def bottleneck_block(data, name, filters, stride=1, dilate=1,
identity_map=True, shortcut_kernel=1, dropout_rate=0.):
branch2a = res_conv_stage(data, '{}_branch2a'.format(name),
filters,
kernel=1,
stride=stride)
branch2b = res_conv_stage(branch2a, '{}_branch2b'.format(name),
filters,
kernel=3,
dilate=dilate)
branch2c = res_conv_stage(branch2b, '{}_branch2c'.format(name),
filters*4,
kernel=1,
has_relu=False,
dropout_rate=dropout_rate)
if identity_map:
return relu(data + branch2c, 'res{}_relu'.format(name))
else:
assert shortcut_kernel in (1, 2)
branch1 = res_conv_stage(data, '{}_branch1'.format(name),
filters*4,
kernel=shortcut_kernel,
stride=stride,
dilate=dilate,
pad=-1 if shortcut_kernel % 2 else 0,
has_relu=False)
return relu(branch1 + branch2c, 'res{}_relu'.format(name))
def conv_stage(data,
names,
filters,
kernel=3,
stride=1,
dilate=1,
pad=-1,
groups=1,
no_bias=True,
dropout_rate=0.):
i = 0
bn1 = bn(data, names[i])
i += 1
relu1 = relu(bn1, names[i])
i += 1
if dropout_rate > 0.:
dropout1 = dropout(relu1, names[i], p=dropout_rate)
i += 1
top = conv(dropout1 if dropout_rate > 0. else relu1,
names[i],
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
return relu1, top
def conv_stage(data,
name,
filters,
kernel=3,
stride=1,
dilate=1,
pad=-1,
groups=1,
no_bias=False,
has_bn=False,
dropout_rate=0.,
has_relu=True):
top = conv(data,
name,
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
if has_bn:
top = bn(top, name='{}_bn'.format(name))
if dropout_rate > 0.0:
top = dropout(top, '{}_do'.format(name), p=dropout_rate)
if has_relu:
top = relu(top, '{}_relu'.format(name))
return top
def res_conv_stage(data,
name,
filters,
kernel=3,
stride=1,
dilate=1,
pad=-1,
groups=1,
no_bias=True,
has_relu=True,
dropout_rate=0.):
top = conv(data,
'res{}'.format(name),
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
top = bn(data=top, name='bn{}'.format(name))
if dropout_rate > 0.0:
top = dropout(top, 'res{}_do'.format(name), p=dropout_rate)
if has_relu:
top = relu(top, 'res{}_relu'.format(name))
return top
def bottleneck_block(data,
name,
filters,
stride=1,
dilate=1,
identity_map=True,
shortcut_kernel=1,
dropout_rate=0.):
branch2a = res_conv_stage(data,
'{}_branch2a'.format(name),
filters,
kernel=1,
stride=stride)
branch2b = res_conv_stage(branch2a,
'{}_branch2b'.format(name),
filters,
kernel=3,
dilate=dilate)
branch2c = res_conv_stage(branch2b,
'{}_branch2c'.format(name),
filters * 4,
kernel=1,
has_relu=False,
dropout_rate=dropout_rate)
if identity_map:
return relu(data + branch2c, 'res{}_relu'.format(name))
else:
assert shortcut_kernel in (1, 2)
branch1 = res_conv_stage(data,
'{}_branch1'.format(name),
filters * 4,
kernel=shortcut_kernel,
stride=stride,
dilate=dilate,
pad=-1 if shortcut_kernel % 2 else 0,
has_relu=False)
return relu(branch1 + branch2c, 'res{}_relu'.format(name))
def res_conv_stage(data, name, filters, kernel=3, stride=1, dilate=1, pad=-1,
groups=1, no_bias=True, has_relu=True, dropout_rate=0.):
top = conv(data, 'res{}'.format(name),
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
top = bn(data=top, name='bn{}'.format(name))
if dropout_rate > 0.0:
top = dropout(top, 'res{}_do'.format(name), p=dropout_rate)
if has_relu:
top = relu(top, 'res{}_relu'.format(name))
return top
def conv_stage(data, name, filters, kernel=3, stride=1, dilate=1, pad=-1,
groups=1, no_bias=False, has_bn=False, dropout_rate=0., has_relu=True):
top = conv(data, name,
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
if has_bn:
top = bn(top, name='{}_bn'.format(name))
if dropout_rate > 0.0:
top = dropout(top, '{}_do'.format(name), p=dropout_rate)
if has_relu:
top = relu(top, '{}_relu'.format(name))
return top
def conv_stage(data, names, filters, kernel=3, stride=1, dilate=1, pad=-1,
groups=1, no_bias=True, dropout_rate=0.):
i = 0
bn1 = bn(data, names[i])
i += 1
relu1 = relu(bn1, names[i])
i += 1
if dropout_rate > 0.:
dropout1 = dropout(relu1, names[i], p=dropout_rate)
i += 1
top = conv(dropout1 if dropout_rate > 0. else relu1, names[i],
filters, kernel=kernel,
stride=stride,
dilate=dilate,
pad=pad,
groups=groups,
no_bias=no_bias)
return relu1, top
def rna_feat(conv1_layers, level_blocks):
'''RNA features'''
def _widen(filters, max_filters, min_filters=-1):
filters = min(max_filters, int(filters * level_blocks[level].width))
if min_filters > 0:
filters = max(min_filters, filters)
return filters
# TODO: these may only work for b33
def _ds_by_pl(data, name, level, dilate):
assert level > 0
level_block = level_blocks[level]
if not level_block.downsample == 'p':
return data, dilate
pool_stride = stride = 2
inc_dilate = level_block.dilate
pad = -1
if inc_dilate:
pool_stride = 1
pad = dilate
print 'Pooling stride: {}, dilate: {}, pad: {}'.format(
pool_stride, dilate, pad)
top = pool(data,
name,
stride=pool_stride,
dilate=dilate,
pad=pad,
pool_type=pool_type)
if inc_dilate:
dilate *= stride
return top, dilate
def _ds_by_cv(data,
name,
filters,
level,
dilate,
kernel=3,
dropout_rate=0.):
assert level > 1
level_block = level_blocks[level - 1]
if not level_block.downsample == 'c':
print 'First block on level {}, dilate: {}'.format(level, dilate)
top = res_block(data,
name,
filters,
kernel=kernel,
dilate=dilate,
identity_map=False,
dropout_rate=dropout_rate)
else:
stride = 2
inc_dilate = level_block.dilate
print 'First block on level {}, stride: {}, dilate: {}'.format(
level, stride, dilate)
top = res_block(data,
name,
filters,
kernel=kernel,
stride=stride,
dilate=dilate,
inc_dilate=inc_dilate,
identity_map=False,
dropout_rate=dropout_rate)
if inc_dilate:
dilate *= stride
return top, dilate
dilate = 1
pool_type = 'max'
crop_size = 224
data = mx.sym.Variable('data')
# 224^2 3
level = 0
print 'Level {}'.format(level)
conv0 = data
# 224^2 3
print conv0.infer_shape(data=(64, 3, crop_size, crop_size))[1]
conv0._set_attr(mirror_stage='True')
level = 1
print 'Level {}'.format(level)
filters = _widen(64, 1024)
res1 = conv(conv0,
'conv1a',
filters,
kernel=conv1_layers[0].kernel,
stride=1,
no_bias=True)
for i, conv1_layer in enumerate(conv1_layers[1:]):
names = [
'conv1b{}_bn'.format(i),
'conv1b{}_relu'.format(i),
'conv1b{}_do'.format(i),
'conv1b{}'.format(i),
]
res1 = conv_stage(res1,
names,
filters,
kernel=conv1_layer.kernel,
stride=1,
no_bias=True)
for i in xrange(level_blocks[level].num):
res1 = res_block(res1, '1b{}'.format(i + 1), filters)
# pass dilate from now on
assert dilate == 1
res1, dilate = _ds_by_pl(res1, 'pool1', level, dilate)
# 112^2 64
print res1.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res1._set_attr(mirror_stage='True')
level = 2
print 'Level {}'.format(level)
filters = _widen(128, 1024)
res2, dilate = _ds_by_cv(res1, '2a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res2 = res_block(res2, '2b{}'.format(i), filters, dilate=dilate)
res2._set_attr(mirror_stage='True')
res2, dilate = _ds_by_pl(res2, 'pool2', level, dilate)
# 56^2 128
print res2.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res2._set_attr(mirror_stage='True')
level = 3
print 'Level {}'.format(level)
filters = _widen(256, 1024)
res3, dilate = _ds_by_cv(res2, '3a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res3 = res_block(res3, '3b{}'.format(i), filters, dilate=dilate)
res3._set_attr(mirror_stage='True')
res3, dilate = _ds_by_pl(res3, 'pool3', level, dilate)
# 28^2 256
print res3.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res3._set_attr(mirror_stage='True')
level = 4
print 'Level {}'.format(level)
filters = _widen(512, 1024)
res4, dilate = _ds_by_cv(res3, '4a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res4 = res_block(res4, '4b{}'.format(i), filters, dilate=dilate)
res4._set_attr(mirror_stage='True')
res4, dilate = _ds_by_pl(res4, 'pool4', level, dilate)
# 14^2 512
print res4.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res4._set_attr(mirror_stage='True')
level = 5
print 'Level {}'.format(level)
# wd = 0.5 : (512, 512)
# wd = 1.0 : (512, 1024)
# wd = 2.0 : (1024, 1024)
filters_l1 = _widen(512, 1024, filters)
filters = [filters_l1, _widen(1024, 1024, filters_l1)]
res5, dilate = _ds_by_cv(res4, '5a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res5 = res_block(res5, '5b{}'.format(i), filters, dilate=dilate)
res5._set_attr(mirror_stage='True')
res5, dilate = _ds_by_pl(res5, 'pool5', level, dilate)
# 7^2 1024
print res5.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res5._set_attr(mirror_stage='True')
level = 6
print 'Level {}'.format(level)
res6, dilate = _ds_by_cv(res5,
'6a', [512, 1024, 2048],
level,
dilate,
kernel=[1, 3, 1],
dropout_rate=level_blocks[level].dropout)
# 7^2 2048
print res6.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res6._set_attr(mirror_stage='True')
level = 7
print 'Level {}'.format(level)
res7 = res_block(res6,
'7a', [1024, 2048, 4096],
kernel=[1, 3, 1],
dilate=dilate,
identity_map=False,
dropout_rate=level_blocks[level].dropout)
# 7^2 4096
print res7.infer_shape(data=(64, 3, crop_size, crop_size))[1]
res7._set_attr(mirror_stage='True')
bn7 = bn(res7, 'bn7')
relu7 = relu(bn7, 'relu7')
relu7._set_attr(mirror_stage='True')
return relu7
def rna_feat(conv1_layers, level_blocks):
'''RNA features'''
def _widen(filters, max_filters, min_filters=-1):
filters = min(max_filters, int(filters * level_blocks[level].width))
if min_filters > 0:
filters = max(min_filters, filters)
return filters
# TODO: these may only work for b33
def _ds_by_pl(data, name, level, dilate):
assert level > 0
level_block = level_blocks[level]
if not level_block.downsample == 'p':
return data, dilate
pool_stride = stride = 2
inc_dilate = level_block.dilate
pad = -1
if inc_dilate:
pool_stride = 1
pad = dilate
print 'Pooling stride: {}, dilate: {}, pad: {}'.format(pool_stride, dilate, pad)
top = pool(data, name, stride=pool_stride, dilate=dilate, pad=pad, pool_type=pool_type)
if inc_dilate:
dilate *= stride
return top, dilate
def _ds_by_cv(data, name, filters, level, dilate, kernel=3, dropout_rate=0.):
assert level > 1
level_block = level_blocks[level-1]
if not level_block.downsample == 'c':
print 'First block on level {}, dilate: {}'.format(level, dilate)
top = res_block(data, name, filters, kernel=kernel,
dilate=dilate, identity_map=False,
dropout_rate=dropout_rate)
else:
stride = 2
inc_dilate = level_block.dilate
print 'First block on level {}, stride: {}, dilate: {}'.format(level, stride, dilate)
top = res_block(data, name, filters, kernel=kernel, stride=stride,
dilate=dilate, inc_dilate=inc_dilate, identity_map=False,
dropout_rate=dropout_rate)
if inc_dilate:
dilate *= stride
return top, dilate
dilate = 1
pool_type = 'max'
crop_size = 224
data = mx.sym.Variable('data')
# 224^2 3
level = 0; print 'Level {}'.format(level)
conv0 = data
# 224^2 3
print conv0.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 1; print 'Level {}'.format(level)
filters = _widen(64, 1024)
res1 = conv(conv0, 'conv1a',
filters,
kernel=conv1_layers[0].kernel,
stride=1,
no_bias=True)
for i, conv1_layer in enumerate(conv1_layers[1:]):
names = ['conv1b{}_bn'.format(i),
'conv1b{}_relu'.format(i),
'conv1b{}_do'.format(i),
'conv1b{}'.format(i),]
res1 = conv_stage(res1, names,
filters,
kernel=conv1_layer.kernel,
stride=1,
no_bias=True)
for i in xrange(level_blocks[level].num):
res1 = res_block(res1, '1b{}'.format(i+1), filters)
# pass dilate from now on
assert dilate == 1
res1, dilate = _ds_by_pl(res1, 'pool1', level, dilate)
# 112^2 64
print res1.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 2; print 'Level {}'.format(level)
filters = _widen(128, 1024)
res2, dilate = _ds_by_cv(res1, '2a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res2 = res_block(res2, '2b{}'.format(i), filters, dilate=dilate)
res2, dilate = _ds_by_pl(res2, 'pool2', level, dilate)
# 56^2 128
print res2.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 3; print 'Level {}'.format(level)
filters = _widen(256, 1024)
res3, dilate = _ds_by_cv(res2, '3a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res3 = res_block(res3, '3b{}'.format(i), filters, dilate=dilate)
res3, dilate = _ds_by_pl(res3, 'pool3', level, dilate)
# 28^2 256
print res3.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 4; print 'Level {}'.format(level)
filters = _widen(512, 1024)
res4, dilate = _ds_by_cv(res3, '4a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res4 = res_block(res4, '4b{}'.format(i), filters, dilate=dilate)
res4, dilate = _ds_by_pl(res4, 'pool4', level, dilate)
# 14^2 512
print res4.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 5; print 'Level {}'.format(level)
# wd = 0.5 : (512, 512)
# wd = 1.0 : (512, 1024)
# wd = 2.0 : (1024, 1024)
filters_l1 = _widen(512, 1024, filters)
filters = [filters_l1, _widen(1024, 1024, filters_l1)]
res5, dilate = _ds_by_cv(res4, '5a', filters, level, dilate)
for i in xrange(1, level_blocks[level].num):
res5 = res_block(res5, '5b{}'.format(i), filters, dilate=dilate)
res5, dilate = _ds_by_pl(res5, 'pool5', level, dilate)
# 7^2 1024
print res5.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 6; print 'Level {}'.format(level)
res6, dilate = _ds_by_cv(res5, '6a', [512, 1024, 2048],
level, dilate, kernel=[1, 3, 1],
dropout_rate=level_blocks[level].dropout)
# 7^2 2048
print res6.infer_shape(data=(64, 3, crop_size, crop_size))[1]
level = 7; print 'Level {}'.format(level)
res7 = res_block(res6, '7a', [1024, 2048, 4096], kernel=[1, 3, 1],
dilate=dilate, identity_map=False,
dropout_rate=level_blocks[level].dropout)
# 7^2 4096
print res7.infer_shape(data=(64, 3, crop_size, crop_size))[1]
bn7 = bn(res7, 'bn7')
relu7 = relu(bn7, 'relu7')
return relu7
