Module imageapply.tools

Functions

def apply_model(model: Callable[[~T], ~T], data: ~T, batch_size=None) ‑> ~T

Applies the model to a batch of data, running at most batch_size entries through at once.

Args

model : Callable[[T], T]

The model to apply

data : T

The data to apply the model to

batch_size : int

The maximum number of entries to run through the model at once. If None, runs all entries through at once.

Returns

T

The output of the model

Expand source code

def apply_model(model : Callable[[T], T], data : T, batch_size=None) -> T:
    """
    Applies the model to a batch of data, running at most batch_size entries through at once.
    
    Args:
        model (Callable[[T], T]): The model to apply
        data (T): The data to apply the model to
        batch_size (int): The maximum number of entries to run through the model at once. If None, \
            runs all entries through at once.
        
    Returns:
        T: The output of the model 
    """

    if batch_size is None or batch_size > data.shape[0]:
        return model(data)

    assert type(batch_size) is int and batch_size > 0, "Batch size must be a positive integer"

    results = []
    for i in range(0, data.shape[0], batch_size):
        results.append(model(data[i:i+min(batch_size, data.shape[0]-i)]))

    return np.concatenate(results, axis=0)

def combine_regions(data: ~T, region_size: tuple, original_size: tuple) ‑> ~T

Combines the regions of the data into the original shape.

Args

data : T

The data to combine

region_size : tuple

The size of the regions

original_size : tuple

The original shape of the data

Returns

T

The combined data

Expand source code

def combine_regions(data : T, region_size : tuple, original_size : tuple) -> T:
    """
    Combines the regions of the data into the original shape.

    Args:
        data (T): The data to combine
        region_size (tuple): The size of the regions
        original_size (tuple): The original shape of the data

    Returns:
        T: The combined data
    """
    n_splits = [original_size[i+1] // region_size[i] for i in range(len(region_size))]
    for dim in range(len(region_size)-1, -1, -1):
        data = np.concatenate(np.split(data, n_splits[dim], axis=0), axis=dim+1)

    return data

def crop_to_original(data: ~T, original_shape: tuple, pad_position: str = 'end') ‑> ~T

Crops the data to the original shape.

Args

data : T

The data to crop

original_shape : tuple

The original shape of the data

pad_position : str

The position to pad. One of "end" or "centre"

Returns

T

The cropped data

Expand source code

def crop_to_original(data : T, original_shape : tuple, pad_position : str = "end") -> T:
    """
    Crops the data to the original shape.
    
    Args:
        data (T): The data to crop
        original_shape (tuple): The original shape of the data
        pad_position (str): The position to pad. One of "end" or "centre"
        
    Returns:
        T: The cropped data
    """
    if pad_position == "end":
        crop_indices = tuple([slice(None)] + [slice(0, original_shape[i+1]) for i in range(len(original_shape)-1)])
    elif pad_position == "centre":
        pad_amount = [data.shape[i+1] - original_shape[i+1] for i in range(len(original_shape)-1)]
        crop_indices = tuple([slice(None)] + [slice(pad_amount[i]//2, pad_amount[i]//2 + original_shape[i+1]) for i in range(len(original_shape)-1)])
    return data[crop_indices]

def divide_into_regions(data: ~T, region_size: tuple) ‑> ~T

Divides the data into regions of size region_size.

Args

data : T

The data to divide

region_size : tuple

The size of the regions

Returns

T

The divided data

Expand source code

def divide_into_regions(data : T, region_size : tuple) -> T:
    """
    Divides the data into regions of size region_size.
    
    Args:
        data (T): The data to divide
        region_size (tuple): The size of the regions
    
    Returns:
        T: The divided data
    """

    # Assumes data is a multiple of region_size
    # data is array of shape (Batch_size, ...)
    # region_size is tuple of length data.ndim - 1

    n_splits = [data.shape[i+1] // region_size[i] for i in range(len(region_size))]
    for dim in range(len(region_size)):
        data = np.concatenate(np.split(data, n_splits[dim], axis=dim+1), axis=0)
    
    return data

def pad_to_multiple(data: ~T, input_size: tuple, pad_mode: str = 'zeros', pad_position: str = 'end') ‑> ~T

Pads the data to a multiple of the input size.

Args

data : T

The data to pad

input_size : tuple

The input size of the model

pad_mode : str

The mode to pad with. One of "zeros", "reflect", "symmetric"

pad_position : str

The position to pad. One of "end" or "centre"

Returns

T

The combined data

Expand source code

def pad_to_multiple(data:T, input_size:tuple, pad_mode:str="zeros", pad_position:str="end") -> T:
    """
    Pads the data to a multiple of the input size.

    Args:
        data (T): The data to pad
        input_size (tuple): The input size of the model
        pad_mode (str): The mode to pad with. One of "zeros", "reflect", "symmetric"
        pad_position (str): The position to pad. One of "end" or "centre"

    Returns:
        T: The combined data
    """

    rem_per_dim = [data.shape[i+1] % input_size[i] for i in range(len(input_size))]
    pad_amount = [input_size[i] - rem_per_dim[i] if rem_per_dim[i] > 0 else 0 for i in range(len(input_size))]
    if pad_position == "end":
        pad_amount = [(0, 0)] + [(0, pad_amount[i]) for i in range(len(pad_amount))]
    elif pad_position == "centre":
        pad_amount = [(0, 0)] + [(pad_amount[i]//2, pad_amount[i] - pad_amount[i]//2) for i in range(len(pad_amount))]
    else:
        raise ValueError("Invalid pad position. Must be 'end' or 'centre'")

    if pad_mode == "zeros":
        return np.pad(data, pad_amount, mode='constant', constant_values=0)
    elif pad_mode == "reflect":
        return np.pad(data, pad_amount, mode='reflect')
    elif pad_mode == "symmetric":
        return np.pad(data, pad_amount, mode='symmetric')
    else:
        raise ValueError("Invalid pad mode. Must be 'zeros', 'reflect', or 'symmetric'")