def interp(w, i, channel_dim):
'''
Input:
w: A 4D block tensor of shape (n, h, w, c)
i: A list of 3-tuples [(x_1, y_1, z_1), (x_2, y_2, z_2), ...],
each having type (int, float, float)
The 4D block represents a batch of 3D image feature volumes with c channels.
The input i is a list of points to index into w via interpolation. Direct
indexing is not possible due to y_1 and z_1 being float values.
Output:
A list of the values: [
w[x_1, y_1, z_1, :]
w[x_2, y_2, z_2, :]
...
w[x_k, y_k, z_k, :]
]
of the same length == len(i)
'''
w_as_vector = tf.reshape(w,
[-1, channel_dim]) # gather expects w to be 1-d
upper_l = tf.to_int32(
tf_concat(
1, [i[:, 0:1], tf.floor(i[:, 1:2]),
tf.floor(i[:, 2:3])]))
upper_r = tf.to_int32(
tf_concat(
1, [i[:, 0:1], tf.floor(i[:, 1:2]),
tf.ceil(i[:, 2:3])]))
lower_l = tf.to_int32(
tf_concat(
1, [i[:, 0:1], tf.ceil(i[:, 1:2]),
tf.floor(i[:, 2:3])]))
lower_r = tf.to_int32(
tf_concat(
1, [i[:, 0:1], tf.ceil(i[:, 1:2]),
tf.ceil(i[:, 2:3])]))
upper_l_idx = to_idx(upper_l, tf.shape(w))
upper_r_idx = to_idx(upper_r, tf.shape(w))
lower_l_idx = to_idx(lower_l, tf.shape(w))
lower_r_idx = to_idx(lower_r, tf.shape(w))
upper_l_value = tf.gather(w_as_vector, upper_l_idx)
upper_r_value = tf.gather(w_as_vector, upper_r_idx)
lower_l_value = tf.gather(w_as_vector, lower_l_idx)
lower_r_value = tf.gather(w_as_vector, lower_r_idx)
alpha_lr = tf.expand_dims(i[:, 2] - tf.floor(i[:, 2]), 1)
alpha_ud = tf.expand_dims(i[:, 1] - tf.floor(i[:, 1]), 1)
upper_value = (1 - alpha_lr) * upper_l_value + (alpha_lr) * upper_r_value
lower_value = (1 - alpha_lr) * lower_l_value + (alpha_lr) * lower_r_value
value = (1 - alpha_ud) * upper_value + (alpha_ud) * lower_value
return value
def to_x1y1x2y2(box):
w = tf.maximum(box[:, 2:3], 1)
h = tf.maximum(box[:, 3:4], 1)
x1 = box[:, 0:1] - w / 2
x2 = box[:, 0:1] + w / 2
y1 = box[:, 1:2] - h / 2
y2 = box[:, 1:2] + h / 2
return tf_concat(1, [x1, y1, x2, y2])
def to_x1y1x2y2(box):
w = tf.maximum(box[:, 2:3], 1)
h = tf.maximum(box[:, 3:4], 1)
x1 = box[:, 0:1] - w / 2
x2 = box[:, 0:1] + w / 2
y1 = box[:, 1:2] - h / 2
y2 = box[:, 1:2] + h / 2
return tf_concat(1, [x1, y1, x2, y2])
def to_x1y1x2y2(box):
w = tf.maximum(box[:, 2:3], 1)
h = tf.maximum(box[:, 3:4], 1)
x1 = box[:, 0:1] - old_div(w, 2)
x2 = box[:, 0:1] + old_div(w, 2)
y1 = box[:, 1:2] - old_div(h, 2)
y2 = box[:, 1:2] + old_div(h, 2)
return tf_concat(1, [x1, y1, x2, y2])
def interp(w, i, channel_dim):
'''
Input:
w: A 4D block tensor of shape (n, h, w, c)
i: A list of 3-tuples [(x_1, y_1, z_1), (x_2, y_2, z_2), ...],
each having type (int, float, float)
The 4D block represents a batch of 3D image feature volumes with c channels.
The input i is a list of points to index into w via interpolation. Direct
indexing is not possible due to y_1 and z_1 being float values.
Output:
A list of the values: [
w[x_1, y_1, z_1, :]
w[x_2, y_2, z_2, :]
...
w[x_k, y_k, z_k, :]
]
of the same length == len(i)
'''
w_as_vector = tf.reshape(w, [-1, channel_dim]) # gather expects w to be 1-d
upper_l = tf.to_int32(tf_concat(1, [i[:, 0:1], tf.floor(i[:, 1:2]), tf.floor(i[:, 2:3])]))
upper_r = tf.to_int32(tf_concat(1, [i[:, 0:1], tf.floor(i[:, 1:2]), tf.ceil(i[:, 2:3])]))
lower_l = tf.to_int32(tf_concat(1, [i[:, 0:1], tf.ceil(i[:, 1:2]), tf.floor(i[:, 2:3])]))
lower_r = tf.to_int32(tf_concat(1, [i[:, 0:1], tf.ceil(i[:, 1:2]), tf.ceil(i[:, 2:3])]))
upper_l_idx = to_idx(upper_l, tf.shape(w))
upper_r_idx = to_idx(upper_r, tf.shape(w))
lower_l_idx = to_idx(lower_l, tf.shape(w))
lower_r_idx = to_idx(lower_r, tf.shape(w))
upper_l_value = tf.gather(w_as_vector, upper_l_idx)
upper_r_value = tf.gather(w_as_vector, upper_r_idx)
lower_l_value = tf.gather(w_as_vector, lower_l_idx)
lower_r_value = tf.gather(w_as_vector, lower_r_idx)
alpha_lr = tf.expand_dims(i[:, 2] - tf.floor(i[:, 2]), 1)
alpha_ud = tf.expand_dims(i[:, 1] - tf.floor(i[:, 1]), 1)
upper_value = (1 - alpha_lr) * upper_l_value + (alpha_lr) * upper_r_value
lower_value = (1 - alpha_lr) * lower_l_value + (alpha_lr) * lower_r_value
value = (1 - alpha_ud) * upper_value + (alpha_ud) * lower_value
return value
def bilinear_select(H, pred_boxes, early_feat, early_feat_channels, w_offset,
h_offset):
'''
Function used for rezooming high level feature maps. Uses bilinear interpolation
to select all channels at index (x, y) for a high level feature map, where x and y are floats.
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
fine_stride = 8. # pixels per 60x80 grid cell in 480x640 image
coarse_stride = H[
'region_size'] # pixels per 15x20 grid cell in 480x640 image
batch_ids = []
x_offsets = []
y_offsets = []
for n in range(H['batch_size']):
for i in range(H['grid_height']):
for j in range(H['grid_width']):
for k in range(H['rnn_len']):
batch_ids.append([n])
x_offsets.append(
[old_div(coarse_stride, 2.) + coarse_stride * j])
y_offsets.append(
[old_div(coarse_stride, 2.) + coarse_stride * i])
batch_ids = tf.constant(batch_ids)
x_offsets = tf.constant(x_offsets)
y_offsets = tf.constant(y_offsets)
pred_boxes_r = tf.reshape(pred_boxes, [outer_size * H['rnn_len'], 4])
scale_factor = old_div(
coarse_stride,
fine_stride) # scale difference between 15x20 and 60x80 features
pred_x_center = old_div(
(pred_boxes_r[:, 0:1] + w_offset * pred_boxes_r[:, 2:3] + x_offsets),
fine_stride)
pred_x_center_clip = tf.clip_by_value(pred_x_center, 0,
scale_factor * H['grid_width'] - 1)
pred_y_center = old_div(
(pred_boxes_r[:, 1:2] + h_offset * pred_boxes_r[:, 3:4] + y_offsets),
fine_stride)
pred_y_center_clip = tf.clip_by_value(pred_y_center, 0,
scale_factor * H['grid_height'] - 1)
interp_indices = tf_concat(
1, [tf.to_float(batch_ids), pred_y_center_clip, pred_x_center_clip])
return interp_indices
def rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets, h_offsets):
'''
Rezoom into a feature map at multiple interpolation points in a grid.
If the predicted object center is at X, len(w_offsets) == 3, and len(h_offsets) == 5,
the rezoom grid will look as follows:
[o o o]
[o o o]
[o X o]
[o o o]
[o o o]
Where each letter indexes into the feature map with bilinear interpolation
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(train_utils.bilinear_select(H,
pred_boxes,
early_feat,
early_feat_channels,
w_offset, h_offset))
interp_indices = tf_concat(0, indices)
rezoom_features = train_utils.interp(early_feat,
interp_indices,
early_feat_channels)
rezoom_features_r = tf.reshape(rezoom_features,
[len(w_offsets) * len(h_offsets),
outer_size,
H['rnn_len'],
early_feat_channels])
rezoom_features_t = tf.transpose(rezoom_features_r, [1, 2, 0, 3])
return tf.reshape(rezoom_features_t,
[outer_size,
H['rnn_len'],
len(w_offsets) * len(h_offsets) * early_feat_channels])
def rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets, h_offsets):
'''
Rezoom into a feature map at multiple interpolation points in a grid.
If the predicted object center is at X, len(w_offsets) == 3, and len(h_offsets) == 5,
the rezoom grid will look as follows:
[o o o]
[o o o]
[o X o]
[o o o]
[o o o]
Where each letter indexes into the feature map with bilinear interpolation
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
indices = []
for w_offset in w_offsets:
for h_offset in h_offsets:
indices.append(train_utils.bilinear_select(H,
pred_boxes,
early_feat,
early_feat_channels,
w_offset, h_offset))
interp_indices = tf_concat(0, indices)
rezoom_features = train_utils.interp(early_feat,
interp_indices,
early_feat_channels)
rezoom_features_r = tf.reshape(rezoom_features,
[len(w_offsets) * len(h_offsets),
outer_size,
H['rnn_len'],
early_feat_channels])
rezoom_features_t = tf.transpose(rezoom_features_r, [1, 2, 0, 3])
return tf.reshape(rezoom_features_t,
[outer_size,
H['rnn_len'],
len(w_offsets) * len(h_offsets) * early_feat_channels])
def bilinear_select(H, pred_boxes, early_feat, early_feat_channels, w_offset, h_offset):
'''
Function used for rezooming high level feature maps. Uses bilinear interpolation
to select all channels at index (x, y) for a high level feature map, where x and y are floats.
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
fine_stride = 8. # pixels per 60x80 grid cell in 480x640 image
coarse_stride = H['region_size'] # pixels per 15x20 grid cell in 480x640 image
batch_ids = []
x_offsets = []
y_offsets = []
for n in range(H['batch_size']):
for i in range(H['grid_height']):
for j in range(H['grid_width']):
for k in range(H['rnn_len']):
batch_ids.append([n])
x_offsets.append([coarse_stride / 2. + coarse_stride * j])
y_offsets.append([coarse_stride / 2. + coarse_stride * i])
batch_ids = tf.constant(batch_ids)
x_offsets = tf.constant(x_offsets)
y_offsets = tf.constant(y_offsets)
pred_boxes_r = tf.reshape(pred_boxes, [outer_size * H['rnn_len'], 4])
scale_factor = coarse_stride / fine_stride # scale difference between 15x20 and 60x80 features
pred_x_center = (pred_boxes_r[:, 0:1] + w_offset * pred_boxes_r[:, 2:3] + x_offsets) / fine_stride
pred_x_center_clip = tf.clip_by_value(pred_x_center,
0,
scale_factor * H['grid_width'] - 1)
pred_y_center = (pred_boxes_r[:, 1:2] + h_offset * pred_boxes_r[:, 3:4] + y_offsets) / fine_stride
pred_y_center_clip = tf.clip_by_value(pred_y_center,
0,
scale_factor * H['grid_height'] - 1)
interp_indices = tf_concat(1, [tf.to_float(batch_ids), pred_y_center_clip, pred_x_center_clip])
return interp_indices
def build_forward(H, x, phase, reuse):
'''
Construct the forward model
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
input_mean = 117.
x -= input_mean
cnn, early_feat = googlenet_load.model(x, H, reuse)
early_feat_channels = H['early_feat_channels']
early_feat = early_feat[:, :, :, :early_feat_channels]
if H['deconv']:
size = 3
stride = 2
pool_size = 5
with tf.variable_scope("deconv", reuse=reuse):
w = tf.get_variable(
'conv_pool_w',
shape=[
size, size, H['later_feat_channels'],
H['later_feat_channels']
],
initializer=tf.random_normal_initializer(stddev=0.01))
cnn_s = tf.nn.conv2d(cnn,
w,
strides=[1, stride, stride, 1],
padding='SAME')
cnn_s_pool = tf.nn.avg_pool(cnn_s[:, :, :, :256],
ksize=[1, pool_size, pool_size, 1],
strides=[1, 1, 1, 1],
padding='SAME')
cnn_s_with_pool = tf_concat(3, [cnn_s_pool, cnn_s[:, :, :, 256:]])
cnn_deconv = deconv(cnn_s_with_pool,
output_shape=[
H['batch_size'], H['grid_height'],
H['grid_width'], 256
],
channels=[H['later_feat_channels'], 256])
cnn = tf_concat(3, (cnn_deconv, cnn[:, :, :, 256:]))
elif H['avg_pool_size'] > 1:
pool_size = H['avg_pool_size']
cnn1 = cnn[:, :, :, :700]
cnn2 = cnn[:, :, :, 700:]
cnn2 = tf.nn.avg_pool(cnn2,
ksize=[1, pool_size, pool_size, 1],
strides=[1, 1, 1, 1],
padding='SAME')
cnn = tf_concat(3, [cnn1, cnn2])
cnn = tf.reshape(cnn, [
H['batch_size'] * H['grid_width'] * H['grid_height'],
H['later_feat_channels']
])
initializer = tf.random_uniform_initializer(-0.1, 0.1)
with tf.variable_scope('decoder', reuse=reuse, initializer=initializer):
scale_down = 0.01
lstm_input = tf.reshape(
cnn * scale_down,
(H['batch_size'] * grid_size, H['later_feat_channels']))
if H['use_lstm']:
lstm_outputs = build_lstm_inner(H, lstm_input)
else:
lstm_outputs = build_overfeat_inner(H, lstm_input)
pred_boxes = []
pred_logits = []
for k in range(H['rnn_len']):
output = lstm_outputs[k]
if phase == 'train':
output = tf.nn.dropout(output, 0.5)
box_weights = tf.get_variable('box_ip%d' % k,
shape=(H['lstm_size'], 4))
conf_weights = tf.get_variable('conf_ip%d' % k,
shape=(H['lstm_size'],
H['num_classes']))
pred_boxes_step = tf.reshape(
tf.matmul(output, box_weights) * 50, [outer_size, 1, 4])
pred_boxes.append(pred_boxes_step)
pred_logits.append(
tf.reshape(tf.matmul(output, conf_weights),
[outer_size, 1, H['num_classes']]))
pred_boxes = tf_concat(1, pred_boxes)
pred_logits = tf_concat(1, pred_logits)
pred_logits_squash = tf.reshape(
pred_logits, [outer_size * H['rnn_len'], H['num_classes']])
pred_confidences_squash = tf.nn.softmax(pred_logits_squash)
pred_confidences = tf.reshape(
pred_confidences_squash,
[outer_size, H['rnn_len'], H['num_classes']])
if H['use_rezoom']:
pred_confs_deltas = []
pred_boxes_deltas = []
w_offsets = H['rezoom_w_coords']
h_offsets = H['rezoom_h_coords']
num_offsets = len(w_offsets) * len(h_offsets)
rezoom_features = rezoom(H, pred_boxes, early_feat,
early_feat_channels, w_offsets, h_offsets)
if phase == 'train':
rezoom_features = tf.nn.dropout(rezoom_features, 0.5)
for k in range(H['rnn_len']):
delta_features = tf_concat(
1, [lstm_outputs[k], rezoom_features[:, k, :] / 1000.])
dim = 128
delta_weights1 = tf.get_variable(
'delta_ip1%d' % k,
shape=[
H['lstm_size'] + early_feat_channels * num_offsets, dim
])
# TODO: add dropout here ?
ip1 = tf.nn.relu(tf.matmul(delta_features, delta_weights1))
if phase == 'train':
ip1 = tf.nn.dropout(ip1, 0.5)
delta_confs_weights = tf.get_variable(
'delta_ip2%d' % k, shape=[dim, H['num_classes']])
if H['reregress']:
delta_boxes_weights = tf.get_variable('delta_ip_boxes%d' %
k,
shape=[dim, 4])
pred_boxes_deltas.append(
tf.reshape(
tf.matmul(ip1, delta_boxes_weights) * 5,
[outer_size, 1, 4]))
scale = H.get('rezoom_conf_scale', 50)
pred_confs_deltas.append(
tf.reshape(
tf.matmul(ip1, delta_confs_weights) * scale,
[outer_size, 1, H['num_classes']]))
pred_confs_deltas = tf_concat(1, pred_confs_deltas)
if H['reregress']:
pred_boxes_deltas = tf_concat(1, pred_boxes_deltas)
return pred_boxes, pred_logits, pred_confidences, pred_confs_deltas, pred_boxes_deltas
return pred_boxes, pred_logits, pred_confidences
def inception_v1_base(inputs,
final_endpoint='Mixed_5c',
scope='InceptionV1'):
"""Defines the Inception V1 base architecture.
This architecture is defined in:
Going deeper with convolutions
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich.
http://arxiv.org/pdf/1409.4842v1.pdf.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
final_endpoint: specifies the endpoint to construct the network up to. It
can be one of ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c',
'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d', 'Mixed_4e',
'Mixed_4f', 'MaxPool_5a_2x2', 'Mixed_5b', 'Mixed_5c']
scope: Optional variable_scope.
Returns:
A dictionary from components of the network to the corresponding activation.
Raises:
ValueError: if final_endpoint is not set to one of the predefined values.
"""
end_points = {}
with tf.variable_scope(scope, 'InceptionV1', [inputs]):
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
weights_initializer=trunc_normal(0.01)):
with slim.arg_scope([slim.conv2d, slim.max_pool2d],
stride=1, padding='SAME'):
end_point = 'Conv2d_1a_7x7'
net = slim.conv2d(inputs, 64, [7, 7], stride=2, scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'MaxPool_2a_3x3'
net = slim.max_pool2d(net, [3, 3], stride=2, scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Conv2d_2b_1x1'
net = slim.conv2d(net, 64, [1, 1], scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Conv2d_2c_3x3'
net = slim.conv2d(net, 192, [3, 3], scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'MaxPool_3a_3x3'
net = slim.max_pool2d(net, [3, 3], stride=2, scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_3b'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 64, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 96, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 128, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 16, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 32, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 32, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_3c'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 192, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'MaxPool_4a_3x3'
net = slim.max_pool2d(net, [3, 3], stride=2, scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_4b'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 96, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 208, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 16, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 48, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_4c'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 112, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 224, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 24, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 64, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_4d'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 256, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 24, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 64, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_4e'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 112, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 144, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 288, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 64, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_4f'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 256, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 320, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 128, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 128, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'MaxPool_5a_2x2'
net = slim.max_pool2d(net, [2, 2], stride=2, scope=end_point)
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_5b'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 256, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 320, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 128, [3, 3], scope='Conv2d_0a_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 128, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
end_point = 'Mixed_5c'
with tf.variable_scope(end_point):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 384, [1, 1], scope='Conv2d_0a_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 192, [1, 1], scope='Conv2d_0a_1x1')
branch_1 = slim.conv2d(branch_1, 384, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 48, [1, 1], scope='Conv2d_0a_1x1')
branch_2 = slim.conv2d(branch_2, 128, [3, 3], scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_3'):
branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
branch_3 = slim.conv2d(branch_3, 128, [1, 1], scope='Conv2d_0b_1x1')
net = tf_concat(3, [branch_0, branch_1, branch_2, branch_3])
end_points[end_point] = net
if final_endpoint == end_point: return net, end_points
raise ValueError('Unknown final endpoint %s' % final_endpoint)
def build_forward(H, x, phase, reuse):
'''
Construct the forward model
'''
grid_size = H['grid_width'] * H['grid_height']
outer_size = grid_size * H['batch_size']
input_mean = 117.
x -= input_mean
cnn, early_feat = googlenet_load.model(x, H, reuse)
early_feat_channels = H['early_feat_channels']
early_feat = early_feat[:, :, :, :early_feat_channels]
if H['deconv']:
size = 3
stride = 2
pool_size = 5
with tf.variable_scope("deconv", reuse=reuse):
w = tf.get_variable('conv_pool_w', shape=[size, size, H['later_feat_channels'], H['later_feat_channels']],
initializer=tf.random_normal_initializer(stddev=0.01))
cnn_s = tf.nn.conv2d(cnn, w, strides=[1, stride, stride, 1], padding='SAME')
cnn_s_pool = tf.nn.avg_pool(cnn_s[:, :, :, :256], ksize=[1, pool_size, pool_size, 1],
strides=[1, 1, 1, 1], padding='SAME')
cnn_s_with_pool = tf_concat(3, [cnn_s_pool, cnn_s[:, :, :, 256:]])
cnn_deconv = deconv(cnn_s_with_pool, output_shape=[H['batch_size'], H['grid_height'], H['grid_width'], 256], channels=[H['later_feat_channels'], 256])
cnn = tf_concat(3, (cnn_deconv, cnn[:, :, :, 256:]))
elif H['avg_pool_size'] > 1:
pool_size = H['avg_pool_size']
cnn1 = cnn[:, :, :, :700]
cnn2 = cnn[:, :, :, 700:]
cnn2 = tf.nn.avg_pool(cnn2, ksize=[1, pool_size, pool_size, 1],
strides=[1, 1, 1, 1], padding='SAME')
cnn = tf_concat(3, [cnn1, cnn2])
cnn = tf.reshape(cnn,
[H['batch_size'] * H['grid_width'] * H['grid_height'], H['later_feat_channels']])
initializer = tf.random_uniform_initializer(-0.1, 0.1)
with tf.variable_scope('decoder', reuse=reuse, initializer=initializer):
scale_down = 0.01
lstm_input = tf.reshape(cnn * scale_down, (H['batch_size'] * grid_size, H['later_feat_channels']))
if H['use_lstm']:
lstm_outputs = build_lstm_inner(H, lstm_input)
else:
lstm_outputs = build_overfeat_inner(H, lstm_input)
pred_boxes = []
pred_logits = []
for k in range(H['rnn_len']):
output = lstm_outputs[k]
if phase == 'train':
output = tf.nn.dropout(output, 0.5)
box_weights = tf.get_variable('box_ip%d' % k,
shape=(H['lstm_size'], 4))
conf_weights = tf.get_variable('conf_ip%d' % k,
shape=(H['lstm_size'], H['num_classes']))
pred_boxes_step = tf.reshape(tf.matmul(output, box_weights) * 50,
[outer_size, 1, 4])
pred_boxes.append(pred_boxes_step)
pred_logits.append(tf.reshape(tf.matmul(output, conf_weights),
[outer_size, 1, H['num_classes']]))
pred_boxes = tf_concat(1, pred_boxes)
pred_logits = tf_concat(1, pred_logits)
pred_logits_squash = tf.reshape(pred_logits,
[outer_size * H['rnn_len'], H['num_classes']])
pred_confidences_squash = tf.nn.softmax(pred_logits_squash)
pred_confidences = tf.reshape(pred_confidences_squash,
[outer_size, H['rnn_len'], H['num_classes']])
if H['use_rezoom']:
pred_confs_deltas = []
pred_boxes_deltas = []
w_offsets = H['rezoom_w_coords']
h_offsets = H['rezoom_h_coords']
num_offsets = len(w_offsets) * len(h_offsets)
rezoom_features = rezoom(H, pred_boxes, early_feat, early_feat_channels, w_offsets, h_offsets)
if phase == 'train':
rezoom_features = tf.nn.dropout(rezoom_features, 0.5)
for k in range(H['rnn_len']):
delta_features = tf_concat(1, [lstm_outputs[k], rezoom_features[:, k, :] / 1000.])
dim = 128
delta_weights1 = tf.get_variable(
'delta_ip1%d' % k,
shape=[H['lstm_size'] + early_feat_channels * num_offsets, dim])
# TODO: add dropout here ?
ip1 = tf.nn.relu(tf.matmul(delta_features, delta_weights1))
if phase == 'train':
ip1 = tf.nn.dropout(ip1, 0.5)
delta_confs_weights = tf.get_variable(
'delta_ip2%d' % k,
shape=[dim, H['num_classes']])
if H['reregress']:
delta_boxes_weights = tf.get_variable(
'delta_ip_boxes%d' % k,
shape=[dim, 4])
pred_boxes_deltas.append(tf.reshape(tf.matmul(ip1, delta_boxes_weights) * 5,
[outer_size, 1, 4]))
scale = H.get('rezoom_conf_scale', 50)
pred_confs_deltas.append(tf.reshape(tf.matmul(ip1, delta_confs_weights) * scale,
[outer_size, 1, H['num_classes']]))
pred_confs_deltas = tf_concat(1, pred_confs_deltas)
if H['reregress']:
pred_boxes_deltas = tf_concat(1, pred_boxes_deltas)
return pred_boxes, pred_logits, pred_confidences, pred_confs_deltas, pred_boxes_deltas
return pred_boxes, pred_logits, pred_confidences
