Skip to content

200-add-mini-unet-for-segmentation #367

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
nmaarnio merged 6 commits into from
Oct 22, 2024
Merged

200-add-mini-unet-for-segmentation #367

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
682766c
docstring update
RichardScottOZ Apr 9, 2024
1a908c9
update docstrings
RichardScottOZ Apr 9, 2024
bb6bbac
Unet parameters
RichardScottOZ Apr 9, 2024
898392a
dice loss strings
RichardScottOZ Apr 9, 2024
1c4cf01
output strings
RichardScottOZ Apr 9, 2024
237cc4c
Merge branch '200-add-mini-unet-for-segmentation' into 200-add-mini-u…
nmaarnio Oct 22, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
87 changes: 43 additions & 44 deletions eis_toolkit/unet/mini_unet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,20 +15,19 @@
@beartype
def img_loader(image_dir: str) -> (np.ndarray, list, np.ndarray):
"""
Fetches all the tiffs in the given directory and creates a numpy ndarray of the common shape.

Do the Fetches all the tiffs in the given directory and creates a numpy ndarray of shape.

Shape type (image_count, bands, width, height) from them. Returns the array, tiff metadata as list and associated
nodatamasks in shape (image_count, width, height) Tiffs are assumed to be same size and named {number}.tif
Calculates shape (image_count, bands, width, height) from them. Returns the array, tiff metadata as a list and associated
nodatamasks in shape (image_count, width, height). Tiffs are assumed to be same size and are named as {number}.tif
starting from 0

Parameter:
image_dir: the directory containing the images

Returns:
the numpy ndarray of the tiffs
the list containing the meta
no data mask
numpy ndarray of the tiffs
list containing the metadata
nodata mask
"""
# fetching the filepaths
paths = []
Expand Down Expand Up @@ -92,18 +91,17 @@ def img_loader(image_dir: str) -> (np.ndarray, list, np.ndarray):
@beartype
def label_loader(label_dir: str) -> (np.ndarray, int):
"""
Do the Fetches all tiffs in the given directory and return a numpy ndarray of shape (image_count, width, height).
Fetches all the tiffs in the given directory and creates a numpy ndarray of the shape (image_count, bands, width, height).

The images are assumed to be same size and contain the labels with numbers showing classes.
The images are assumed to be the same size and contain the labels with numbers showing classes.
The tiffs should have only one band.

Parameters:
label_dir: the directory containing the label files

Return:
a numpy ndarray containing labels
no data value

nodata value
"""
# fetching the filepaths
paths = []
Expand All @@ -112,7 +110,7 @@ def label_loader(label_dir: str) -> (np.ndarray, int):

img_count = len(paths)

# Getting the size of the images
# Getting the size of the rasters
with rasterio.open(paths[0]) as src:
meta = src.meta.copy()
nodata_value = meta["nodata"]
Expand Down Expand Up @@ -149,23 +147,23 @@ def build_autoencoder_multichannel_with_skip_connection(
) -> tf.keras.Model:
"""

Do build the Unet.
Build the Unet.

Parameters:
input_shape The shap of the input used by the Unet:
input_shape The shape of the input used by the Unet:
kernel_shape The shape of the convolution kernel:
list_of_convolutional_layers The list of convolutional layers of the Unet. This list will be reversed
for the decoder.
dropout: This is the dropout rate assigned.
dropout: The dropout rate assigned.
pool_size: The size of the max pooling layer.
up_sampling_factor: the decoder need up sampling factor to enlarge the features layer by layer.
output_filters: the number of filters of the output.
output_kernel: the dimension of the output.
last_activation: last activation sigmoid by default.
data_augmentation: if you want to include data augmentation right before the input layer.
data_augmentation_params_crop: if data augmentation is true fill this value.
data_augmentation_params_rotation: if data augmentation is true fill this value.
regularization: type of regularization for each layer.
up_sampling_factor: The decoder need up sampling factor to enlarge the features layer by layer.
output_filters: The number of filters of the output.
output_kernel: The dimension of the output.
last_activation: Last activation sigmoid by default.
data_augmentation: If you want to include data augmentation right before the input layer.
data_augmentation_params_crop: If data augmentation is True fill this value.
data_augmentation_params_rotation: if data augmentation is True fill this value.
regularization: Type of regularization for each layer.

Return:
the built Unet.
Expand Down Expand Up @@ -263,7 +261,7 @@ def dice_coeff_uncertain(
smooth: the smooth factor.

Return:
the dice loss value and the mask.
The dice loss value and the mask.
"""

y_true = tf.cast(y_true, dtype=tf.float32)
Expand All @@ -287,14 +285,14 @@ def dice_coeff_uncertain(
def regularization_loss(y_true: tf.Tensor, y_pred: tf.Tensor) -> float:
"""

Do compute the dice loss.
Computes the dice loss.

Parameters:
y_true: the true labels of the dataset.
y_pred: the predicted labels from the model.

Return:
a floating poit representing the loss.
a floating point representing the loss.
"""

y_pred_uncertain = y_pred[:, :, :, 1]
Expand All @@ -305,14 +303,14 @@ def regularization_loss(y_true: tf.Tensor, y_pred: tf.Tensor) -> float:
def dice_coefficient(prediction: tf.Tensor, true_label: tf.Tensor) -> tf.Tensor:
"""

Do calculate the dice loss coefficient.
Calculate the dice loss coefficient.

Parameters:
true_label: the true labels of the dataset.
prediction: the predicted labels from the model.

Return:
a floating poit representing the loss.
a floating point representing the loss.
"""
prediction = tf.cast(prediction, tf.float32)
true_label = tf.cast(true_label, tf.float32)
Expand All @@ -324,14 +322,14 @@ def dice_coefficient(prediction: tf.Tensor, true_label: tf.Tensor) -> tf.Tensor:
def dice_loss(prediction: tf.Tensor, true_label: tf.Tensor) -> tf.Tensor:
"""

Do calculate the dice loss.
Calculate the dice loss.

Parameters:
true_label: the true labels of the dataset.
prediction: the predicted labels from the model.

Return:
a floating poit representing the loss.
a floating point representing the loss.

"""

Expand Down Expand Up @@ -404,28 +402,29 @@ def train_and_predict_the_model(
Train and predict the Unet.

Parameters:
x_train: a numpy array with the training sample inside
y_train: labels of the training dataset,
x_test: a numpy array with the testing sample inside,
y_test: labels of the testing dataset,
batch_size: how many sample per epochs should be used for fitting the model,
x_train: a numpy array with the training samplse
y_train: labels of the training dataset
x_test: a numpy array with the testing samples
y_test: labels of the testing dataset
batch_size: how many sample per epochs should be used for fitting the model
epochs: how many epochs used for the training phase.
is_uncertainty: bool = True, if you want to add uncertainty estimation to your Mini-Unet.
list_of_convolutional_layers: how many convolutional layer suggestion -> [32, 64, 128, 256],
dropout: This is the dropout rate assigned. It is used to randomly remove some predictions.
pool_size: The size of the max pooling layer.
up_sampling_factor: the decoder need up sampling factor to enlarge the features layer by layer.
output_filters: the number of output nodes.
output_kernel: the dimension of the output, it should be the same of the input.
last_activation: last activation sigmoid by default.
up_sampling_factor: the decoder upsampling factor to enlarge the features layer by layer.
output_filters: the number of filters of the output.
output_kernel: the dimension of the output.
last_activation: last activation, sigmoid by default.
data_augmentation: if you want to include data augmentation right before the input layer.
data_augmentation_params_crop: if data augmentation is true fill this value (crop range).
data_augmentation_params_rotation: if data augmentation is true fill this value (rotation range).
regularization: type of regularization for each layer. (L1, L2, L1L2, or None).
data_augmentation_params_crop: if data augmentation is True fill this value (crop range).
data_augmentation_params_rotation: if data augmentation is True fill this value (rotation range).
regularization: type of regularization for each layer.

Raise:
InvalidInputException: when one input is null.
InvalidInputException: when an input is null.
Return:
the predicted numpy array
The predicted numpy array
"""

if x_train.shape[0] == 0 or x_train is None:
Expand Down