batch_size is set, the input data is processed in batches to avoid memory issues.

Source code in src/autoden/algorithms/noise2noise.py 
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
def infer(self, inp: NDArray, average_splits: bool = True, channel_axis_dst: int | None = None) -> NDArray:
    """
    Perform inference on the input data.

    Parameters
    ----------
    inp : NDArray
        The input data to perform inference on. It is expected to have an extra dimension including the different splits.
    average_splits : bool, optional
        If True, the splits are averaged. Default is True.
    channel_axis_dst : int | None, optional
        The desired channel axis for the output. If None, the output will have the same channel axis as the input.

    Returns
    -------
    NDArray
        The inferred output data. If `average_splits` is True, the splits are averaged.

    Notes
    -----
    If `self.batch_size` is set, the input data is processed in batches to avoid memory issues.
    """
    inp_shape = inp.shape
    dims_to_preserve = self.n_dims + (self.n_channels_in > 1)
    inp = inp.reshape([-1, *inp_shape[-dims_to_preserve:]])

    if self.batch_size is not None:
        out = []
        for b in tqdm(range(0, inp.shape[0], self.batch_size), desc="Inference batch"):
            out.append(super().infer(inp[b : b + self.batch_size]))
        out = np.concatenate(out, axis=0)
    else:
        out = super().infer(inp)
    out = out.reshape(inp_shape)

    if average_splits:
        realization_batch_axis = -dims_to_preserve - 1
        out = out.mean(axis=realization_batch_axis)

    if channel_axis_dst is not None:
        out = self._move_output_channel_axis(out, channel_axis_dst)

    return out

prepare_data 

prepare_data(
    inp: NDArray,
    num_tst_ratio: float = 0.2,
    strategy: str = "1:X",
    channel_axis: int | None = None,
    realizations_axis: int = 0,
) -> tuple[NDArray, NDArray, NDArray]

Prepare input data for training.

Parameters:

  • inp (NDArray) –

    The input data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the number of samples, and H and W are the height and width of each sample, respectively.

  • num_tst_ratio (float, default: 0.2 ) –

    The ratio of the input data to be used for testing. The remaining data will be used for training. Default is 0.2.

  • strategy (str, default: '1:X' ) –

    The strategy to be used for creating input-target pairs. The available strategies are: - "1:X": Use the mean of the remaining samples as the target for each sample. - "X:1": Use the mean of the remaining samples as the input for each sample. Default is "1:X".

  • channel_axis (int | None, default: None ) –

    The axis of the input array that corresponds to the spectral dimension. If None, the spectral dimension is assumed to not be present. Default is None.

  • realizations_axis (int, default: 0 ) –

    The axis of the input array that corresponds to the redundant realizations dimension. Default is 0.

Returns:

  • tuple[NDArray, NDArray, NDArray] –

    A tuple containing: - The input data array. - The target data array. - The mask array indicating the training pixels.

Notes

This function generates input-target pairs based on the specified strategy. It also generates a mask array indicating the training pixels based on the provided ratio.

Source code in src/autoden/algorithms/noise2noise.py 
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
def prepare_data(
    self,
    inp: NDArray,
    num_tst_ratio: float = 0.2,
    strategy: str = "1:X",
    channel_axis: int | None = None,
    realizations_axis: int = 0,
) -> tuple[NDArray, NDArray, NDArray]:
    """
    Prepare input data for training.

    Parameters
    ----------
    inp : NDArray
        The input data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the
        number of samples, and H and W are the height and width of each sample, respectively.
    num_tst_ratio : float, optional
        The ratio of the input data to be used for testing. The remaining data will be used for training.
        Default is 0.2.
    strategy : str, optional
        The strategy to be used for creating input-target pairs. The available strategies are:
        - "1:X": Use the mean of the remaining samples as the target for each sample.
        - "X:1": Use the mean of the remaining samples as the input for each sample.
        Default is "1:X".
    channel_axis : int | None, optional
        The axis of the input array that corresponds to the spectral dimension.
        If None, the spectral dimension is assumed to not be present.
        Default is None.
    realizations_axis : int, optional
        The axis of the input array that corresponds to the redundant realizations dimension. Default is 0.

    Returns
    -------
    tuple[NDArray, NDArray, NDArray]
        A tuple containing:
        - The input data array.
        - The target data array.
        - The mask array indicating the training pixels.

    Notes
    -----
    This function generates input-target pairs based on the specified strategy. It also generates a mask array
    indicating the training pixels based on the provided ratio.
    """
    inp, channel_axis = self._prepare_channel_axis(inp, channel_axis)
    self._check_channel_axis_size(inp, channel_axis, "n_channels_in")
    self._check_channel_axis_size(inp, channel_axis, "n_channels_out")

    model_n_axes = self.n_dims + (channel_axis is not None)
    if inp.ndim < model_n_axes:
        raise ValueError(f"Target data should at least be of {model_n_axes + 1} dimensions, but its shape is {inp.shape}")

    # This will be the destination position of the axis
    realizations_dst_axis = inp.ndim - model_n_axes - 1

    inp_x = np.stack(
        [np.delete(inp, obj=ii, axis=realizations_axis).mean(axis=realizations_axis) for ii in range(len(inp))],
        axis=realizations_dst_axis,
    )
    inp = np.moveaxis(inp, source=realizations_axis, destination=realizations_dst_axis)

    if strategy.upper() == "1:X":
        tmp_inp = inp
        tmp_tgt = inp_x
    elif strategy.upper() == "X:1":
        tmp_inp = inp_x
        tmp_tgt = inp
    else:
        raise ValueError(f"Strategy {strategy} not implemented. Please choose one of: ['1:X', 'X:1']")

    mask_trn = get_random_pixel_mask(inp.shape, mask_pixel_ratio=num_tst_ratio)

    return tmp_inp, tmp_tgt, mask_trn

train 

train(
    inp: NDArray,
    tgt: NDArray,
    mask_trn: NDArray,
    *,
    epochs: int,
    learning_rate: float = 0.001,
    optimizer: str = "adam",
    lower_limit: float | NDArray | None = None,
    restarts: int | None = None,
    accum_grads: bool = False
) -> dict[str, NDArray]

Train the denoiser using the Noise2Noise self-supervised approach.

Parameters:

  • inp (NDArray) –

    The input data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the number of samples, and H and W are the height and width of each sample, respectively.

  • tgt (NDArray) –

    The target data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the number of samples, and H and W are the height and width of each sample, respectively.

  • mask_trn (NDArray) –

    The mask array indicating the pixels used for training.

  • epochs (int) –

    The number of epochs to train the model.

  • learning_rate (float, default: 0.001 ) –

    The learning rate for the optimizer. Default is 1e-3.

  • optimizer (str, default: 'adam' ) –

    The optimization algorithm to be used for training. Default is "adam".

  • lower_limit (float | NDArray | None, default: None ) –

    The lower limit for the input data. If provided, the input data will be clipped to this limit. Default is None.

  • restarts (int | None, default: None ) –

    The number of times to restart the cosine annealing of the learning rate. If provided, the cosine annealing of the learning rate will be restarted the specified number of times. Default is None.

  • accum_grads (bool, default: False ) –

    Whether to accumulate gradients over multiple batches. If True, gradients will be accumulated over multiple batches before updating the model parameters. Default is False.

Returns:

  • dict[str, NDArray] –

    A dictionary containing the training losses.

Notes

This method uses the Noise2Noise self-supervised approach to train the denoiser. The input data is used to generate target data based on the specified strategy. The training process involves creating pairs of input and target data and then training the model to minimize the difference between the predicted and target data.

Source code in src/autoden/algorithms/noise2noise.py 
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
def train(
    self,
    inp: NDArray,
    tgt: NDArray,
    mask_trn: NDArray,
    *,
    epochs: int,
    learning_rate: float = 1e-3,
    optimizer: str = "adam",
    lower_limit: float | NDArray | None = None,
    restarts: int | None = None,
    accum_grads: bool = False,
) -> dict[str, NDArray]:
    """
    Train the denoiser using the Noise2Noise self-supervised approach.

    Parameters
    ----------
    inp : NDArray
        The input data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the
        number of samples, and H and W are the height and width of each sample, respectively.
    tgt : NDArray
        The target data to be used for training. This should be a NumPy array of shape (N, H, W), where N is the
        number of samples, and H and W are the height and width of each sample, respectively.
    mask_trn : NDArray
        The mask array indicating the pixels used for training.
    epochs : int
        The number of epochs to train the model.
    learning_rate : float, optional
        The learning rate for the optimizer. Default is 1e-3.
    optimizer : str, optional
        The optimization algorithm to be used for training. Default is "adam".
    lower_limit : float | NDArray | None, optional
        The lower limit for the input data. If provided, the input data will be clipped to this limit.
        Default is None.
    restarts : int | None, optional
        The number of times to restart the cosine annealing of the learning rate. If provided, the cosine annealing
        of the learning rate will be restarted the specified number of times. Default is None.
    accum_grads : bool, optional
        Whether to accumulate gradients over multiple batches. If True, gradients will be accumulated over multiple
        batches before updating the model parameters. Default is False.

    Returns
    -------
    dict[str, NDArray]
        A dictionary containing the training losses.

    Notes
    -----
    This method uses the Noise2Noise self-supervised approach to train the denoiser. The input data is used to
    generate target data based on the specified strategy. The training process involves creating pairs of input
    and target data and then training the model to minimize the difference between the predicted and target data.
    """
    if self.data_sb is None:
        self.data_sb = compute_scaling_selfsupervised(inp)

    # Rescale the datasets
    inp = inp * self.data_sb.scale_inp - self.data_sb.bias_inp
    tgt = tgt * self.data_sb.scale_tgt - self.data_sb.bias_tgt

    reg = self._get_regularization()
    losses = self._train_pixelmask_batched(
        inp=inp,
        tgt=tgt,
        mask_tst=mask_trn,
        epochs=epochs,
        learning_rate=learning_rate,
        optimizer=optimizer,
        regularizer=reg,
        lower_limit=lower_limit,
        restarts=restarts,
        accum_grads=accum_grads,
    )

    if self.verbose:
        self._plot_loss_curves(losses, f"Self-supervised {self.__class__.__name__} {optimizer.upper()}")

    return losses