keras_layer_AnchorBoxes.py

'''
A custom Keras layer to generate anchor boxes.

Copyright (C) 2017 Pierluigi Ferrari

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program.  If not, see <http://www.gnu.org/licenses/>.
'''

import keras.backend as K
from keras.engine.topology import InputSpec
from keras.engine.topology import Layer
import numpy as np

from ssd_box_encode_decode_utils import convert_coordinates

class AnchorBoxes(Layer):
    '''
    A Keras layer to create an output tensor containing anchor box coordinates based on the
    input tensor and the passed arguments.

    A set of 2D anchor boxes of different aspect ratios is created for each spatial unit of
    the input tensor. The number of anchor boxes created per unit depends on the arguments
    `aspect_ratios` and `two_boxes_for_ar1`, in the default case it is 4. The boxes
    are parameterized by the coordinate tuple `(xmin, xmax, ymin, ymax)`.

    The logic implemented by this layer is identical to the logic in the module
    `ssd_box_encode_decode_utils.py`.

    The purpose of having this layer in the network is to make the model self-sufficient
    at inference time. Since the model is predicting offsets to the anchor boxes
    (rather than predicting absolute box coordinates directly), one needs to know the anchor
    box coordinates in order to construct the final prediction boxes from the predicted offsets.
    If the model's output tensor did not contain the anchor box coordinates, the necessary
    information to convert the predicted offsets back to absolute coordinates would be missing
    in the model output. The reason why it is necessary to predict offsets to the anchor boxes
    rather than to predict absolute box coordinates directly is explained in `README.md`.

    Input shape:
        4D tensor of shape `(batch, channels, height, width)` if `dim_ordering = 'th'`
        or `(batch, height, width, channels)` if `dim_ordering = 'tf'`.

    Output shape:
        5D tensor of shape `(batch, height, width, n_boxes, 4)`.
    '''

    def __init__(self,
                 img_height,
                 img_width,
                 this_scale,
                 next_scale,
                 aspect_ratios=[0.5, 1.0, 2.0],
                 two_boxes_for_ar1=True,
                 limit_boxes=True,
                 variances=[1.0, 1.0, 1.0, 1.0],
                 coords='centroids',
                 normalize_coords=False,
                 **kwargs):
        '''
        All arguments need to be set to the same values as in the box encoding process, otherwise the behavior is undefined.

        Arguments:
            img_height (int): The height of the input images.
            img_width (int): The width of the input images.
            this_scale (float): A float in [0, 1], the scaling factor for the size of the generate default boxes
                as a fraction of the shorter side of the input image.
            next_scale (float): A float in [0, 1], the next larger scaling factor. Only relevant if
                `self.two_boxes_for_ar1 == True`.
            aspect_ratios (list, optional): The list of aspect ratios for which default boxes are to be
                generated for this layer. Defaults to [0.5, 1.0, 2.0].
            two_boxes_for_ar1 (bool, optional): Only relevant if `aspect_ratios` contains 1.
                If `True`, two default boxes will be generated for aspect ratio 1. The first will be generated
                using the scaling factor for the respective layer, the second one will be generated using
                geometric mean of said scaling factor and next bigger scaling factor. Defaults to `True`.
            limit_boxes (bool, optional): If `True`, limits box coordinates to stay within image boundaries.
                Defaults to `True`.
            variances (list, optional): A list of 4 floats >0 with scaling factors (actually it's not factors but divisors
                to be precise) for the encoded predicted box coordinates. A variance value of 1.0 would apply
                no scaling at all to the predictions, while values in (0,1) upscale the encoded predictions and values greater
                than 1.0 downscale the encoded predictions. If you want to reproduce the configuration of the original SSD,
                set this to `[0.1, 0.1, 0.2, 0.2]`, provided the coordinate format is 'centroids'. Defaults to `[1.0, 1.0, 1.0, 1.0]`.
            coords (str, optional): The box coordinate format to be used. Can be either 'centroids' for the format
                `(cx, cy, w, h)` (box center coordinates, width, and height) or 'minmax' for the format
                `(xmin, xmax, ymin, ymax)`. Defaults to 'centroids'.
            normalize_coords (bool, optional): Set to `True` if the model uses relative instead of absolute coordinates,
                i.e. if the model predicts box coordinates within [0,1] instead of absolute coordinates. Defaults to `False`.
        '''
        if K.backend() != 'tensorflow':
            raise TypeError("This layer only supports TensorFlow at the moment, but you are using the {} backend.".format(K.backend()))

        if (this_scale < 0) or (next_scale < 0) or (this_scale > 1):
            raise ValueError("`this_scale` must be in [0, 1] and `next_scale` must be >0, but `this_scale` == {}, `next_scale` == {}".format(this_scale, next_scale))

        if len(variances) != 4:
            raise ValueError("4 variance values must be pased, but {} values were received.".format(len(variances)))
        variances = np.array(variances)
        if np.any(variances <= 0):
            raise ValueError("All variances must be >0, but the variances given are {}".format(variances))

        self.img_height = img_height
        self.img_width = img_width
        self.this_scale = this_scale
        self.next_scale = next_scale
        self.aspect_ratios = aspect_ratios
        self.two_boxes_for_ar1 = two_boxes_for_ar1
        self.limit_boxes = limit_boxes
        self.variances = variances
        self.coords = coords
        self.normalize_coords = normalize_coords
        # Compute the number of boxes per cell
        if (1 in aspect_ratios) & two_boxes_for_ar1:
            self.n_boxes = len(aspect_ratios) + 1
        else:
            self.n_boxes = len(aspect_ratios)
        super(AnchorBoxes, self).__init__(**kwargs)

    def build(self, input_shape):
        self.input_spec = [InputSpec(shape=input_shape)]
        super(AnchorBoxes, self).build(input_shape)

    def call(self, x, mask=None):
        '''
        Return an anchor box tensor based on the shape of the input tensor.

        The logic implemented here is identical to the logic in the module `ssd_box_encode_decode_utils.py`.

        Note that this tensor does not participate in any graph computations at runtime. It is being created
        as a constant once during graph creation and is just being output along with the rest of the model output
        during runtime. Because of this, all logic is implemented as Numpy array operations and it is sufficient
        to convert the resulting Numpy array into a Keras tensor at the very end before outputting it.

        Arguments:
            x (tensor): 4D tensor of shape `(batch, channels, height, width)` if `dim_ordering = 'th'`
                or `(batch, height, width, channels)` if `dim_ordering = 'tf'`. The input for this
                layer must be the output of the localization predictor layer.
        '''

        # Compute box width and height for each aspect ratio
        # The shorter side of the image will be used to compute `w` and `h` using `scale` and `aspect_ratios`.
        self.aspect_ratios = np.sort(self.aspect_ratios)
        size = min(self.img_height, self.img_width)
        # Compute the box widths and and heights for all aspect ratios
        wh_list = []
        for ar in self.aspect_ratios:
            if (ar == 1) & self.two_boxes_for_ar1:
                # Compute the regular default box for aspect ratio 1 and...
                w = self.this_scale * size * np.sqrt(ar)
                h = self.this_scale * size / np.sqrt(ar)
                wh_list.append((w,h))
                # ...also compute one slightly larger version using the geometric mean of this scale value and the next
                w = np.sqrt(self.this_scale * self.next_scale) * size * np.sqrt(ar)
                h = np.sqrt(self.this_scale * self.next_scale) * size / np.sqrt(ar)
                wh_list.append((w,h))
            else:
                w = self.this_scale * size * np.sqrt(ar)
                h = self.this_scale * size / np.sqrt(ar)
                wh_list.append((w,h))
        wh_list = np.array(wh_list)

        # We need the shape of the input tensor
        batch_size, feature_map_height, feature_map_width, feature_map_channels = x._keras_shape

        # Compute the grid of box center points. They are identical for all aspect ratios
        cell_height = self.img_height / feature_map_height
        cell_width = self.img_width / feature_map_width
        cx = np.linspace(cell_width/2, self.img_width-cell_width/2, feature_map_width)
        cy = np.linspace(cell_height/2, self.img_height-cell_height/2, feature_map_height)
        cx_grid, cy_grid = np.meshgrid(cx, cy)
        cx_grid = np.expand_dims(cx_grid, -1) # This is necessary for np.tile() to do what we want further down
        cy_grid = np.expand_dims(cy_grid, -1) # This is necessary for np.tile() to do what we want further down

        # Create a 4D tensor template of shape `(feature_map_height, feature_map_width, n_boxes, 4)`
        # where the last dimension will contain `(cx, cy, w, h)`
        boxes_tensor = np.zeros((feature_map_height, feature_map_width, self.n_boxes, 4))

        boxes_tensor[:, :, :, 0] = np.tile(cx_grid, (1, 1, self.n_boxes)) # Set cx
        boxes_tensor[:, :, :, 1] = np.tile(cy_grid, (1, 1, self.n_boxes)) # Set cy
        boxes_tensor[:, :, :, 2] = wh_list[:, 0] # Set w
        boxes_tensor[:, :, :, 3] = wh_list[:, 1] # Set h

        # Convert `(cx, cy, w, h)` to `(xmin, xmax, ymin, ymax)`
        boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='centroids2minmax')

        # If `limit_boxes` is enabled, clip the coordinates to lie within the image boundaries
        if self.limit_boxes:
            x_coords = boxes_tensor[:,:,:,[0, 1]]
            x_coords[x_coords >= self.img_width] = self.img_width - 1
            x_coords[x_coords < 0] = 0
            boxes_tensor[:,:,:,[0, 1]] = x_coords
            y_coords = boxes_tensor[:,:,:,[2, 3]]
            y_coords[y_coords >= self.img_height] = self.img_height - 1
            y_coords[y_coords < 0] = 0
            boxes_tensor[:,:,:,[2, 3]] = y_coords

        # `normalize_coords` is enabled, normalize the coordinates to be within [0,1]
        if self.normalize_coords:
            boxes_tensor[:, :, :, :2] /= self.img_width
            boxes_tensor[:, :, :, 2:] /= self.img_height

        if self.coords == 'centroids':
            # TODO: Implement box limiting directly for `(cx, cy, w, h)` so that we don't have to unnecessarily convert back and forth
            # Convert `(xmin, xmax, ymin, ymax)` back to `(cx, cy, w, h)`
            boxes_tensor = convert_coordinates(boxes_tensor, start_index=0, conversion='minmax2centroids')

        # 4: Create a tensor to contain the variances and append it to `boxes_tensor`. This tensor has the same shape
        #    as `boxes_tensor` and simply contains the same 4 variance values for every position in the last axis.
        variances_tensor = np.zeros_like(boxes_tensor) # Has shape `(feature_map_height, feature_map_width, n_boxes, 4)`
        variances_tensor += self.variances # Long live broadcasting
        # Now `boxes_tensor` becomes a tensor of shape `(feature_map_height, feature_map_width, n_boxes, 8)`
        boxes_tensor = np.concatenate((boxes_tensor, variances_tensor), axis=-1)

        # Now prepend one dimension to `boxes_tensor` to account for the batch size and tile it along
        # The result will be a 5D tensor of shape `(batch_size, feature_map_height, feature_map_width, n_boxes, 8)`
        boxes_tensor = np.expand_dims(boxes_tensor, axis=0)
        boxes_tensor = K.tile(K.constant(boxes_tensor, dtype='float32'), (K.shape(x)[0], 1, 1, 1, 1))

        return boxes_tensor

    def compute_output_shape(self, input_shape):
        batch_size, feature_map_height, feature_map_width, feature_map_channels = input_shape
        return (batch_size, feature_map_height, feature_map_width, self.n_boxes, 8)