def ssd_300(image_size,
n_classes,
min_scale=0.1,
max_scale=0.9,
scales=None,
aspect_ratios_global=None,
aspect_ratios_per_layer=[[0.5, 1.0, 2.0],
[1.0 / 3.0, 0.5, 1.0, 2.0, 3.0],
[1.0 / 3.0, 0.5, 1.0, 2.0, 3.0],
[1.0 / 3.0, 0.5, 1.0, 2.0, 3.0],
[0.5, 1.0, 2.0], [0.5, 1.0, 2.0]],
two_boxes_for_ar1=True,
limit_boxes=False,
variances=[0.1, 0.1, 0.2, 0.2],
coords='centroids',
normalize_coords=False):
'''
Build a Keras model with SSD_300 architecture, see references.
The base network is a reduced atrous VGG-16, extended by the SSD architecture,
as described in the paper.
In case you're wondering why this function has so many arguments: All arguments except
the first two (`image_size` and `n_classes`) are only needed so that the anchor box
layers can produce the correct anchor boxes. In case you're training the network, the
parameters passed here must be the same as the ones used to set up `SSDBoxEncoder`.
In case you're loading trained weights, the parameters passed here must be the same
as the ones used to produce the trained weights.
Note: Requires Keras v2.0 or later. Currently works only with the
TensorFlow backend (v1.0 or later).
Arguments:
image_size (tuple): The input image size in the format `(height, width, channels)`.
n_classes (int): The number of categories for classification including
the background class (i.e. the number of positive classes +1 for
the background calss).
min_scale (float, optional): The smallest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images. Defaults to 0.1.
max_scale (float, optional): The largest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images. All scaling factors between the smallest and the
largest will be linearly interpolated. Note that the second to last of the linearly interpolated
scaling factors will actually be the scaling factor for the last predictor layer, while the last
scaling factor is used for the second box for aspect ratio 1 in the last predictor layer
if `two_boxes_for_ar1` is `True`. Defaults to 0.9.
scales (list, optional): A list of floats containing scaling factors per convolutional predictor layer.
This list must be one element longer than the number of predictor layers. The first `k` elements are the
scaling factors for the `k` predictor layers, while the last element is used for the second box
for aspect ratio 1 in the last predictor layer if `two_boxes_for_ar1` is `True`. This additional
last scaling factor must be passed either way, even if it is not being used.
Defaults to `None`. If a list is passed, this argument overrides `min_scale` and
`max_scale`. All scaling factors must be greater than zero.
aspect_ratios_global (list, optional): The list of aspect ratios for which anchor boxes are to be
generated. This list is valid for all prediction layers. Defaults to None.
aspect_ratios_per_layer (list, optional): A list containing one aspect ratio list for each prediction layer.
This allows you to set the aspect ratios for each predictor layer individually, which is the case for the
original SSD300 implementation. If a list is passed, it overrides `aspect_ratios_global`.
Defaults to the aspect ratios used in the original SSD300 architecture, i.e.:
[[0.5, 1.0, 2.0],
[1.0/3.0, 0.5, 1.0, 2.0, 3.0],
[1.0/3.0, 0.5, 1.0, 2.0, 3.0],
[1.0/3.0, 0.5, 1.0, 2.0, 3.0],
[0.5, 1.0, 2.0],
[0.5, 1.0, 2.0]]
two_boxes_for_ar1 (bool, optional): Only relevant for aspect ratio lists that contain 1. Will be ignored otherwise.
If `True`, two anchor 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`, following the original
implementation.
limit_boxes (bool, optional): If `True`, limits box coordinates to stay within image boundaries.
This would normally be set to `True`, but here it defaults to `False`, following the original
implementation.
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. Defaults to `[0.1, 0.1, 0.2, 0.2]`, following the original implementation.
The coordinate format must be 'centroids'.
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', following the original implementation.
normalize_coords (bool, optional): Set to `True` if the model is supposed to use relative instead of absolute coordinates,
i.e. if the model predicts box coordinates within [0,1] instead of absolute coordinates. Defaults to `False`.
Returns:
model: The Keras SSD model.
predictor_sizes: A Numpy array containing the `(height, width)` portion
of the output tensor shape for each convolutional predictor layer. During
training, the generator function needs this in order to transform
the ground truth labels into tensors of identical structure as the
output tensors of the model, which is in turn needed for the cost
function.
References:
https://arxiv.org/abs/1512.02325v5
'''
n_predictor_layers = 6 # The number of predictor conv layers in the network is 6 for the original SSD300
# Get a few exceptions out of the way first
if aspect_ratios_global is None and aspect_ratios_per_layer is None:
raise ValueError(
"`aspect_ratios_global` and `aspect_ratios_per_layer` cannot both be None. At least one needs to be specified."
)
if aspect_ratios_per_layer:
if len(aspect_ratios_per_layer) != n_predictor_layers:
raise ValueError(
"It must be either aspect_ratios_per_layer is None or len(aspect_ratios_per_layer) == {}, but len(aspect_ratios_per_layer) == {}."
.format(n_predictor_layers, len(aspect_ratios_per_layer)))
if (min_scale is None or max_scale is None) and scales is None:
raise ValueError(
"Either `min_scale` and `max_scale` or `scales` need to be specified."
)
if scales:
if len(scales) != n_predictor_layers + 1:
raise ValueError(
"It must be either scales is None or len(scales) == {}, but len(scales) == {}."
.format(n_predictor_layers + 1, len(scales)))
else: # If no explicit list of scaling factors was passed, compute the list of scaling factors from `min_scale` and `max_scale`
scales = np.linspace(min_scale, max_scale, n_predictor_layers + 1)
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))
# Set the aspect ratios for each predictor layer. These are only needed for the anchor box layers.
if aspect_ratios_per_layer:
aspect_ratios_conv4_3 = aspect_ratios_per_layer[0]
aspect_ratios_fc7 = aspect_ratios_per_layer[1]
aspect_ratios_conv6_2 = aspect_ratios_per_layer[2]
aspect_ratios_conv7_2 = aspect_ratios_per_layer[3]
aspect_ratios_conv8_2 = aspect_ratios_per_layer[4]
aspect_ratios_conv9_2 = aspect_ratios_per_layer[5]
else:
aspect_ratios_conv4_3 = aspect_ratios_global
aspect_ratios_fc7 = aspect_ratios_global
aspect_ratios_conv6_2 = aspect_ratios_global
aspect_ratios_conv7_2 = aspect_ratios_global
aspect_ratios_conv8_2 = aspect_ratios_global
aspect_ratios_conv9_2 = aspect_ratios_global
# Compute the number of boxes to be predicted per cell for each predictor layer.
# We need this so that we know how many channels the predictor layers need to have.
if aspect_ratios_per_layer:
n_boxes = []
for aspect_ratios in aspect_ratios_per_layer:
if (1 in aspect_ratios) & two_boxes_for_ar1:
n_boxes.append(len(aspect_ratios) +
1) # +1 for the second box for aspect ratio 1
else:
n_boxes.append(len(aspect_ratios))
n_boxes_conv4_3 = n_boxes[
0] # 4 boxes per cell for the original implementation
n_boxes_fc7 = n_boxes[
1] # 6 boxes per cell for the original implementation
n_boxes_conv6_2 = n_boxes[
2] # 6 boxes per cell for the original implementation
n_boxes_conv7_2 = n_boxes[
3] # 6 boxes per cell for the original implementation
n_boxes_conv8_2 = n_boxes[
4] # 4 boxes per cell for the original implementation
n_boxes_conv9_2 = n_boxes[
5] # 4 boxes per cell for the original implementation
else: # If only a global aspect ratio list was passed, then the number of boxes is the same for each predictor layer
if (1 in aspect_ratios_global) & two_boxes_for_ar1:
n_boxes = len(aspect_ratios_global) + 1
else:
n_boxes = len(aspect_ratios_global)
n_boxes_conv4_3 = n_boxes
n_boxes_fc7 = n_boxes
n_boxes_conv6_2 = n_boxes
n_boxes_conv7_2 = n_boxes
n_boxes_conv8_2 = n_boxes
n_boxes_conv9_2 = n_boxes
# Input image format
img_height, img_width, img_channels = image_size[0], image_size[
1], image_size[2]
### Design the actual network
x = Input(shape=(img_height, img_width, img_channels))
normed = Lambda(
lambda z: z / 127.5 - 1.0, # Convert input feature range to [-1,1]
output_shape=(img_height, img_width, img_channels),
name='lambda1')(x)
conv1_1 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='conv1_1')(normed)
conv1_2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='conv1_2')(conv1_1)
pool1 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid',
name='pool1')(conv1_2)
conv2_1 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='conv2_1')(pool1)
conv2_2 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='conv2_2')(conv2_1)
pool2 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid',
name='pool2')(conv2_2)
conv3_1 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='conv3_1')(pool2)
conv3_2 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='conv3_2')(conv3_1)
conv3_3 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='conv3_3')(conv3_2)
pool3 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid',
name='pool3')(conv3_3)
conv4_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv4_1')(pool3)
conv4_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv4_2')(conv4_1)
conv4_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv4_3')(conv4_2)
pool4 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid',
name='pool4')(conv4_3)
conv5_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv5_1')(pool4)
conv5_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv5_2')(conv5_1)
conv5_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='conv5_3')(conv5_2)
pool5 = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
padding='same',
name='pool5')(conv5_3)
fc6 = Conv2D(1024, (3, 3),
dilation_rate=(6, 6),
activation='relu',
padding='same',
name='fc6')(pool5)
fc7 = Conv2D(1024, (1, 1), activation='relu', padding='same',
name='fc7')(fc6)
conv6_1 = Conv2D(256, (1, 1),
activation='relu',
padding='same',
name='conv6_1')(fc7)
conv6_2 = Conv2D(512, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
name='conv6_2')(conv6_1)
conv7_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
name='conv7_1')(conv6_2)
conv7_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
name='conv7_2')(conv7_1)
conv8_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
name='conv8_1')(conv7_2)
conv8_2 = Conv2D(256, (3, 3),
strides=(1, 1),
activation='relu',
padding='valid',
name='conv8_2')(conv8_1)
conv9_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
name='conv9_1')(conv8_2)
conv9_2 = Conv2D(256, (3, 3),
strides=(1, 1),
activation='relu',
padding='valid',
name='conv9_2')(conv9_1)
# Feed conv4_3 into the L2 normalization layer
conv4_3_norm = L2Normalization(gamma_init=20, name='conv4_3_norm')(conv4_3)
### Build the convolutional predictor layers on top of the base network
# We precidt `n_classes` confidence values for each box, hence the confidence predictors have depth `n_boxes * n_classes`
# Output shape of the confidence layers: `(batch, height, width, n_boxes * n_classes)`
conv4_3_norm_mbox_conf = Conv2D(
n_boxes_conv4_3 * n_classes, (3, 3),
padding='same',
name='conv4_3_norm_mbox_conf')(conv4_3_norm)
fc7_mbox_conf = Conv2D(n_boxes_fc7 * n_classes, (3, 3),
padding='same',
name='fc7_mbox_conf')(fc7)
conv6_2_mbox_conf = Conv2D(n_boxes_conv6_2 * n_classes, (3, 3),
padding='same',
name='conv6_2_mbox_conf')(conv6_2)
conv7_2_mbox_conf = Conv2D(n_boxes_conv7_2 * n_classes, (3, 3),
padding='same',
name='conv7_2_mbox_conf')(conv7_2)
conv8_2_mbox_conf = Conv2D(n_boxes_conv8_2 * n_classes, (3, 3),
padding='same',
name='conv8_2_mbox_conf')(conv8_2)
conv9_2_mbox_conf = Conv2D(n_boxes_conv9_2 * n_classes, (3, 3),
padding='same',
name='conv9_2_mbox_conf')(conv9_2)
# ****************************************************************************
# 用于生成角度
conv4_3_norm_mbox_angle = Conv2D(
n_boxes_conv4_3 * 2, (3, 3),
padding='same',
name='conv4_3_norm_mbox_angle')(conv4_3_norm)
fc7_mbox_angle = Conv2D(n_boxes_fc7 * 2, (3, 3),
padding='same',
name='fc7_mbox_angle')(fc7)
conv6_2_mbox_angle = Conv2D(n_boxes_conv6_2 * 2, (3, 3),
padding='same',
name='conv6_2_mbox_angle')(conv6_2)
conv7_2_mbox_angle = Conv2D(n_boxes_conv7_2 * 2, (3, 3),
padding='same',
name='conv7_2_mbox_angle')(conv7_2)
conv8_2_mbox_angle = Conv2D(n_boxes_conv8_2 * 2, (3, 3),
padding='same',
name='conv8_2_mbox_angle')(conv8_2)
conv9_2_mbox_angle = Conv2D(n_boxes_conv9_2 * 2, (3, 3),
padding='same',
name='conv9_2_mbox_angle')(conv9_2)
# ****************************************************************************
# We predict 4 box coordinates for each box, hence the localization predictors have depth `n_boxes * 4`
# Output shape of the localization layers: `(batch, height, width, n_boxes * 4)`
conv4_3_norm_mbox_loc = Conv2D(n_boxes_conv4_3 * 4, (3, 3),
padding='same',
name='conv4_3_norm_mbox_loc')(conv4_3_norm)
fc7_mbox_loc = Conv2D(n_boxes_fc7 * 4, (3, 3),
padding='same',
name='fc7_mbox_loc')(fc7)
conv6_2_mbox_loc = Conv2D(n_boxes_conv6_2 * 4, (3, 3),
padding='same',
name='conv6_2_mbox_loc')(conv6_2)
conv7_2_mbox_loc = Conv2D(n_boxes_conv7_2 * 4, (3, 3),
padding='same',
name='conv7_2_mbox_loc')(conv7_2)
conv8_2_mbox_loc = Conv2D(n_boxes_conv8_2 * 4, (3, 3),
padding='same',
name='conv8_2_mbox_loc')(conv8_2)
conv9_2_mbox_loc = Conv2D(n_boxes_conv9_2 * 4, (3, 3),
padding='same',
name='conv9_2_mbox_loc')(conv9_2)
### Generate the anchor boxes (called "priors" in the original Caffe/C++ implementation, so I'll keep their layer names)
# Output shape of anchors: `(batch, height, width, n_boxes, 8)`
conv4_3_norm_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[0],
next_scale=scales[1],
aspect_ratios=aspect_ratios_conv4_3,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv4_3_norm_mbox_priorbox')(conv4_3_norm)
fc7_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[1],
next_scale=scales[2],
aspect_ratios=aspect_ratios_fc7,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='fc7_mbox_priorbox')(fc7)
conv6_2_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[2],
next_scale=scales[3],
aspect_ratios=aspect_ratios_conv6_2,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv6_2_mbox_priorbox')(conv6_2)
conv7_2_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[3],
next_scale=scales[4],
aspect_ratios=aspect_ratios_conv7_2,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv7_2_mbox_priorbox')(conv7_2)
conv8_2_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[4],
next_scale=scales[5],
aspect_ratios=aspect_ratios_conv8_2,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv8_2_mbox_priorbox')(conv8_2)
conv9_2_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[5],
next_scale=scales[6],
aspect_ratios=aspect_ratios_conv9_2,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv9_2_mbox_priorbox')(conv9_2)
### Reshape
# Reshape the class predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, n_classes)`
# We want the classes isolated in the last axis to perform softmax on them
conv4_3_norm_mbox_conf_reshape = Reshape(
(-1, n_classes),
name='conv4_3_norm_mbox_conf_reshape')(conv4_3_norm_mbox_conf)
fc7_mbox_conf_reshape = Reshape(
(-1, n_classes), name='fc7_mbox_conf_reshape')(fc7_mbox_conf)
conv6_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv6_2_mbox_conf_reshape')(conv6_2_mbox_conf)
conv7_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv7_2_mbox_conf_reshape')(conv7_2_mbox_conf)
conv8_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv8_2_mbox_conf_reshape')(conv8_2_mbox_conf)
conv9_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv9_2_mbox_conf_reshape')(conv9_2_mbox_conf)
# ****************************************************************************
# reshape角度
conv4_3_norm_mbox_angle_reshape = Reshape(
(-1, 2),
name='conv4_3_norm_mbox_angle_reshape')(conv4_3_norm_mbox_angle)
fc7_mbox_angle_reshape = Reshape(
(-1, 2), name='fc7_mbox_angle_reshape')(fc7_mbox_angle)
conv6_2_mbox_angle_reshape = Reshape(
(-1, 2), name='conv6_2_mbox_angle_reshape')(conv6_2_mbox_angle)
conv7_2_mbox_angle_reshape = Reshape(
(-1, 2), name='conv7_2_mbox_angle_reshape')(conv7_2_mbox_angle)
conv8_2_mbox_angle_reshape = Reshape(
(-1, 2), name='conv8_2_mbox_angle_reshape')(conv8_2_mbox_angle)
conv9_2_mbox_angle_reshape = Reshape(
(-1, 2), name='conv9_2_mbox_angle_reshape')(conv9_2_mbox_angle)
# ****************************************************************************
# Reshape the box predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, 4)`
# We want the four box coordinates isolated in the last axis to compute the smooth L1 loss
conv4_3_norm_mbox_loc_reshape = Reshape(
(-1, 4), name='conv4_3_norm_mbox_loc_reshape')(conv4_3_norm_mbox_loc)
fc7_mbox_loc_reshape = Reshape((-1, 4),
name='fc7_mbox_loc_reshape')(fc7_mbox_loc)
conv6_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv6_2_mbox_loc_reshape')(conv6_2_mbox_loc)
conv7_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv7_2_mbox_loc_reshape')(conv7_2_mbox_loc)
conv8_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv8_2_mbox_loc_reshape')(conv8_2_mbox_loc)
conv9_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv9_2_mbox_loc_reshape')(conv9_2_mbox_loc)
# Reshape the anchor box tensors, yielding 3D tensors of shape `(batch, height * width * n_boxes, 8)`
conv4_3_norm_mbox_priorbox_reshape = Reshape(
(-1, 8),
name='conv4_3_norm_mbox_priorbox_reshape')(conv4_3_norm_mbox_priorbox)
fc7_mbox_priorbox_reshape = Reshape(
(-1, 8), name='fc7_mbox_priorbox_reshape')(fc7_mbox_priorbox)
conv6_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv6_2_mbox_priorbox_reshape')(conv6_2_mbox_priorbox)
conv7_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv7_2_mbox_priorbox_reshape')(conv7_2_mbox_priorbox)
conv8_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv8_2_mbox_priorbox_reshape')(conv8_2_mbox_priorbox)
conv9_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv9_2_mbox_priorbox_reshape')(conv9_2_mbox_priorbox)
### Concatenate the predictions from the different layers
# Axis 0 (batch) and axis 2 (n_classes or 4, respectively) are identical for all layer predictions,
# so we want to concatenate along axis 1, the number of boxes per layer
# Output shape of `mbox_conf`: (batch, n_boxes_total, n_classes)
mbox_conf = Concatenate(axis=1, name='mbox_conf')([
conv4_3_norm_mbox_conf_reshape, fc7_mbox_conf_reshape,
conv6_2_mbox_conf_reshape, conv7_2_mbox_conf_reshape,
conv8_2_mbox_conf_reshape, conv9_2_mbox_conf_reshape
])
# ****************************************************************************
mbox_angle = Concatenate(axis=1, name='mbox_angle')([
conv4_3_norm_mbox_angle_reshape, fc7_mbox_angle_reshape,
conv6_2_mbox_angle_reshape, conv7_2_mbox_angle_reshape,
conv8_2_mbox_angle_reshape, conv9_2_mbox_angle_reshape
])
# ****************************************************************************
# Output shape of `mbox_loc`: (batch, n_boxes_total, 4)
mbox_loc = Concatenate(axis=1, name='mbox_loc')([
conv4_3_norm_mbox_loc_reshape, fc7_mbox_loc_reshape,
conv6_2_mbox_loc_reshape, conv7_2_mbox_loc_reshape,
conv8_2_mbox_loc_reshape, conv9_2_mbox_loc_reshape
])
# Output shape of `mbox_priorbox`: (batch, n_boxes_total, 8)
mbox_priorbox = Concatenate(axis=1, name='mbox_priorbox')([
conv4_3_norm_mbox_priorbox_reshape, fc7_mbox_priorbox_reshape,
conv6_2_mbox_priorbox_reshape, conv7_2_mbox_priorbox_reshape,
conv8_2_mbox_priorbox_reshape, conv9_2_mbox_priorbox_reshape
])
# The box coordinate predictions will go into the loss function just the way they are,
# but for the class predictions, we'll apply a softmax activation layer first
mbox_conf_softmax = Activation('softmax',
name='mbox_conf_softmax')(mbox_conf)
# Concatenate the class and box predictions and the anchors to one large predictions vector
# Output shape of `predictions`: (batch, n_boxes_total, n_classes + 4 + 8)
#############################################################################
predictions = Concatenate(axis=2, name='predictions')(
[mbox_conf_softmax, mbox_angle, mbox_loc, mbox_priorbox])
#############################################################################
model = Model(inputs=x, outputs=predictions)
# Get the spatial dimensions (height, width) of the predictor conv layers, we need them to
# be able to generate the default boxes for the matching process outside of the model during training.
# Note that the original implementation performs anchor box matching inside the loss function. We don't do that.
# Instead, we'll do it in the batch generator function.
# The spatial dimensions are the same for the confidence and localization predictors, so we just take those of the conf layers.
predictor_sizes = np.array([
conv4_3_norm_mbox_conf._keras_shape[1:3],
fc7_mbox_conf._keras_shape[1:3], conv6_2_mbox_conf._keras_shape[1:3],
conv7_2_mbox_conf._keras_shape[1:3],
conv8_2_mbox_conf._keras_shape[1:3],
conv9_2_mbox_conf._keras_shape[1:3]
])
return model, predictor_sizes
def mn_model(image_size,
n_classes,
min_scale=0.1,
max_scale=0.9,
scales=None,
aspect_ratios_global=[0.5, 1.0, 2.0],
aspect_ratios_per_layer=None,
two_boxes_for_ar1=True,
limit_boxes=True,
variances=[1.0, 1.0, 1.0, 1.0],
coords='centroids',
normalize_coords=False):
n_predictor_layers = 6 # The number of predictor conv layers in the network is 6 for the original SSD300
# Get a few exceptions out of the way first
if aspect_ratios_global is None and aspect_ratios_per_layer is None:
raise ValueError("`aspect_ratios_global` and `aspect_ratios_per_layer` cannot both be None. At least one needs to be specified.")
if aspect_ratios_per_layer:
if len(aspect_ratios_per_layer) != n_predictor_layers:
raise ValueError("It must be either aspect_ratios_per_layer is None or len(aspect_ratios_per_layer) == {}, but len(aspect_ratios_per_layer) == {}.".format(n_predictor_layers, len(aspect_ratios_per_layer)))
if (min_scale is None or max_scale is None) and scales is None:
raise ValueError("Either `min_scale` and `max_scale` or `scales` need to be specified.")
if scales:
if len(scales) != n_predictor_layers+1:
raise ValueError("It must be either scales is None or len(scales) == {}, but len(scales) == {}.".format(n_predictor_layers+1, len(scales)))
else: # If no explicit list of scaling factors was passed, compute the list of scaling factors from `min_scale` and `max_scale`
scales = np.linspace(min_scale, max_scale, n_predictor_layers+1)
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))
# Set the aspect ratios for each predictor layer. These are only needed for the anchor box layers.
if aspect_ratios_per_layer:
aspect_ratios_conv4_3 = aspect_ratios_per_layer[0]
aspect_ratios_fc7 = aspect_ratios_per_layer[1]
aspect_ratios_conv6_2 = aspect_ratios_per_layer[2]
aspect_ratios_conv7_2 = aspect_ratios_per_layer[3]
aspect_ratios_conv8_2 = aspect_ratios_per_layer[4]
aspect_ratios_conv9_2 = aspect_ratios_per_layer[5]
else:
aspect_ratios_conv4_3 = aspect_ratios_global
aspect_ratios_fc7 = aspect_ratios_global
aspect_ratios_conv6_2 = aspect_ratios_global
aspect_ratios_conv7_2 = aspect_ratios_global
aspect_ratios_conv8_2 = aspect_ratios_global
aspect_ratios_conv9_2 = aspect_ratios_global
# Compute the number of boxes to be predicted per cell for each predictor layer.
# We need this so that we know how many channels the predictor layers need to have.
if aspect_ratios_per_layer:
n_boxes = []
for aspect_ratios in aspect_ratios_per_layer:
if (1 in aspect_ratios) & two_boxes_for_ar1:
n_boxes.append(len(aspect_ratios) + 1) # +1 for the second box for aspect ratio 1
else:
n_boxes.append(len(aspect_ratios))
n_boxes_conv4_3 = n_boxes[0] # 4 boxes per cell for the original implementation
n_boxes_fc7 = n_boxes[1] # 6 boxes per cell for the original implementation
n_boxes_conv6_2 = n_boxes[2] # 6 boxes per cell for the original implementation
n_boxes_conv7_2 = n_boxes[3] # 6 boxes per cell for the original implementation
n_boxes_conv8_2 = n_boxes[4] # 4 boxes per cell for the original implementation
n_boxes_conv9_2 = n_boxes[5] # 4 boxes per cell for the original implementation
else: # If only a global aspect ratio list was passed, then the number of boxes is the same for each predictor layer
if (1 in aspect_ratios_global) & two_boxes_for_ar1:
n_boxes = len(aspect_ratios_global) + 1
else:
n_boxes = len(aspect_ratios_global)
n_boxes_conv4_3 = n_boxes
n_boxes_fc7 = n_boxes
n_boxes_conv6_2 = n_boxes
n_boxes_conv7_2 = n_boxes
n_boxes_conv8_2 = n_boxes
n_boxes_conv9_2 = n_boxes
print ("Height, Width, Channels :", image_size[0], image_size[1], image_size[2])
# Input image format
img_height, img_width, img_channels = image_size[0], image_size[1], image_size[2]
input_shape = (img_height, img_width, img_channels)
img_input = Input(shape=input_shape)
alpha = 1.0
depth_multiplier = 1
x = Lambda(lambda z: z/255., # Convert input feature range to [-1,1]
output_shape=(img_height, img_width, img_channels),
name='lambda1')(img_input)
x = Lambda(lambda z: z - 0.5, # Convert input feature range to [-1,1]
output_shape=(img_height, img_width, img_channels),
name='lambda2')(x)
x = Lambda(lambda z: z *2., # Convert input feature range to [-1,1]
output_shape=(img_height, img_width, img_channels),
name='lambda3')(x)
x = _conv_block(x, 32, alpha, strides=(2, 2))
x = _depthwise_conv_block_classification(x, 64, alpha, depth_multiplier, block_id=1)
x = _depthwise_conv_block_classification(x, 128, alpha, depth_multiplier,
strides=(2, 2), block_id=2)
x = _depthwise_conv_block_classification(x, 128, alpha, depth_multiplier, block_id=3)
x = _depthwise_conv_block_classification(x, 256, alpha, depth_multiplier,
strides=(2, 2), block_id=4)
x = _depthwise_conv_block_classification(x, 256, alpha, depth_multiplier, block_id=5)
x = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier,
strides=(2, 2), block_id=6)
x = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier, block_id=7)
x = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier, block_id=8)
x = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier, block_id=9)
x = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier, block_id=10)
conv4_3 = _depthwise_conv_block_classification(x, 512, alpha, depth_multiplier, block_id=11) #11 conv4_3 (300x300)-> 19x19
x = _depthwise_conv_block_classification(conv4_3, 1024, alpha, depth_multiplier,
strides=(2, 2), block_id=12) # (300x300) -> 10x10
fc7 = _depthwise_conv_block_classification(x, 1024, alpha, depth_multiplier, block_id=13) # 13 fc7 (300x300) -> 10x10
conv6_1 = bn_conv(fc7, 'detection_conv6_1', 256, 1, 1, subsample =(1,1), border_mode ='same', bias=conv_has_bias)
conv6_2 = _depthwise_conv_block_detection(input = conv6_1, layer_name='detection_conv6_2', strides=(2,2),
pointwise_conv_filters=512, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=1)
conv7_1 = bn_conv(conv6_2, 'detection_conv7_1', 128, 1, 1, subsample =(1,1), border_mode ='same', bias=conv_has_bias)
conv7_2 = _depthwise_conv_block_detection(input = conv7_1, layer_name='detection_conv7_2', strides=(2,2),
pointwise_conv_filters=256, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=2)
#conv7_1 = Conv2D(128, (1, 1), activation='relu', padding='same', name='detection_conv7_1')(conv6_2)
#conv7_2 = Conv2D(256, (3, 3), strides=(2, 2), activation='relu', padding='same', name='detection_conv7_2')(conv7_1)
conv8_1 = bn_conv(conv7_2, 'detection_conv8_1', 128, 1, 1, subsample =(1,1), border_mode ='same', bias=conv_has_bias)
conv8_2 = _depthwise_conv_block_detection(input = conv8_1, layer_name='detection_conv8_2', strides=(2,2),
pointwise_conv_filters=256, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=3)
# # conv8_2 = bn_conv(conv8_1, 'detection_conv8_2', 256, 2, 2, subsample =(1,1), border_mode ='same', bias=conv_has_bias)
conv9_1 = bn_conv(conv8_2, 'detection_conv9_1', 64, 1, 1, subsample =(1,1), border_mode ='same', bias=conv_has_bias)
# conv9_2 = bn_conv(conv9_1, 'detection_conv9_2', 128, 3, 3, subsample =(2,2), border_mode ='same', bias=conv_has_bias)
conv9_2 = _depthwise_conv_block_detection(input = conv9_1, layer_name='detection_conv9_2', strides=(2,2),
pointwise_conv_filters=256, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=4)
# Feed conv4_3 into the L2 normalization layer
conv4_3_norm = L2Normalization(gamma_init=20, name='detection_conv4_3_norm')(conv4_3)
conv4_3_norm_mbox_conf = _depthwise_conv_block_detection(input = conv4_3_norm, layer_name='detection_conv4_3_norm_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_conv4_3 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=1)
fc7_mbox_conf = _depthwise_conv_block_detection(input = fc7, layer_name='detection_fc7_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_fc7 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=2)
conv6_2_mbox_conf = _depthwise_conv_block_detection(input = conv6_2, layer_name='detection_conv6_2_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_conv6_2 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=3)
conv7_2_mbox_conf = _depthwise_conv_block_detection(input = conv7_2, layer_name='detection_conv7_2_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_conv7_2 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=4)
conv8_2_mbox_conf = _depthwise_conv_block_detection(input = conv8_2, layer_name='detection_conv8_2_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_conv8_2 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=5)
conv9_2_mbox_conf = _depthwise_conv_block_detection(input = conv9_2, layer_name='detection_conv9_2_mbox_conf', strides=(1,1),
pointwise_conv_filters=n_boxes_conv9_2 * n_classes, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=6)
# We predict 4 box coordinates for each box, hence the localization predictors have depth `n_boxes * 4`
# Output shape of the localization layers: `(batch, height, width, n_boxes * 4)`
conv4_3_norm_mbox_loc = _depthwise_conv_block_detection(input = conv4_3_norm, layer_name='detection_conv4_3_norm_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_conv4_3 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=1)
fc7_mbox_loc = _depthwise_conv_block_detection(input = fc7, layer_name='detection_fc7_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_fc7 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=2)
conv6_2_mbox_loc = _depthwise_conv_block_detection(input = conv6_2, layer_name='detection_conv6_2_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_conv6_2 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=3)
conv7_2_mbox_loc = _depthwise_conv_block_detection(input = conv7_2, layer_name='detection_conv7_2_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_conv7_2 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=4)
conv8_2_mbox_loc = _depthwise_conv_block_detection(input = conv8_2, layer_name='detection_conv8_2_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_conv8_2 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=5)
conv9_2_mbox_loc = _depthwise_conv_block_detection(input = conv9_2, layer_name='detection_conv9_2_mbox_loc', strides=(1,1),
pointwise_conv_filters=n_boxes_conv9_2 * 4, alpha=alpha, depth_multiplier=depth_multiplier,
padding = 'same', use_bias = True, block_id=5)
### Generate the anchor boxes
# Output shape of anchors: `(batch, height, width, n_boxes, 8)`
conv4_3_norm_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[0], next_scale=scales[1], aspect_ratios=aspect_ratios_conv4_3,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_conv4_3_norm_mbox_priorbox')(conv4_3_norm)
fc7_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[1], next_scale=scales[2], aspect_ratios=aspect_ratios_fc7,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_fc7_mbox_priorbox')(fc7)
conv6_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[2], next_scale=scales[3], aspect_ratios=aspect_ratios_conv6_2,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_conv6_2_mbox_priorbox')(conv6_2)
conv7_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[3], next_scale=scales[4], aspect_ratios=aspect_ratios_conv7_2,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_conv7_2_mbox_priorbox')(conv7_2)
conv8_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[4], next_scale=scales[5], aspect_ratios=aspect_ratios_conv8_2,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_conv8_2_mbox_priorbox')(conv8_2)
conv9_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[5], next_scale=scales[6], aspect_ratios=aspect_ratios_conv9_2,
two_boxes_for_ar1=two_boxes_for_ar1, limit_boxes=limit_boxes, variances=variances, coords=coords, normalize_coords=normalize_coords, name='detection_conv9_2_mbox_priorbox')(conv9_2)
### Reshape
# Reshape the class predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, n_classes)`
# We want the classes isolated in the last axis to perform softmax on them
conv4_3_norm_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_conv4_3_norm_mbox_conf_reshape')(conv4_3_norm_mbox_conf)
fc7_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_fc7_mbox_conf_reshape')(fc7_mbox_conf)
conv6_2_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_conv6_2_mbox_conf_reshape')(conv6_2_mbox_conf)
conv7_2_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_conv7_2_mbox_conf_reshape')(conv7_2_mbox_conf)
conv8_2_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_conv8_2_mbox_conf_reshape')(conv8_2_mbox_conf)
conv9_2_mbox_conf_reshape = Reshape((-1, n_classes), name='detection_conv9_2_mbox_conf_reshape')(conv9_2_mbox_conf)
# Reshape the box predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, 4)`
# We want the four box coordinates isolated in the last axis to compute the smooth L1 loss
conv4_3_norm_mbox_loc_reshape = Reshape((-1, 4), name='detection_conv4_3_norm_mbox_loc_reshape')(conv4_3_norm_mbox_loc)
fc7_mbox_loc_reshape = Reshape((-1, 4), name='detection_fc7_mbox_loc_reshape')(fc7_mbox_loc)
conv6_2_mbox_loc_reshape = Reshape((-1, 4), name='detection_conv6_2_mbox_loc_reshape')(conv6_2_mbox_loc)
conv7_2_mbox_loc_reshape = Reshape((-1, 4), name='detection_conv7_2_mbox_loc_reshape')(conv7_2_mbox_loc)
conv8_2_mbox_loc_reshape = Reshape((-1, 4), name='detection_conv8_2_mbox_loc_reshape')(conv8_2_mbox_loc)
conv9_2_mbox_loc_reshape = Reshape((-1, 4), name='detection_conv9_2_mbox_loc_reshape')(conv9_2_mbox_loc)
# Reshape the anchor box tensors, yielding 3D tensors of shape `(batch, height * width * n_boxes, 8)`
conv4_3_norm_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_conv4_3_norm_mbox_priorbox_reshape')(conv4_3_norm_mbox_priorbox)
fc7_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_fc7_mbox_priorbox_reshape')(fc7_mbox_priorbox)
conv6_2_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_conv6_2_mbox_priorbox_reshape')(conv6_2_mbox_priorbox)
conv7_2_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_conv7_2_mbox_priorbox_reshape')(conv7_2_mbox_priorbox)
conv8_2_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_conv8_2_mbox_priorbox_reshape')(conv8_2_mbox_priorbox)
conv9_2_mbox_priorbox_reshape = Reshape((-1, 8), name='detection_conv9_2_mbox_priorbox_reshape')(conv9_2_mbox_priorbox)
### Concatenate the predictions from the different layers
# Axis 0 (batch) and axis 2 (n_classes or 4, respectively) are identical for all layer predictions,
# so we want to concatenate along axis 1, the number of boxes per layer
# Output shape of `mbox_conf`: (batch, n_boxes_total, n_classes)
mbox_conf = Concatenate(axis=1, name='detection_mbox_conf')([conv4_3_norm_mbox_conf_reshape,
fc7_mbox_conf_reshape,
conv6_2_mbox_conf_reshape,
conv7_2_mbox_conf_reshape,
conv8_2_mbox_conf_reshape,
conv9_2_mbox_conf_reshape])
# Output shape of `mbox_loc`: (batch, n_boxes_total, 4)
mbox_loc = Concatenate(axis=1, name='detection_mbox_loc')([conv4_3_norm_mbox_loc_reshape,
fc7_mbox_loc_reshape,
conv6_2_mbox_loc_reshape,
conv7_2_mbox_loc_reshape,
conv8_2_mbox_loc_reshape,
conv9_2_mbox_loc_reshape])
# Output shape of `mbox_priorbox`: (batch, n_boxes_total, 8)
mbox_priorbox = Concatenate(axis=1, name='detection_mbox_priorbox')([conv4_3_norm_mbox_priorbox_reshape,
fc7_mbox_priorbox_reshape,
conv6_2_mbox_priorbox_reshape,
conv7_2_mbox_priorbox_reshape,
conv8_2_mbox_priorbox_reshape,
conv9_2_mbox_priorbox_reshape])
# The box coordinate predictions will go into the loss function just the way they are,
# but for the class predictions, we'll apply a softmax activation layer first
mbox_conf_softmax = Activation('softmax', name='detection_mbox_conf_softmax')(mbox_conf)
# Concatenate the class and box predictions and the anchors to one large predictions vector
# Output shape of `predictions`: (batch, n_boxes_total, n_classes + 4 + 8)
predictions = Concatenate(axis=2, name='detection_predictions')([mbox_conf_softmax, mbox_loc, mbox_priorbox])
model = Model(inputs=img_input, outputs=predictions)
#model = Model(inputs=img_input, outputs=predictions)
# Get the spatial dimensions (height, width) of the predictor conv layers, we need them to
# be able to generate the default boxes for the matching process outside of the model during training.
# Note that the original implementation performs anchor box matching inside the loss function. We don't do that.
# Instead, we'll do it in the batch generator function.
# The spatial dimensions are the same for the confidence and localization predictors, so we just take those of the conf layers.
predictor_sizes = np.array([conv4_3_norm_mbox_conf._keras_shape[1:3],
fc7_mbox_conf._keras_shape[1:3],
conv6_2_mbox_conf._keras_shape[1:3],
conv7_2_mbox_conf._keras_shape[1:3],
conv8_2_mbox_conf._keras_shape[1:3],
conv9_2_mbox_conf._keras_shape[1:3]])
model_layer = dict([(layer.name, layer) for layer in model.layers])
# for key in model_layer:
# model_layer[key].trainable = True
# model = Model(img_input, conv9_2)
# model_layer = dict([(layer.name, layer) for layer in model.layers])
# predictor_sizes = 0
return model, model_layer, img_input, predictor_sizes
def ssd_512_surveillance(image_size,
n_classes,
l2_regularization=0.0005,
min_scale=None,
max_scale=None,
scales=None,
aspect_ratios_global=None,
aspect_ratios_per_layer=[
[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
],
two_boxes_for_ar1=True,
steps=[8, 16, 32, 64, 128],
offsets=None,
limit_boxes=False,
variances=[0.1, 0.1, 0.2, 0.2],
coords='centroids',
normalize_coords=False,
subtract_mean=[123, 117, 104],
divide_by_stddev=None,
swap_channels=True,
return_predictor_sizes=False):
# In Surveillance environment, there are no big objects ( bigger than 1/4 of scene)
# So original ssd512 has 7 layers, but we only need 5 for detecting
n_predictor_layers = 5
# Account for the background class.
n_classes += 1
# Make the internal name shorter.
l2_reg = l2_regularization
img_height, img_width, img_channels = image_size[0], image_size[
1], image_size[2]
############################################################################
# Get a few exceptions out of the way.
############################################################################
if aspect_ratios_global is None and aspect_ratios_per_layer is None:
raise ValueError(
"`aspect_ratios_global` and `aspect_ratios_per_layer` cannot both be None. At least one needs to be specified."
)
if aspect_ratios_per_layer:
if len(aspect_ratios_per_layer) != n_predictor_layers:
raise ValueError(
"It must be either aspect_ratios_per_layer is None or len(aspect_ratios_per_layer) == {}, but len(aspect_ratios_per_layer) == {}."
.format(n_predictor_layers, len(aspect_ratios_per_layer)))
if (min_scale is None or max_scale is None) and scales is None:
raise ValueError(
"Either `min_scale` and `max_scale` or `scales` need to be specified."
)
if scales:
if len(scales) != n_predictor_layers + 1:
raise ValueError(
"It must be either scales is None or len(scales) == {}, but len(scales) == {}."
.format(n_predictor_layers + 1, len(scales)))
# If no explicit list of scaling factors was passed, compute the list of scaling factors from `min_scale` and `max_scale`
else:
scales = np.linspace(min_scale, max_scale, n_predictor_layers + 1)
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))
if (not (steps is None)) and (len(steps) != n_predictor_layers):
raise ValueError(
"You must provide at least one step value per predictor layer.")
if (not (offsets is None)) and (len(offsets) != n_predictor_layers):
raise ValueError(
"You must provide at least one offset value per predictor layer.")
############################################################################
# Compute the anchor box parameters.
############################################################################
# Set the aspect ratios for each predictor layer. These are only needed for the anchor box layers.
if aspect_ratios_per_layer:
aspect_ratios = aspect_ratios_per_layer
else:
aspect_ratios = [aspect_ratios_global] * n_predictor_layers
# Compute the number of boxes to be predicted per cell for each predictor layer.
# We need this so that we know how many channels the predictor layers need to have.
if aspect_ratios_per_layer:
n_boxes = []
for ar in aspect_ratios_per_layer:
if (1 in ar) & two_boxes_for_ar1:
# +1 for the second box for aspect ratio 1
n_boxes.append(len(ar) + 1)
else:
n_boxes.append(len(ar))
# If only a global aspect ratio list was passed, then the number of boxes is the same for each predictor layer
else:
if (1 in aspect_ratios_global) & two_boxes_for_ar1:
n_boxes = len(aspect_ratios_global) + 1
else:
n_boxes = len(aspect_ratios_global)
n_boxes = [n_boxes] * n_predictor_layers
if steps is None:
steps = [None] * n_predictor_layers
if offsets is None:
offsets = [None] * n_predictor_layers
############################################################################
# Build the network.
############################################################################
x = Input(shape=(img_height, img_width, img_channels))
# The following identity layer is only needed so that the subsequent lambda layers can be optional.
x1 = Lambda(lambda z: z,
output_shape=(img_height, img_width, img_channels),
name='identity_layer')(x)
if not (subtract_mean is None):
x1 = Lambda(lambda z: z - np.array(subtract_mean),
output_shape=(img_height, img_width, img_channels),
name='input_mean_normalization')(x1)
if not (divide_by_stddev is None):
x1 = Lambda(lambda z: z / np.array(divide_by_stddev),
output_shape=(img_height, img_width, img_channels),
name='input_stddev_normalization')(x1)
if swap_channels and (img_channels == 3):
x1 = Lambda(lambda z: z[..., ::-1],
output_shape=(img_height, img_width, img_channels),
name='input_channel_swap')(x1)
conv1_1 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv1_1')(x1)
conv1_2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv1_2')(conv1_1)
pool1 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool1')(conv1_2)
conv2_1 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv2_1')(pool1)
conv2_2 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv2_2')(conv2_1)
pool2 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool2')(conv2_2)
conv3_1 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv3_1')(pool2)
conv3_2 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv3_2')(conv3_1)
conv3_3 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv3_3')(conv3_2)
pool3 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool3')(conv3_3)
conv4_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv4_1')(pool3)
conv4_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv4_2')(conv4_1)
conv4_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv4_3')(conv4_2)
pool4 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool4')(conv4_3)
conv5_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv5_1')(pool4)
conv5_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv5_2')(conv5_1)
conv5_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv5_3')(conv5_2)
pool5 = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
padding='same',
name='pool5')(conv5_3)
fc6 = Conv2D(1024, (3, 3),
dilation_rate=(6, 6),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='fc6')(pool5)
fc7 = Conv2D(1024, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='fc7')(fc6)
conv6_1 = Conv2D(256, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv6_1')(fc7)
conv6_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv6_padding')(conv6_1)
conv6_2 = Conv2D(512, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv6_2')(conv6_1)
conv7_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv7_1')(conv6_2)
conv7_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv7_padding')(conv7_1)
conv7_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv7_2')(conv7_1)
conv8_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv8_1')(conv7_2)
conv8_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv8_padding')(conv8_1)
conv8_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv8_2')(conv8_1)
# Delete two layers
# Feed conv4_3 into the L2 normalization layer
conv4_3_norm = L2Normalization(gamma_init=20, name='conv4_3_norm')(conv4_3)
### Build the convolutional predictor layers on top of the base network
# We predict `n_classes` confidence values for each box, hence the confidence predictors have depth `n_boxes * n_classes`
# Output shape of the confidence layers: `(batch, height, width, n_boxes * n_classes)`
conv4_3_norm_mbox_conf = Conv2D(
n_boxes[0] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv4_3_norm_mbox_conf')(conv4_3_norm)
fc7_mbox_conf = Conv2D(n_boxes[1] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='fc7_mbox_conf')(fc7)
conv6_2_mbox_conf = Conv2D(n_boxes[2] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv6_2_mbox_conf')(conv6_2)
conv7_2_mbox_conf = Conv2D(n_boxes[3] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv7_2_mbox_conf')(conv7_2)
conv8_2_mbox_conf = Conv2D(n_boxes[4] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv8_2_mbox_conf')(conv8_2)
# We predict 4 box coordinates for each box, hence the localization predictors have depth `n_boxes * 4`
# Output shape of the localization layers: `(batch, height, width, n_boxes * 4)`
conv4_3_norm_mbox_loc = Conv2D(n_boxes[0] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv4_3_norm_mbox_loc')(conv4_3_norm)
fc7_mbox_loc = Conv2D(n_boxes[1] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='fc7_mbox_loc')(fc7)
conv6_2_mbox_loc = Conv2D(n_boxes[2] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv6_2_mbox_loc')(conv6_2)
conv7_2_mbox_loc = Conv2D(n_boxes[3] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv7_2_mbox_loc')(conv7_2)
conv8_2_mbox_loc = Conv2D(n_boxes[4] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
kernel_regularizer=l2(l2_reg),
name='conv8_2_mbox_loc')(conv8_2)
### Generate the anchor boxes (called "priors" in the original Caffe/C++ implementation, so I'll keep their layer names)
# Output shape of anchors: `(batch, height, width, n_boxes, 8)`
conv4_3_norm_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[0],
next_scale=scales[1],
aspect_ratios=aspect_ratios[0],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[0],
this_offsets=offsets[0],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv4_3_norm_mbox_priorbox')(conv4_3_norm_mbox_loc)
fc7_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[1],
next_scale=scales[2],
aspect_ratios=aspect_ratios[1],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[1],
this_offsets=offsets[1],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='fc7_mbox_priorbox')(fc7_mbox_loc)
conv6_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[2],
next_scale=scales[3],
aspect_ratios=aspect_ratios[2],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[2],
this_offsets=offsets[2],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv6_2_mbox_priorbox')(conv6_2_mbox_loc)
conv7_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[3],
next_scale=scales[4],
aspect_ratios=aspect_ratios[3],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[3],
this_offsets=offsets[3],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv7_2_mbox_priorbox')(conv7_2_mbox_loc)
conv8_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[4],
next_scale=scales[5],
aspect_ratios=aspect_ratios[4],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[4],
this_offsets=offsets[4],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv8_2_mbox_priorbox')(conv8_2_mbox_loc)
### Reshape
# Reshape the class predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, n_classes)`
# We want the classes isolated in the last axis to perform softmax on them
conv4_3_norm_mbox_conf_reshape = Reshape(
(-1, n_classes),
name='conv4_3_norm_mbox_conf_reshape')(conv4_3_norm_mbox_conf)
fc7_mbox_conf_reshape = Reshape(
(-1, n_classes), name='fc7_mbox_conf_reshape')(fc7_mbox_conf)
conv6_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv6_2_mbox_conf_reshape')(conv6_2_mbox_conf)
conv7_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv7_2_mbox_conf_reshape')(conv7_2_mbox_conf)
conv8_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv8_2_mbox_conf_reshape')(conv8_2_mbox_conf)
# Reshape the box predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, 4)`
# We want the four box coordinates isolated in the last axis to compute the smooth L1 loss
conv4_3_norm_mbox_loc_reshape = Reshape(
(-1, 4), name='conv4_3_norm_mbox_loc_reshape')(conv4_3_norm_mbox_loc)
fc7_mbox_loc_reshape = Reshape((-1, 4),
name='fc7_mbox_loc_reshape')(fc7_mbox_loc)
conv6_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv6_2_mbox_loc_reshape')(conv6_2_mbox_loc)
conv7_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv7_2_mbox_loc_reshape')(conv7_2_mbox_loc)
conv8_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv8_2_mbox_loc_reshape')(conv8_2_mbox_loc)
# Reshape the anchor box tensors, yielding 3D tensors of shape `(batch, height * width * n_boxes, 8)`
conv4_3_norm_mbox_priorbox_reshape = Reshape(
(-1, 8),
name='conv4_3_norm_mbox_priorbox_reshape')(conv4_3_norm_mbox_priorbox)
fc7_mbox_priorbox_reshape = Reshape(
(-1, 8), name='fc7_mbox_priorbox_reshape')(fc7_mbox_priorbox)
conv6_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv6_2_mbox_priorbox_reshape')(conv6_2_mbox_priorbox)
conv7_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv7_2_mbox_priorbox_reshape')(conv7_2_mbox_priorbox)
conv8_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv8_2_mbox_priorbox_reshape')(conv8_2_mbox_priorbox)
### Concatenate the predictions from the different layers
# Axis 0 (batch) and axis 2 (n_classes or 4, respectively) are identical for all layer predictions,
# so we want to concatenate along axis 1, the number of boxes per layer
# Output shape of `mbox_conf`: (batch, n_boxes_total, n_classes)
mbox_conf = Concatenate(axis=1, name='mbox_conf')([
conv4_3_norm_mbox_conf_reshape,
fc7_mbox_conf_reshape,
conv6_2_mbox_conf_reshape,
conv7_2_mbox_conf_reshape,
conv8_2_mbox_conf_reshape,
])
# Output shape of `mbox_loc`: (batch, n_boxes_total, 4)
mbox_loc = Concatenate(axis=1, name='mbox_loc')([
conv4_3_norm_mbox_loc_reshape,
fc7_mbox_loc_reshape,
conv6_2_mbox_loc_reshape,
conv7_2_mbox_loc_reshape,
conv8_2_mbox_loc_reshape,
])
# Output shape of `mbox_priorbox`: (batch, n_boxes_total, 8)
mbox_priorbox = Concatenate(axis=1, name='mbox_priorbox')([
conv4_3_norm_mbox_priorbox_reshape,
fc7_mbox_priorbox_reshape,
conv6_2_mbox_priorbox_reshape,
conv7_2_mbox_priorbox_reshape,
conv8_2_mbox_priorbox_reshape,
])
# The box coordinate predictions will go into the loss function just the way they are,
# but for the class predictions, we'll apply a softmax activation layer first
mbox_conf_softmax = Activation('softmax',
name='mbox_conf_softmax')(mbox_conf)
# Concatenate the class and box predictions and the anchors to one large predictions vector
# Output shape of `predictions`: (batch, n_boxes_total, n_classes + 4 + 8)
predictions = Concatenate(axis=2, name='predictions')(
[mbox_conf_softmax, mbox_loc, mbox_priorbox])
model = Model(inputs=x, outputs=predictions)
if return_predictor_sizes:
# Get the spatial dimensions (height, width) of the predictor conv layers, we need them to
# be able to generate the default boxes for the matching process outside of the model during training.
# Note that the original implementation performs anchor box matching inside the loss function. We don't do that.
# Instead, we'll do it in the batch generator function.
# The spatial dimensions are the same for the confidence and localization predictors, so we just take those of the conf layers.
predictor_sizes = np.array([
conv4_3_norm_mbox_conf._keras_shape[1:3],
fc7_mbox_conf._keras_shape[1:3],
conv6_2_mbox_conf._keras_shape[1:3],
conv7_2_mbox_conf._keras_shape[1:3],
conv8_2_mbox_conf._keras_shape[1:3],
])
return model, predictor_sizes
else:
return model
def ssd_512(image_size,
n_classes,
min_scale=None,
max_scale=None,
scales=None,
aspect_ratios_global=None,
aspect_ratios_per_layer=[[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0, 1.0 / 3.0],
[1.0, 2.0, 0.5], [1.0, 2.0, 0.5]],
two_boxes_for_ar1=True,
steps=[8, 16, 32, 64, 128, 256, 512],
offsets=None,
limit_boxes=False,
variances=[0.1, 0.1, 0.2, 0.2],
coords='centroids',
normalize_coords=False,
subtract_mean=None,
divide_by_stddev=None,
swap_channels=False,
return_predictor_sizes=False):
'''
Build a Keras model with SSD512 architecture, see references.
The base network is a reduced atrous VGG-16, extended by the SSD architecture,
as described in the paper.
In case you're wondering why this function has so many arguments: All arguments except
the first two (`image_size` and `n_classes`) are only needed so that the anchor box
layers can produce the correct anchor boxes. In case you're training the network, the
parameters passed here must be the same as the ones used to set up `SSDBoxEncoder`.
In case you're loading trained weights, the parameters passed here must be the same
as the ones used to produce the trained weights.
Some of these arguments are explained in more detail in the documentation of the
`SSDBoxEncoder` class.
Note: Requires Keras v2.0 or later. Currently works only with the
TensorFlow backend (v1.0 or later).
Arguments:
image_size (tuple): The input image size in the format `(height, width, channels)`.
n_classes (int): The number of categories for classification including
the background class (i.e. the number of positive classes +1 for
the background calss).
min_scale (float, optional): The smallest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images.
max_scale (float, optional): The largest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images. All scaling factors between the smallest and the
largest will be linearly interpolated. Note that the second to last of the linearly interpolated
scaling factors will actually be the scaling factor for the last predictor layer, while the last
scaling factor is used for the second box for aspect ratio 1 in the last predictor layer
if `two_boxes_for_ar1` is `True`.
scales (list, optional): A list of floats containing scaling factors per convolutional predictor layer.
This list must be one element longer than the number of predictor layers. The first `k` elements are the
scaling factors for the `k` predictor layers, while the last element is used for the second box
for aspect ratio 1 in the last predictor layer if `two_boxes_for_ar1` is `True`. This additional
last scaling factor must be passed either way, even if it is not being used.
If a list is passed, this argument overrides `min_scale` and `max_scale`. All scaling factors
must be greater than zero.
aspect_ratios_global (list, optional): The list of aspect ratios for which anchor boxes are to be
generated. This list is valid for all prediction layers.
aspect_ratios_per_layer (list, optional): A list containing one aspect ratio list for each prediction layer.
This allows you to set the aspect ratios for each predictor layer individually, which is the case for the
original SSD512 implementation. If a list is passed, it overrides `aspect_ratios_global`.
Defaults to the aspect ratios used in the original SSD512 architecture, i.e.:
[[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5]]
two_boxes_for_ar1 (bool, optional): Only relevant for aspect ratio lists that contain 1. Will be ignored otherwise.
If `True`, two anchor 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`, following the original
implementation.
steps (list, optional): `None` or a list with as many elements as there are predictor layers. The elements can be
either ints/floats or tuples of two ints/floats. These numbers represent for each predictor layer how many
pixels apart the anchor box center points should be vertically and horizontally along the spatial grid over
the image. If the list contains ints/floats, then that value will be used for both spatial dimensions.
If the list contains tuples of two ints/floats, then they represent `(step_height, step_width)`.
If no steps are provided, then they will be computed such that the anchor box center points will form an
equidistant grid within the image dimensions.
offsets (list, optional): `None` or a list with as many elements as there are predictor layers. The elements can be
either floats or tuples of two floats. These numbers represent for each predictor layer how many
pixels from the top and left boarders of the image the top-most and left-most anchor box center points should be
as a fraction of `steps`. The last bit is important: The offsets are not absolute pixel values, but fractions
of the step size specified in the `steps` argument. If the list contains floats, then that value will
be used for both spatial dimensions. If the list contains tuples of two floats, then they represent
`(vertical_offset, horizontal_offset)`. If no offsets are provided, then they will default to 0.5 of the step size.
limit_boxes (bool, optional): If `True`, limits box coordinates to stay within image boundaries.
This would normally be set to `True`, but here it defaults to `False`, following the original
implementation.
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. Defaults to `[0.1, 0.1, 0.2, 0.2]`, following the original implementation.
The coordinate format must be 'centroids'.
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', following the original implementation.
normalize_coords (bool, optional): Set to `True` if the model is supposed to use relative instead of absolute coordinates,
i.e. if the model predicts box coordinates within [0,1] instead of absolute coordinates.
subtract_mean (array-like, optional): `None` or an array-like object of integers or floating point values
of any shape that is broadcast-compatible with the image shape. The elements of this array will be
subtracted from the image pixel intensity values. For example, pass a list of three integers
to perform per-channel mean normalization for color images.
divide_by_stddev (array-like, optional): `None` or an array-like object of non-zero integers or
floating point values of any shape that is broadcast-compatible with the image shape. The image pixel
intensity values will be divided by the elements of this array. For example, pass a list
of three integers to perform per-channel standard deviation normalization for color images.
swap_channels (bool, optional): If `True`, the color channel order of the input images will be reversed,
i.e. if the input color channel order is RGB, the color channels will be swapped to BGR.
return_predictor_sizes (bool, optional): If `True`, this function not only returns the model, but also
a list containing the spatial dimensions of the predictor layers. This isn't strictly necessary since
you can always get their sizes easily via the Keras API, but it's convenient and less error-prone
to get them this way. They are only relevant for training anyway (SSDBoxEncoder needs to know the
spatial dimensions of the predictor layers), for inference you don't need them.
Returns:
model: The Keras SSD512 model.
predictor_sizes (optional): A Numpy array containing the `(height, width)` portion
of the output tensor shape for each convolutional predictor layer. During
training, the generator function needs this in order to transform
the ground truth labels into tensors of identical structure as the
output tensors of the model, which is in turn needed for the cost
function.
References:
https://arxiv.org/abs/1512.02325v5
'''
n_predictor_layers = 7 # The number of predictor conv layers in the network is 7 for the original SSD512
# Get a few exceptions out of the way first.
if aspect_ratios_global is None and aspect_ratios_per_layer is None:
raise ValueError(
"`aspect_ratios_global` and `aspect_ratios_per_layer` cannot both be None. At least one needs to be specified."
)
if aspect_ratios_per_layer:
if len(aspect_ratios_per_layer) != n_predictor_layers:
raise ValueError(
"It must be either aspect_ratios_per_layer is None or len(aspect_ratios_per_layer) == {}, but len(aspect_ratios_per_layer) == {}."
.format(n_predictor_layers, len(aspect_ratios_per_layer)))
if (min_scale is None or max_scale is None) and scales is None:
raise ValueError(
"Either `min_scale` and `max_scale` or `scales` need to be specified."
)
if scales:
if len(scales) != n_predictor_layers + 1:
raise ValueError(
"It must be either scales is None or len(scales) == {}, but len(scales) == {}."
.format(n_predictor_layers + 1, len(scales)))
else: # If no explicit list of scaling factors was passed, compute the list of scaling factors from `min_scale` and `max_scale`
scales = np.linspace(min_scale, max_scale, n_predictor_layers + 1)
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))
if (not (steps is None)) and (len(steps) != n_predictor_layers):
raise ValueError(
"You must provide at least one step value per predictor layer.")
if (not (offsets is None)) and (len(offsets) != n_predictor_layers):
raise ValueError(
"You must provide at least one offset value per predictor layer.")
# Set the aspect ratios for each predictor layer. These are only needed for the anchor box layers.
if aspect_ratios_per_layer:
aspect_ratios = aspect_ratios_per_layer
else:
aspect_ratios = [aspect_ratios_global] * n_predictor_layers
# Compute the number of boxes to be predicted per cell for each predictor layer.
# We need this so that we know how many channels the predictor layers need to have.
if aspect_ratios_per_layer:
n_boxes = []
for ar in aspect_ratios_per_layer:
if (1 in ar) & two_boxes_for_ar1:
n_boxes.append(len(ar) +
1) # +1 for the second box for aspect ratio 1
else:
n_boxes.append(len(ar))
else: # If only a global aspect ratio list was passed, then the number of boxes is the same for each predictor layer
if (1 in aspect_ratios_global) & two_boxes_for_ar1:
n_boxes = len(aspect_ratios_global) + 1
else:
n_boxes = len(aspect_ratios_global)
n_boxes = [n_boxes] * n_predictor_layers
if steps is None:
steps = [None] * n_predictor_layers
if offsets is None:
offsets = [None] * n_predictor_layers
# Input image format
img_height, img_width, img_channels = image_size[0], image_size[
1], image_size[2]
### Build the actual network.
x = Input(shape=(img_height, img_width, img_channels))
# The following identity layer is only needed so that subsequent lambda layers can be optional.
x1 = Lambda(lambda z: z,
output_shape=(img_height, img_width, img_channels),
name='idendity_layer')(x)
if not (subtract_mean is None):
x1 = Lambda(lambda z: z - np.array(subtract_mean),
output_shape=(img_height, img_width, img_channels),
name='input_mean_normalization')(x1)
if not (divide_by_stddev is None):
x1 = Lambda(lambda z: z / np.array(divide_by_stddev),
output_shape=(img_height, img_width, img_channels),
name='input_stddev_normalization')(x1)
if swap_channels and (img_channels == 3):
x1 = Lambda(lambda z: z[..., ::-1],
output_shape=(img_height, img_width, img_channels),
name='input_channel_swap')(x1)
conv1_1 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_1')(x1)
conv1_2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_2')(conv1_1)
pool1 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool1')(conv1_2)
conv2_1 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_1')(pool1)
conv2_2 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_2')(conv2_1)
pool2 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool2')(conv2_2)
conv3_1 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_1')(pool2)
conv3_2 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_2')(conv3_1)
conv3_3 = Conv2D(256, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_3')(conv3_2)
pool3 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool3')(conv3_3)
conv4_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_1')(pool3)
conv4_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_2')(conv4_1)
conv4_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_3')(conv4_2)
pool4 = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='same',
name='pool4')(conv4_3)
conv5_1 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_1')(pool4)
conv5_2 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_2')(conv5_1)
conv5_3 = Conv2D(512, (3, 3),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_3')(conv5_2)
pool5 = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
padding='same',
name='pool5')(conv5_3)
fc6 = Conv2D(1024, (3, 3),
dilation_rate=(6, 6),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='fc6')(pool5)
fc7 = Conv2D(1024, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='fc7')(fc6)
conv6_1 = Conv2D(256, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_1')(fc7)
conv6_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv6_padding')(conv6_1)
conv6_2 = Conv2D(512, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
name='conv6_2')(conv6_1)
conv7_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_1')(conv6_2)
conv7_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv7_padding')(conv7_1)
conv7_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
name='conv7_2')(conv7_1)
conv8_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_1')(conv7_2)
conv8_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv8_padding')(conv8_1)
conv8_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
name='conv8_2')(conv8_1)
conv9_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_1')(conv8_2)
conv9_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv9_padding')(conv9_1)
conv9_2 = Conv2D(256, (3, 3),
strides=(2, 2),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
name='conv9_2')(conv9_1)
conv10_1 = Conv2D(128, (1, 1),
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv10_1')(conv9_2)
conv10_1 = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv10_padding')(conv10_1)
conv10_2 = Conv2D(256, (4, 4),
strides=(1, 1),
activation='relu',
padding='valid',
kernel_initializer='he_normal',
name='conv10_2')(conv10_1)
# Feed conv4_3 into the L2 normalization layer
conv4_3_norm = L2Normalization(gamma_init=20, name='conv4_3_norm')(conv4_3)
### Build the convolutional predictor layers on top of the base network
# We precidt `n_classes` confidence values for each box, hence the confidence predictors have depth `n_boxes * n_classes`
# Output shape of the confidence layers: `(batch, height, width, n_boxes * n_classes)`
conv4_3_norm_mbox_conf = Conv2D(
n_boxes[0] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv4_3_norm_mbox_conf')(conv4_3_norm)
fc7_mbox_conf = Conv2D(n_boxes[1] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='fc7_mbox_conf')(fc7)
conv6_2_mbox_conf = Conv2D(n_boxes[2] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv6_2_mbox_conf')(conv6_2)
conv7_2_mbox_conf = Conv2D(n_boxes[3] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv7_2_mbox_conf')(conv7_2)
conv8_2_mbox_conf = Conv2D(n_boxes[4] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv8_2_mbox_conf')(conv8_2)
conv9_2_mbox_conf = Conv2D(n_boxes[5] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv9_2_mbox_conf')(conv9_2)
conv10_2_mbox_conf = Conv2D(n_boxes[6] * n_classes, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv10_2_mbox_conf')(conv10_2)
# We predict 4 box coordinates for each box, hence the localization predictors have depth `n_boxes * 4`
# Output shape of the localization layers: `(batch, height, width, n_boxes * 4)`
conv4_3_norm_mbox_loc = Conv2D(n_boxes[0] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv4_3_norm_mbox_loc')(conv4_3_norm)
fc7_mbox_loc = Conv2D(n_boxes[1] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='fc7_mbox_loc')(fc7)
conv6_2_mbox_loc = Conv2D(n_boxes[2] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv6_2_mbox_loc')(conv6_2)
conv7_2_mbox_loc = Conv2D(n_boxes[3] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv7_2_mbox_loc')(conv7_2)
conv8_2_mbox_loc = Conv2D(n_boxes[4] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv8_2_mbox_loc')(conv8_2)
conv9_2_mbox_loc = Conv2D(n_boxes[5] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv9_2_mbox_loc')(conv9_2)
conv10_2_mbox_loc = Conv2D(n_boxes[6] * 4, (3, 3),
padding='same',
kernel_initializer='he_normal',
name='conv10_2_mbox_loc')(conv10_2)
### Generate the anchor boxes (called "priors" in the original Caffe/C++ implementation, so I'll keep their layer names)
# Output shape of anchors: `(batch, height, width, n_boxes, 8)`
conv4_3_norm_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[0],
next_scale=scales[1],
aspect_ratios=aspect_ratios[0],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[0],
this_offsets=offsets[0],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv4_3_norm_mbox_priorbox')(conv4_3_norm_mbox_loc)
fc7_mbox_priorbox = AnchorBoxes(img_height,
img_width,
this_scale=scales[1],
next_scale=scales[2],
aspect_ratios=aspect_ratios[1],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[1],
this_offsets=offsets[1],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='fc7_mbox_priorbox')(fc7_mbox_loc)
conv6_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[2],
next_scale=scales[3],
aspect_ratios=aspect_ratios[2],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[2],
this_offsets=offsets[2],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv6_2_mbox_priorbox')(conv6_2_mbox_loc)
conv7_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[3],
next_scale=scales[4],
aspect_ratios=aspect_ratios[3],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[3],
this_offsets=offsets[3],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv7_2_mbox_priorbox')(conv7_2_mbox_loc)
conv8_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[4],
next_scale=scales[5],
aspect_ratios=aspect_ratios[4],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[4],
this_offsets=offsets[4],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv8_2_mbox_priorbox')(conv8_2_mbox_loc)
conv9_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[5],
next_scale=scales[6],
aspect_ratios=aspect_ratios[5],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[5],
this_offsets=offsets[5],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv9_2_mbox_priorbox')(conv9_2_mbox_loc)
conv10_2_mbox_priorbox = AnchorBoxes(
img_height,
img_width,
this_scale=scales[6],
next_scale=scales[7],
aspect_ratios=aspect_ratios[6],
two_boxes_for_ar1=two_boxes_for_ar1,
this_steps=steps[6],
this_offsets=offsets[6],
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
name='conv10_2_mbox_priorbox')(conv10_2_mbox_loc)
### Reshape
# Reshape the class predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, n_classes)`
# We want the classes isolated in the last axis to perform softmax on them
conv4_3_norm_mbox_conf_reshape = Reshape(
(-1, n_classes),
name='conv4_3_norm_mbox_conf_reshape')(conv4_3_norm_mbox_conf)
fc7_mbox_conf_reshape = Reshape(
(-1, n_classes), name='fc7_mbox_conf_reshape')(fc7_mbox_conf)
conv6_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv6_2_mbox_conf_reshape')(conv6_2_mbox_conf)
conv7_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv7_2_mbox_conf_reshape')(conv7_2_mbox_conf)
conv8_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv8_2_mbox_conf_reshape')(conv8_2_mbox_conf)
conv9_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv9_2_mbox_conf_reshape')(conv9_2_mbox_conf)
conv10_2_mbox_conf_reshape = Reshape(
(-1, n_classes), name='conv10_2_mbox_conf_reshape')(conv10_2_mbox_conf)
# Reshape the box predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, 4)`
# We want the four box coordinates isolated in the last axis to compute the smooth L1 loss
conv4_3_norm_mbox_loc_reshape = Reshape(
(-1, 4), name='conv4_3_norm_mbox_loc_reshape')(conv4_3_norm_mbox_loc)
fc7_mbox_loc_reshape = Reshape((-1, 4),
name='fc7_mbox_loc_reshape')(fc7_mbox_loc)
conv6_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv6_2_mbox_loc_reshape')(conv6_2_mbox_loc)
conv7_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv7_2_mbox_loc_reshape')(conv7_2_mbox_loc)
conv8_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv8_2_mbox_loc_reshape')(conv8_2_mbox_loc)
conv9_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv9_2_mbox_loc_reshape')(conv9_2_mbox_loc)
conv10_2_mbox_loc_reshape = Reshape(
(-1, 4), name='conv10_2_mbox_loc_reshape')(conv10_2_mbox_loc)
# Reshape the anchor box tensors, yielding 3D tensors of shape `(batch, height * width * n_boxes, 8)`
conv4_3_norm_mbox_priorbox_reshape = Reshape(
(-1, 8),
name='conv4_3_norm_mbox_priorbox_reshape')(conv4_3_norm_mbox_priorbox)
fc7_mbox_priorbox_reshape = Reshape(
(-1, 8), name='fc7_mbox_priorbox_reshape')(fc7_mbox_priorbox)
conv6_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv6_2_mbox_priorbox_reshape')(conv6_2_mbox_priorbox)
conv7_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv7_2_mbox_priorbox_reshape')(conv7_2_mbox_priorbox)
conv8_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv8_2_mbox_priorbox_reshape')(conv8_2_mbox_priorbox)
conv9_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv9_2_mbox_priorbox_reshape')(conv9_2_mbox_priorbox)
conv10_2_mbox_priorbox_reshape = Reshape(
(-1, 8), name='conv10_2_mbox_priorbox_reshape')(conv10_2_mbox_priorbox)
### Concatenate the predictions from the different layers
# Axis 0 (batch) and axis 2 (n_classes or 4, respectively) are identical for all layer predictions,
# so we want to concatenate along axis 1, the number of boxes per layer
# Output shape of `mbox_conf`: (batch, n_boxes_total, n_classes)
mbox_conf = Concatenate(axis=1, name='mbox_conf')([
conv4_3_norm_mbox_conf_reshape, fc7_mbox_conf_reshape,
conv6_2_mbox_conf_reshape, conv7_2_mbox_conf_reshape,
conv8_2_mbox_conf_reshape, conv9_2_mbox_conf_reshape,
conv10_2_mbox_conf_reshape
])
# Output shape of `mbox_loc`: (batch, n_boxes_total, 4)
mbox_loc = Concatenate(axis=1, name='mbox_loc')([
conv4_3_norm_mbox_loc_reshape, fc7_mbox_loc_reshape,
conv6_2_mbox_loc_reshape, conv7_2_mbox_loc_reshape,
conv8_2_mbox_loc_reshape, conv9_2_mbox_loc_reshape,
conv10_2_mbox_loc_reshape
])
# Output shape of `mbox_priorbox`: (batch, n_boxes_total, 8)
mbox_priorbox = Concatenate(axis=1, name='mbox_priorbox')([
conv4_3_norm_mbox_priorbox_reshape, fc7_mbox_priorbox_reshape,
conv6_2_mbox_priorbox_reshape, conv7_2_mbox_priorbox_reshape,
conv8_2_mbox_priorbox_reshape, conv9_2_mbox_priorbox_reshape,
conv10_2_mbox_priorbox_reshape
])
# The box coordinate predictions will go into the loss function just the way they are,
# but for the class predictions, we'll apply a softmax activation layer first
mbox_conf_softmax = Activation('softmax',
name='mbox_conf_softmax')(mbox_conf)
# Concatenate the class and box predictions and the anchors to one large predictions vector
# Output shape of `predictions`: (batch, n_boxes_total, n_classes + 4 + 8)
predictions = Concatenate(axis=2, name='predictions')(
[mbox_conf_softmax, mbox_loc, mbox_priorbox])
model = Model(inputs=x, outputs=predictions)
if return_predictor_sizes:
# Get the spatial dimensions (height, width) of the predictor conv layers, we need them to
# be able to generate the default boxes for the matching process outside of the model during training.
# Note that the original implementation performs anchor box matching inside the loss function. We don't do that.
# Instead, we'll do it in the batch generator function.
# The spatial dimensions are the same for the confidence and localization predictors, so we just take those of the conf layers.
predictor_sizes = np.array([
conv4_3_norm_mbox_conf._keras_shape[1:3],
fc7_mbox_conf._keras_shape[1:3],
conv6_2_mbox_conf._keras_shape[1:3],
conv7_2_mbox_conf._keras_shape[1:3],
conv8_2_mbox_conf._keras_shape[1:3],
conv9_2_mbox_conf._keras_shape[1:3],
conv10_2_mbox_conf._keras_shape[1:3]
])
return model, predictor_sizes
else:
return model
def ssd_300(image_size,
n_classes,
mode='training',
l2_regularization=0.0005,
min_scale=None,
max_scale=None,
scales=None,
aspect_ratios_global=None,
aspect_ratios_per_layer=[[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5, 3.0, 1.0/3.0],
[1.0, 2.0, 0.5],
[1.0, 2.0, 0.5]],
two_boxes_for_ar1=True,
steps=[8, 16, 32, 64, 100, 300],
offsets=None,
clip_boxes=False,
variances=[0.1, 0.1, 0.2, 0.2],
coords='centroids',
normalize_coords=True,
subtract_mean=[123, 117, 104],
divide_by_stddev=None,
swap_channels=[2, 1, 0],
confidence_thresh=0.01,
iou_threshold=0.45,
top_k=200,
nms_max_output_size=400,
return_predictor_sizes=False):
'''
Build a Keras model with SSD300 architecture, see references.
The base network is a reduced atrous VGG-16, extended by the SSD architecture,
as described in the paper.
Most of the arguments that this function takes are only needed for the anchor
box layers. In case you're training the network, the parameters passed here must
be the same as the ones used to set up `SSDBoxEncoder`. In case you're loading
trained weights, the parameters passed here must be the same as the ones used
to produce the trained weights.
Some of these arguments are explained in more detail in the documentation of the
`SSDBoxEncoder` class.
Note: Requires Keras v2.0 or later. Currently works only with the
TensorFlow backend (v1.0 or later).
Arguments:
image_size (tuple): The input image size in the format `(height, width, channels)`.
n_classes (int): The number of positive classes, e.g. 20 for Pascal VOC, 80 for MS COCO.
mode (str, optional): One of 'training', 'inference' and 'inference_fast'. In 'training' mode,
the model outputs the raw prediction tensor, while in 'inference' and 'inference_fast' modes,
the raw predictions are decoded into absolute coordinates and filtered via confidence thresholding,
non-maximum suppression, and top-k filtering. The difference between latter two modes is that
'inference' follows the exact procedure of the original Caffe implementation, while
'inference_fast' uses a faster prediction decoding procedure.
l2_regularization (float, optional): The L2-regularization rate. Applies to all convolutional layers.
Set to zero to deactivate L2-regularization.
min_scale (float, optional): The smallest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images.
max_scale (float, optional): The largest scaling factor for the size of the anchor boxes as a fraction
of the shorter side of the input images. All scaling factors between the smallest and the
largest will be linearly interpolated. Note that the second to last of the linearly interpolated
scaling factors will actually be the scaling factor for the last predictor layer, while the last
scaling factor is used for the second box for aspect ratio 1 in the last predictor layer
if `two_boxes_for_ar1` is `True`.
scales (list, optional): A list of floats containing scaling factors per convolutional predictor layer.
This list must be one element longer than the number of predictor layers. The first `k` elements are the
scaling factors for the `k` predictor layers, while the last element is used for the second box
for aspect ratio 1 in the last predictor layer if `two_boxes_for_ar1` is `True`. This additional
last scaling factor must be passed either way, even if it is not being used. If a list is passed,
this argument overrides `min_scale` and `max_scale`. All scaling factors must be greater than zero.
aspect_ratios_global (list, optional): The list of aspect ratios for which anchor boxes are to be
generated. This list is valid for all prediction layers.
aspect_ratios_per_layer (list, optional): A list containing one aspect ratio list for each prediction layer.
This allows you to set the aspect ratios for each predictor layer individually, which is the case for the
original SSD300 implementation. If a list is passed, it overrides `aspect_ratios_global`.
two_boxes_for_ar1 (bool, optional): Only relevant for aspect ratio lists that contain 1. Will be ignored otherwise.
If `True`, two anchor 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.
steps (list, optional): `None` or a list with as many elements as there are predictor layers. The elements can be
either ints/floats or tuples of two ints/floats. These numbers represent for each predictor layer how many
pixels apart the anchor box center points should be vertically and horizontally along the spatial grid over
the image. If the list contains ints/floats, then that value will be used for both spatial dimensions.
If the list contains tuples of two ints/floats, then they represent `(step_height, step_width)`.
If no steps are provided, then they will be computed such that the anchor box center points will form an
equidistant grid within the image dimensions.
offsets (list, optional): `None` or a list with as many elements as there are predictor layers. The elements can be
either floats or tuples of two floats. These numbers represent for each predictor layer how many
pixels from the top and left boarders of the image the top-most and left-most anchor box center points should be
as a fraction of `steps`. The last bit is important: The offsets are not absolute pixel values, but fractions
of the step size specified in the `steps` argument. If the list contains floats, then that value will
be used for both spatial dimensions. If the list contains tuples of two floats, then they represent
`(vertical_offset, horizontal_offset)`. If no offsets are provided, then they will default to 0.5 of the step size.
clip_boxes (bool, optional): If `True`, clips the anchor box coordinates to stay within image boundaries.
variances (list, optional): A list of 4 floats >0. The anchor box offset for each coordinate will be divided by
its respective variance value.
coords (str, optional): The box coordinate format to be used internally by the model (i.e. this is not the input format
of the ground truth labels). Can be either 'centroids' for the format `(cx, cy, w, h)` (box center coordinates, width,
and height), 'minmax' for the format `(xmin, xmax, ymin, ymax)`, or 'corners' for the format `(xmin, ymin, xmax, ymax)`.
normalize_coords (bool, optional): Set to `True` if the model is supposed to use relative instead of absolute coordinates,
i.e. if the model predicts box coordinates within [0,1] instead of absolute coordinates.
subtract_mean (array-like, optional): `None` or an array-like object of integers or floating point values
of any shape that is broadcast-compatible with the image shape. The elements of this array will be
subtracted from the image pixel intensity values. For example, pass a list of three integers
to perform per-channel mean normalization for color images.
divide_by_stddev (array-like, optional): `None` or an array-like object of non-zero integers or
floating point values of any shape that is broadcast-compatible with the image shape. The image pixel
intensity values will be divided by the elements of this array. For example, pass a list
of three integers to perform per-channel standard deviation normalization for color images.
swap_channels (list, optional): Either `False` or a list of integers representing the desired order in which the input
image channels should be swapped.
confidence_thresh (float, optional): A float in [0,1), the minimum classification confidence in a specific
positive class in order to be considered for the non-maximum suppression stage for the respective class.
A lower value will result in a larger part of the selection process being done by the non-maximum suppression
stage, while a larger value will result in a larger part of the selection process happening in the confidence
thresholding stage.
iou_threshold (float, optional): A float in [0,1]. All boxes that have a Jaccard similarity of greater than `iou_threshold`
with a locally maximal box will be removed from the set of predictions for a given class, where 'maximal' refers
to the box's confidence score.
top_k (int, optional): The number of highest scoring predictions to be kept for each batch item after the
non-maximum suppression stage.
nms_max_output_size (int, optional): The maximal number of predictions that will be left over after the NMS stage.
return_predictor_sizes (bool, optional): If `True`, this function not only returns the model, but also
a list containing the spatial dimensions of the predictor layers. This isn't strictly necessary since
you can always get their sizes easily via the Keras API, but it's convenient and less error-prone
to get them this way. They are only relevant for training anyway (SSDBoxEncoder needs to know the
spatial dimensions of the predictor layers), for inference you don't need them.
Returns:
model: The Keras SSD300 model.
predictor_sizes (optional): A Numpy array containing the `(height, width)` portion
of the output tensor shape for each convolutional predictor layer. During
training, the generator function needs this in order to transform
the ground truth labels into tensors of identical structure as the
output tensors of the model, which is in turn needed for the cost
function.
References:
https://arxiv.org/abs/1512.02325v5
'''
n_predictor_layers = 6 # The number of predictor conv layers in the network is 6 for the original SSD300.
n_classes += 1 # Account for the background class.
l2_reg = l2_regularization # Make the internal name shorter.
img_height, img_width, img_channels = image_size[0], image_size[1], image_size[2]
############################################################################
# Get a few exceptions out of the way.
############################################################################
if aspect_ratios_global is None and aspect_ratios_per_layer is None:
raise ValueError("`aspect_ratios_global` and `aspect_ratios_per_layer` cannot both be None. At least one needs to be specified.")
if aspect_ratios_per_layer:
if len(aspect_ratios_per_layer) != n_predictor_layers:
raise ValueError("It must be either aspect_ratios_per_layer is None or len(aspect_ratios_per_layer) == {}, but len(aspect_ratios_per_layer) == {}.".format(n_predictor_layers, len(aspect_ratios_per_layer)))
if (min_scale is None or max_scale is None) and scales is None:
raise ValueError("Either `min_scale` and `max_scale` or `scales` need to be specified.")
if scales:
if len(scales) != n_predictor_layers+1:
raise ValueError("It must be either scales is None or len(scales) == {}, but len(scales) == {}.".format(n_predictor_layers+1, len(scales)))
else: # If no explicit list of scaling factors was passed, compute the list of scaling factors from `min_scale` and `max_scale`
scales = np.linspace(min_scale, max_scale, n_predictor_layers+1)
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))
if (not (steps is None)) and (len(steps) != n_predictor_layers):
raise ValueError("You must provide at least one step value per predictor layer.")
if (not (offsets is None)) and (len(offsets) != n_predictor_layers):
raise ValueError("You must provide at least one offset value per predictor layer.")
############################################################################
# Compute the anchor box parameters.
############################################################################
# Set the aspect ratios for each predictor layer. These are only needed for the anchor box layers.
if aspect_ratios_per_layer:
aspect_ratios = aspect_ratios_per_layer
else:
aspect_ratios = [aspect_ratios_global] * n_predictor_layers
# Compute the number of boxes to be predicted per cell for each predictor layer.
# We need this so that we know how many channels the predictor layers need to have.
if aspect_ratios_per_layer:
n_boxes = []
for ar in aspect_ratios_per_layer:
if (1 in ar) & two_boxes_for_ar1:
n_boxes.append(len(ar) + 1) # +1 for the second box for aspect ratio 1
else:
n_boxes.append(len(ar))
else: # If only a global aspect ratio list was passed, then the number of boxes is the same for each predictor layer
if (1 in aspect_ratios_global) & two_boxes_for_ar1:
n_boxes = len(aspect_ratios_global) + 1
else:
n_boxes = len(aspect_ratios_global)
n_boxes = [n_boxes] * n_predictor_layers
if steps is None:
steps = [None] * n_predictor_layers
if offsets is None:
offsets = [None] * n_predictor_layers
############################################################################
# Define functions for the Lambda layers below.
############################################################################
def identity_layer(tensor):
return tensor
def input_mean_normalization(tensor):
return tensor - np.array(subtract_mean)
def input_stddev_normalization(tensor):
return tensor / np.array(divide_by_stddev)
def input_channel_swap(tensor):
if len(swap_channels) == 3:
return K.stack([tensor[...,swap_channels[0]], tensor[...,swap_channels[1]], tensor[...,swap_channels[2]]], axis=-1)
elif len(swap_channels) == 4:
return K.stack([tensor[...,swap_channels[0]], tensor[...,swap_channels[1]], tensor[...,swap_channels[2]], tensor[...,swap_channels[3]]], axis=-1)
############################################################################
# Build the network.
############################################################################
x = Input(shape=(img_height, img_width, img_channels))
# The following identity layer is only needed so that the subsequent lambda layers can be optional.
x1 = Lambda(identity_layer, output_shape=(img_height, img_width, img_channels), name='identity_layer')(x)
if not (subtract_mean is None):
x1 = Lambda(input_mean_normalization, output_shape=(img_height, img_width, img_channels), name='input_mean_normalization')(x1)
if not (divide_by_stddev is None):
x1 = Lambda(input_stddev_normalization, output_shape=(img_height, img_width, img_channels), name='input_stddev_normalization')(x1)
if swap_channels:
x1 = Lambda(input_channel_swap, output_shape=(img_height, img_width, img_channels), name='input_channel_swap')(x1)
conv1_1 = Conv2D(64, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv1_1')(x1)
conv1_2 = Conv2D(64, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv1_2')(conv1_1)
pool1 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same', name='pool1')(conv1_2)
conv2_1 = Conv2D(128, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv2_1')(pool1)
conv2_2 = Conv2D(128, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv2_2')(conv2_1)
pool2 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same', name='pool2')(conv2_2)
conv3_1 = Conv2D(256, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv3_1')(pool2)
conv3_2 = Conv2D(256, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv3_2')(conv3_1)
conv3_3 = Conv2D(256, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv3_3')(conv3_2)
pool3 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same', name='pool3')(conv3_3)
conv4_1 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv4_1')(pool3)
conv4_2 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv4_2')(conv4_1)
conv4_3 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv4_3')(conv4_2)
pool4 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same', name='pool4')(conv4_3)
conv5_1 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv5_1')(pool4)
conv5_2 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv5_2')(conv5_1)
conv5_3 = Conv2D(512, (3, 3), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv5_3')(conv5_2)
pool5 = MaxPooling2D(pool_size=(3, 3), strides=(1, 1), padding='same', name='pool5')(conv5_3)
fc6 = Conv2D(1024, (3, 3), dilation_rate=(6, 6), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc6')(pool5)
fc7 = Conv2D(1024, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc7')(fc6)
conv6_1 = Conv2D(256, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv6_1')(fc7)
conv6_1 = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv6_padding')(conv6_1)
conv6_2 = Conv2D(512, (3, 3), strides=(2, 2), activation='relu', padding='valid', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv6_2')(conv6_1)
conv7_1 = Conv2D(128, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv7_1')(conv6_2)
conv7_1 = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv7_padding')(conv7_1)
conv7_2 = Conv2D(256, (3, 3), strides=(2, 2), activation='relu', padding='valid', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv7_2')(conv7_1)
conv8_1 = Conv2D(128, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv8_1')(conv7_2)
conv8_2 = Conv2D(256, (3, 3), strides=(1, 1), activation='relu', padding='valid', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv8_2')(conv8_1)
conv9_1 = Conv2D(128, (1, 1), activation='relu', padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv9_1')(conv8_2)
conv9_2 = Conv2D(256, (3, 3), strides=(1, 1), activation='relu', padding='valid', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv9_2')(conv9_1)
# Feed conv4_3 into the L2 normalization layer
conv4_3_norm = L2Normalization(gamma_init=20, name='conv4_3_norm')(conv4_3)
### Build the convolutional predictor layers on top of the base network
# We precidt `n_classes` confidence values for each box, hence the confidence predictors have depth `n_boxes * n_classes`
# Output shape of the confidence layers: `(batch, height, width, n_boxes * n_classes)`
conv4_3_norm_mbox_conf = Conv2D(n_boxes[0] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv4_3_norm_mbox_conf')(conv4_3_norm)
fc7_mbox_conf = Conv2D(n_boxes[1] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc7_mbox_conf')(fc7)
conv6_2_mbox_conf = Conv2D(n_boxes[2] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv6_2_mbox_conf')(conv6_2)
conv7_2_mbox_conf = Conv2D(n_boxes[3] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv7_2_mbox_conf')(conv7_2)
conv8_2_mbox_conf = Conv2D(n_boxes[4] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv8_2_mbox_conf')(conv8_2)
conv9_2_mbox_conf = Conv2D(n_boxes[5] * n_classes, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv9_2_mbox_conf')(conv9_2)
# We predict 4 box coordinates for each box, hence the localization predictors have depth `n_boxes * 4`
# Output shape of the localization layers: `(batch, height, width, n_boxes * 4)`
conv4_3_norm_mbox_loc = Conv2D(n_boxes[0] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv4_3_norm_mbox_loc')(conv4_3_norm)
fc7_mbox_loc = Conv2D(n_boxes[1] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='fc7_mbox_loc')(fc7)
conv6_2_mbox_loc = Conv2D(n_boxes[2] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv6_2_mbox_loc')(conv6_2)
conv7_2_mbox_loc = Conv2D(n_boxes[3] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv7_2_mbox_loc')(conv7_2)
conv8_2_mbox_loc = Conv2D(n_boxes[4] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv8_2_mbox_loc')(conv8_2)
conv9_2_mbox_loc = Conv2D(n_boxes[5] * 4, (3, 3), padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(l2_reg), name='conv9_2_mbox_loc')(conv9_2)
### Generate the anchor boxes (called "priors" in the original Caffe/C++ implementation, so I'll keep their layer names)
# Output shape of anchors: `(batch, height, width, n_boxes, 8)`
conv4_3_norm_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[0], next_scale=scales[1], aspect_ratios=aspect_ratios[0],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[0], this_offsets=offsets[0], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='conv4_3_norm_mbox_priorbox')(conv4_3_norm_mbox_loc)
fc7_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[1], next_scale=scales[2], aspect_ratios=aspect_ratios[1],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[1], this_offsets=offsets[1], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='fc7_mbox_priorbox')(fc7_mbox_loc)
conv6_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[2], next_scale=scales[3], aspect_ratios=aspect_ratios[2],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[2], this_offsets=offsets[2], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='conv6_2_mbox_priorbox')(conv6_2_mbox_loc)
conv7_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[3], next_scale=scales[4], aspect_ratios=aspect_ratios[3],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[3], this_offsets=offsets[3], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='conv7_2_mbox_priorbox')(conv7_2_mbox_loc)
conv8_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[4], next_scale=scales[5], aspect_ratios=aspect_ratios[4],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[4], this_offsets=offsets[4], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='conv8_2_mbox_priorbox')(conv8_2_mbox_loc)
conv9_2_mbox_priorbox = AnchorBoxes(img_height, img_width, this_scale=scales[5], next_scale=scales[6], aspect_ratios=aspect_ratios[5],
two_boxes_for_ar1=two_boxes_for_ar1, this_steps=steps[5], this_offsets=offsets[5], clip_boxes=clip_boxes,
variances=variances, coords=coords, normalize_coords=normalize_coords, name='conv9_2_mbox_priorbox')(conv9_2_mbox_loc)
### Reshape
# Reshape the class predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, n_classes)`
# We want the classes isolated in the last axis to perform softmax on them
conv4_3_norm_mbox_conf_reshape = Reshape((-1, n_classes), name='conv4_3_norm_mbox_conf_reshape')(conv4_3_norm_mbox_conf)
fc7_mbox_conf_reshape = Reshape((-1, n_classes), name='fc7_mbox_conf_reshape')(fc7_mbox_conf)
conv6_2_mbox_conf_reshape = Reshape((-1, n_classes), name='conv6_2_mbox_conf_reshape')(conv6_2_mbox_conf)
conv7_2_mbox_conf_reshape = Reshape((-1, n_classes), name='conv7_2_mbox_conf_reshape')(conv7_2_mbox_conf)
conv8_2_mbox_conf_reshape = Reshape((-1, n_classes), name='conv8_2_mbox_conf_reshape')(conv8_2_mbox_conf)
conv9_2_mbox_conf_reshape = Reshape((-1, n_classes), name='conv9_2_mbox_conf_reshape')(conv9_2_mbox_conf)
# Reshape the box predictions, yielding 3D tensors of shape `(batch, height * width * n_boxes, 4)`
# We want the four box coordinates isolated in the last axis to compute the smooth L1 loss
conv4_3_norm_mbox_loc_reshape = Reshape((-1, 4), name='conv4_3_norm_mbox_loc_reshape')(conv4_3_norm_mbox_loc)
fc7_mbox_loc_reshape = Reshape((-1, 4), name='fc7_mbox_loc_reshape')(fc7_mbox_loc)
conv6_2_mbox_loc_reshape = Reshape((-1, 4), name='conv6_2_mbox_loc_reshape')(conv6_2_mbox_loc)
conv7_2_mbox_loc_reshape = Reshape((-1, 4), name='conv7_2_mbox_loc_reshape')(conv7_2_mbox_loc)
conv8_2_mbox_loc_reshape = Reshape((-1, 4), name='conv8_2_mbox_loc_reshape')(conv8_2_mbox_loc)
conv9_2_mbox_loc_reshape = Reshape((-1, 4), name='conv9_2_mbox_loc_reshape')(conv9_2_mbox_loc)
# Reshape the anchor box tensors, yielding 3D tensors of shape `(batch, height * width * n_boxes, 8)`
conv4_3_norm_mbox_priorbox_reshape = Reshape((-1, 8), name='conv4_3_norm_mbox_priorbox_reshape')(conv4_3_norm_mbox_priorbox)
fc7_mbox_priorbox_reshape = Reshape((-1, 8), name='fc7_mbox_priorbox_reshape')(fc7_mbox_priorbox)
conv6_2_mbox_priorbox_reshape = Reshape((-1, 8), name='conv6_2_mbox_priorbox_reshape')(conv6_2_mbox_priorbox)
conv7_2_mbox_priorbox_reshape = Reshape((-1, 8), name='conv7_2_mbox_priorbox_reshape')(conv7_2_mbox_priorbox)
conv8_2_mbox_priorbox_reshape = Reshape((-1, 8), name='conv8_2_mbox_priorbox_reshape')(conv8_2_mbox_priorbox)
conv9_2_mbox_priorbox_reshape = Reshape((-1, 8), name='conv9_2_mbox_priorbox_reshape')(conv9_2_mbox_priorbox)
### Concatenate the predictions from the different layers
# Axis 0 (batch) and axis 2 (n_classes or 4, respectively) are identical for all layer predictions,
# so we want to concatenate along axis 1, the number of boxes per layer
# Output shape of `mbox_conf`: (batch, n_boxes_total, n_classes)
mbox_conf = Concatenate(axis=1, name='mbox_conf')([conv4_3_norm_mbox_conf_reshape,
fc7_mbox_conf_reshape,
conv6_2_mbox_conf_reshape,
conv7_2_mbox_conf_reshape,
conv8_2_mbox_conf_reshape,
conv9_2_mbox_conf_reshape])
# Output shape of `mbox_loc`: (batch, n_boxes_total, 4)
mbox_loc = Concatenate(axis=1, name='mbox_loc')([conv4_3_norm_mbox_loc_reshape,
fc7_mbox_loc_reshape,
conv6_2_mbox_loc_reshape,
conv7_2_mbox_loc_reshape,
conv8_2_mbox_loc_reshape,
conv9_2_mbox_loc_reshape])
# Output shape of `mbox_priorbox`: (batch, n_boxes_total, 8)
mbox_priorbox = Concatenate(axis=1, name='mbox_priorbox')([conv4_3_norm_mbox_priorbox_reshape,
fc7_mbox_priorbox_reshape,
conv6_2_mbox_priorbox_reshape,
conv7_2_mbox_priorbox_reshape,
conv8_2_mbox_priorbox_reshape,
conv9_2_mbox_priorbox_reshape])
# The box coordinate predictions will go into the loss function just the way they are,
# but for the class predictions, we'll apply a softmax activation layer first
mbox_conf_softmax = Activation('softmax', name='mbox_conf_softmax')(mbox_conf)
# Concatenate the class and box predictions and the anchors to one large predictions vector
# Output shape of `predictions`: (batch, n_boxes_total, n_classes + 4 + 8)
predictions = Concatenate(axis=2, name='predictions')([mbox_conf_softmax, mbox_loc, mbox_priorbox])
if mode == 'training':
model = Model(inputs=x, outputs=predictions)
elif mode == 'inference':
decoded_predictions = DecodeDetections(confidence_thresh=confidence_thresh,
iou_threshold=iou_threshold,
top_k=top_k,
nms_max_output_size=nms_max_output_size,
coords=coords,
normalize_coords=normalize_coords,
img_height=img_height,
img_width=img_width,
name='decoded_predictions')(predictions)
model = Model(inputs=x, outputs=decoded_predictions)
elif mode == 'inference_fast':
decoded_predictions = DecodeDetectionsFast(confidence_thresh=confidence_thresh,
iou_threshold=iou_threshold,
top_k=top_k,
nms_max_output_size=nms_max_output_size,
coords=coords,
normalize_coords=normalize_coords,
img_height=img_height,
img_width=img_width,
name='decoded_predictions')(predictions)
model = Model(inputs=x, outputs=decoded_predictions)
else:
raise ValueError("`mode` must be one of 'training', 'inference' or 'inference_fast', but received '{}'.".format(mode))
if return_predictor_sizes:
predictor_sizes = np.array([conv4_3_norm_mbox_conf._keras_shape[1:3],
fc7_mbox_conf._keras_shape[1:3],
conv6_2_mbox_conf._keras_shape[1:3],
conv7_2_mbox_conf._keras_shape[1:3],
conv8_2_mbox_conf._keras_shape[1:3],
conv9_2_mbox_conf._keras_shape[1:3]])
return model, predictor_sizes
else:
return model