def keynn_64(image, pose_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(prefix,
1,
5,
image,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(prefix,
5,
10,
inet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
10,
15,
inet,
3,
256,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
15,
20,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
20,
25,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=False)
vout = net2d.GAP('vGAP', inet)
pred_pose = net2d.FC('pose', vout, pose_dim)
return pred_pose
def prep_sHG_3d_encode_info_v2(image, k3d_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(
prefix,
1,
5,
image,
3,
64,
# int image: shape=(100, 64, 64, 3), out shape=(100, 64, 64, 64)
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(
prefix,
5,
10,
inet,
3,
128, # out shape=(100, 32, 32, 128)
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(
prefix,
10,
15,
inet,
3,
256, # out shape=(100, 16, 16, 256) # good enough for hourglass input.. i guess
is_training=tp,
dropout=0.7,
down_first=True)
hg_input = inet
inet2 = net2d.stack_modern_conv(
prefix,
15,
20,
inet,
3,
512, # out shape=(100, 8, 8, 512)
is_training=tp,
dropout=0.7,
down_first=True)
# check
out_3d = net2d.GAP('vGAP', inet2)
pred3d = net2d.FC('3d', out_3d, k3d_dim)
return hg_input, pred3d
def prep_sHG_3d_encode_info(image, k2d_dim, k3d_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(
prefix,
1,
5,
image,
3,
64,
# int image: shape=(100, 64, 64, 3), out shape=(100, 64, 64, 64)
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(
prefix,
5,
10,
inet,
3,
128, # out shape=(100, 32, 32, 128)
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(
prefix,
10,
15,
inet,
3,
256, # out shape=(100, 16, 16, 256)
is_training=tp,
dropout=0.7,
down_first=True)
# inet = net2d.stack_modern_conv(prefix, 15, 20, inet, 3, 512, # out shape=(100, 8, 8, 512)
# is_training=tp, dropout=0.7, down_first=True)
# check
return inet
def base_64_ds_key_vnn(image, depth, pose_dim, mask_dim, key_dim, reuse_flag,
tp):
n = image.get_shape().as_list()[0]
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(prefix,
1,
5,
image,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(prefix,
5,
10,
inet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
10,
15,
inet,
3,
256,
is_training=tp,
dropout=0.7,
down_first=True)
dnet = net2d.stack_modern_conv(prefix + '_depth',
1,
5,
depth,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
dnet = net2d.stack_modern_conv(prefix + '_depth',
5,
10,
dnet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
dnet = net2d.stack_modern_conv(prefix + '_depth',
10,
15,
dnet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
idnet = tf.concat([inet, dnet], axis=3)
mask_out = tf.reshape(idnet, [n, -1])
pred_mask = net2d.FC('mout', mask_out, mask_dim * 2)
idnet = net2d.stack_modern_conv(prefix,
15,
18,
idnet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True)
vout = net2d.GAP('vGAP', idnet)
pred_bin = net2d.FC('bin', vout, pose_dim)
idnet = net2d.stack_modern_conv(prefix,
18,
22,
idnet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=False)
key_out = net2d.GAP('kGAP', idnet)
pred_key = net2d.FC('kout', key_out, key_dim)
return pred_bin, pred_mask, pred_key
def base_64_key_vnn(image, depth, key_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(prefix,
1,
5,
image,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(prefix,
5,
10,
inet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
10,
15,
inet,
3,
256,
is_training=tp,
dropout=0.7,
down_first=True)
dnet = net2d.stack_modern_conv(prefix + '_depth',
1,
5,
depth,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
dnet = net2d.stack_modern_conv(prefix + '_depth',
5,
10,
dnet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
dnet = net2d.stack_modern_conv(prefix + '_depth',
10,
15,
dnet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
idnet = tf.concat([inet, dnet], axis=3)
idnet = net2d.stack_modern_conv(prefix,
15,
22,
idnet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True)
key_out = net2d.GAP('kGAP', idnet)
pred_key = net2d.FC('kout', key_out, key_dim)
return pred_key
def keyos_64_for_hg(image, key_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(prefix,
1,
5,
image,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(
prefix,
5,
10,
inet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True) # [h, w, c] = [64, 64, 64]
inet = net2d.stack_modern_conv(prefix,
10,
15,
inet,
3,
256,
is_training=tp,
dropout=0.7,
down_first=True) # [16, 16, 256]
inet = net2d.stack_modern_conv(prefix,
15,
18,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True) # [8, 8, 512]
inet1 = tf.image.resize_nearest_neighbor(inet,
size=[64, 64]) # 64, 64, 512
out_os = net2d.GAP('osGAP', inet)
pred_os = net2d.FC('os', out_os, key_dim)
inet = net2d.stack_modern_conv(prefix,
18,
21,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True) # [4, 4, 512]
inet2 = tf.image.resize_nearest_neighbor(inet, size=[64, 64])
out_3d = net2d.GAP('3dGAP', inet)
pred3d = net2d.FC('3d', out_3d, key_dim * 3)
inet = net2d.stack_modern_conv(prefix,
21,
25,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=False) # [4, 4, 512]
inet3 = tf.image.resize_nearest_neighbor(inet, size=[64, 64])
out_2d = net2d.GAP('2dGAP', inet)
pred2d = net2d.FC('2d', out_2d, key_dim * 2)
return pred2d, pred3d, pred_os, tf.add_n([inet1, inet2, inet3])
def keyos_64(image, key_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(prefix,
1,
5,
image,
3,
64,
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(prefix,
5,
10,
inet,
3,
128,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
10,
15,
inet,
3,
256,
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(prefix,
15,
18,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True)
out_os = net2d.GAP('osGAP', inet)
pred_os = net2d.FC('os', out_os, key_dim)
inet = net2d.stack_modern_conv(prefix,
18,
21,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=True)
out_3d = net2d.GAP('3dGAP', inet)
pred3d = net2d.FC('3d', out_3d, key_dim * 3)
inet = net2d.stack_modern_conv(prefix,
21,
25,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=False)
out_2d = net2d.GAP('2dGAP', inet)
pred2d = net2d.FC('2d', out_2d, key_dim * 2)
return pred2d, pred3d, pred_os
def key23d_64(image, k2d_dim, k3d_dim, reuse_flag, tp):
prefix = 'conv'
with tf.variable_scope('Base_CNN', reuse=reuse_flag):
inet = net2d.stack_modern_conv(
prefix,
1,
5,
image,
3,
64, # int image: shape=(100, 64, 64, 3), out shape=(100, 64, 64, 64)
is_training=tp,
dropout=0.7,
down_first=False)
inet = net2d.stack_modern_conv(
prefix,
5,
10,
inet,
3,
128, # out shape=(100, 32, 32, 128)
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(
prefix,
10,
15,
inet,
3,
256, # out shape=(100, 16, 16, 256)
is_training=tp,
dropout=0.7,
down_first=True)
inet = net2d.stack_modern_conv(
prefix,
15,
20,
inet,
3,
512, # out shape=(100, 8, 8, 512)
is_training=tp,
dropout=0.7,
down_first=True)
# check
out_3d = net2d.GAP('vGAP', inet)
pred3d = net2d.FC('3d', out_3d, k3d_dim)
inet = net2d.stack_modern_conv(prefix,
20,
25,
inet,
3,
512,
is_training=tp,
dropout=0.7,
down_first=False)
out_2d = net2d.GAP('vGAP', inet)
pred2d = net2d.FC('2d', out_2d, k2d_dim)
return pred2d, pred3d