def change_lr_mult(model, output):
with open(model) as n:
str1 = n.read()
msg = caffe_pb2.NetParameter()
text_format.Merge(str1, msg)
for i, l in enumerate(msg.layer):
if l.type == "Convolution" or l.type == "DepthwiseConvolution" or l.type == "ConvolutionDepthwise":
if l.param._values == []:
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 0
l.param._values.append(lr_param)
if l.convolution_param.bias_term is True:
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 0
l.param._values.append(lr_param)
# You should ensure param num and blobs num is equal
else:
for num in range(len(l.param)):
l.param[num].lr_mult = 0
print("Saving output model %s..." % output)
with open(output, 'w') as m:
m.write(text_format.MessageToString(msg))
return output
def add_BN_layer(net_msg,name,bottom):
# norm layer
batchnormlayer = net_msg.layer.add()
batchnormlayer.name = name+'_norm'
batchnormlayer.type = 'BatchNorm'
batchnormlayer.bottom._values.append(bottom)
batchnormlayer.top._values.append(batchnormlayer.name)
for i in range(0,3):
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 0
batchnormlayer.param._values.append(lr_param)
# scale layer
scalelayer = net_msg.layer.add()
scalelayer.name = name+'_scale'
scalelayer.type = 'Scale'
scalelayer.bottom._values.append(batchnormlayer.name)
scalelayer.top._values.append(name)
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 1
scalelayer.param._values.append(lr_param)
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 2
lr_param.decay_mult = 0
scalelayer.param._values.append(lr_param)
scalelayer.scale_param.bias_term = True
scalelayer.scale_param.filler.type = 'msra'
def add_conv_layer(net_msg,name,bottom,num_output,pad,kernel_size,stride,bias_term=True,learn_depth=False,input_channel=1,connectivity_mode=0):
conv_layer = net_msg.layer.add()
conv_layer.name = name
conv_layer.type = 'Convolution'
conv_layer.bottom._values.append(bottom)
conv_layer.top._values.append(conv_layer.name)
if 1==connectivity_mode:
conv_layer.connectivity_mode = caffe_pb2.LayerParameter.DISCONNECTED_ELTWISE
elif 2==connectivity_mode:
conv_layer.connectivity_mode = caffe_pb2.LayerParameter.DISCONNECTED_GRPWISE
# param info for weight and bias
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 1
if learn_depth:
blk_param = caffe_pb2.BlockGroupLassoSpec()
blk_param.xdimen = kernel_size*kernel_size*input_channel
blk_param.ydimen = num_output
lr_param.block_group_lasso._values.append(blk_param)
conv_layer.param._values.append(lr_param)
if bias_term:
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 2
conv_layer.param._values.append(lr_param)
# conv parameters
conv_layer.convolution_param.num_output = num_output
conv_layer.convolution_param.pad._values.append(pad)
conv_layer.convolution_param.kernel_size._values.append(kernel_size)
conv_layer.convolution_param.stride._values.append(stride)
conv_layer.convolution_param.weight_filler.type = 'msra'
conv_layer.convolution_param.bias_term = bias_term
if bias_term:
conv_layer.convolution_param.bias_filler.type = 'constant'
def add_BN_layer(net_msg, bottom):
# norm layer
batchnormlayer = net_msg.layer.add()
batchnormlayer.name = bottom + '_bn'
batchnormlayer.type = 'BatchNorm'
batchnormlayer.bottom._values.append(bottom)
batchnormlayer.top._values.append(bottom)
# compatible with old caffe version
for i in range(0, 3):
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 0
lr_param.decay_mult = 0
batchnormlayer.param._values.append(lr_param)
# scale layer
scalelayer = net_msg.layer.add()
scalelayer.name = bottom + '_scale'
scalelayer.type = 'Scale'
scalelayer.bottom._values.append(bottom)
scalelayer.top._values.append(bottom)
# optional: modify lr mult
# lr_param = caffe_pb2.ParamSpec()
# lr_param.lr_mult = 1
# scalelayer.param._values.append(lr_param)
# lr_param = caffe_pb2.ParamSpec()
# lr_param.lr_mult = 2
# lr_param.decay_mult = 0
# scalelayer.param._values.append(lr_param)
scalelayer.scale_param.bias_term = True
scalelayer.scale_param.filler.value = 1
scalelayer.scale_param.bias_filler.value = 0
return bottom
def rename_Upsample(self, source_node):
attr = source_node.attrs
layer = pb2.LayerParameter()
layer.type = "Deconvolution"
assert attr['height_scale'] == attr['width_scale']
factor = int(attr['height_scale'])
c = int(attr['channel'])
k = 2 * factor - factor % 2
layer.convolution_param.num_output = c
layer.convolution_param.kernel_size.extend([k])
layer.convolution_param.stride.extend([factor])
layer.convolution_param.pad.extend([int(math.ceil((factor - 1) / 2.))])
layer.convolution_param.group = c
layer.convolution_param.weight_filler.type = 'bilinear'
layer.convolution_param.bias_term = False
learning_param = pb2.ParamSpec()
learning_param.lr_mult = 0
learning_param.decay_mult = 0
layer.param.extend([learning_param])
""" Init weight blob of filter kernel """
blobs_weight = FillBilinear(c, k)
layer.blobs.extend([as_blob(blobs_weight)])
for b in source_node.in_edges:
layer.bottom.append(b)
layer.top.append(source_node.name)
layer.name = source_node.real_name
return layer
def UpsampleBilinear(pytorch_layer):
layer = pb2.LayerParameter()
layer.type = "Deconvolution"
assert pytorch_layer.scale_factor[0] == pytorch_layer.scale_factor[1]
factor = int(pytorch_layer.scale_factor[0])
c = int(pytorch_layer.input_size[1])
k = 2 * factor - factor % 2
layer.convolution_param.num_output = c
layer.convolution_param.kernel_size.extend([k])
layer.convolution_param.stride.extend([factor])
layer.convolution_param.pad.extend([int(math.ceil((factor - 1) / 2.))])
layer.convolution_param.group = c
layer.convolution_param.weight_filler.type = 'bilinear'
layer.convolution_param.bias_term = False
learning_param = pb2.ParamSpec()
learning_param.lr_mult = 0
learning_param.decay_mult = 0
layer.param.extend([learning_param])
""" Init weight blob of filter kernel """
blobs_weight = FillBilinear(c, k)
layer.blobs.extend([as_blob(blobs_weight)])
return layer
def bn(torch_layer):
layer = pb2.LayerParameter()
layer.type = "BN"
param_scale = pb2.ParamSpec()
param_scale.lr_mult = 1
param_scale.decay_mult = 0
param_bias = pb2.ParamSpec()
param_bias.lr_mult = 1
param_bias.decay_mult = 0
layer.param.extend([param_scale, param_bias])
layer.bn_param.slope_filler.value = 1
layer.bn_param.bias_filler.value = 0
layer.blobs.extend([
as_blob(torch_layer[name][None, :])
for name in ['weight', 'bias', 'running_mean', 'running_var']
])
return layer
def _get_param(num_param, lr_mult=1):
if num_param == 1:
# only weight
param = caffe_pb2.ParamSpec()
param.lr_mult = 1 * lr_mult
param.decay_mult = 1 * lr_mult
return [param]
elif num_param == 2:
# weight and bias
param_w = caffe_pb2.ParamSpec()
param_w.lr_mult = 1 * lr_mult
param_w.decay_mult = 1 * lr_mult
param_b = caffe_pb2.ParamSpec()
param_b.lr_mult = 2 * lr_mult
param_b.decay_mult = 0
return [param_w, param_b]
else:
raise ValueError("Unknown num_param {}".format(num_param))
def __init__(self, name, num_filters,
kernel_size, stride=1, pad=0,
weight_filler=GaussianFiller(std=0.01),
lr_mult=1):
super(ConvolutionLayer, self).__init__(name, 'Convolution', 1, 1)
self._inplace = False
self._params.convolution_param.num_output = num_filters
self._params.convolution_param.kernel_size.extend([kernel_size])
self._params.convolution_param.pad.extend([pad])
self._params.convolution_param.stride.extend([stride])
self._params.convolution_param.weight_filler.MergeFrom(weight_filler.to_proto())
self._params.convolution_param.bias_filler.MergeFrom(ConstantFiller().to_proto())
weight_blob_param = caffe_pb2.ParamSpec(lr_mult=1 * lr_mult)
bias_blob_param = caffe_pb2.ParamSpec(lr_mult=2 * lr_mult)
self._params.param.extend([weight_blob_param, bias_blob_param])
def add_ip_layer(net_msg,name,bottom,num):
ip_layer = net_msg.layer.add()
ip_layer.name = name
ip_layer.type = 'InnerProduct'
ip_layer.bottom._values.append(bottom)
ip_layer.top._values.append(name)
# param info for weight and bias
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 1
lr_param.decay_mult = 1
ip_layer.param._values.append(lr_param)
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 2
lr_param.decay_mult = 0
ip_layer.param._values.append(lr_param)
# inner product parameters
ip_layer.inner_product_param.num_output = num
ip_layer.inner_product_param.weight_filler.type = 'msra'
ip_layer.inner_product_param.bias_filler.type = 'constant'
ip_layer.inner_product_param.bias_filler.value = 0.0
def _simple_conv_layer(name,
bottom,
top,
num_output,
kernel_size,
pad,
dilation=1,
std=0.01,
bias=0.0,
param_type=0):
"""
param_type:
0 -> do nothing
1 -> (lr 1, decay 0; lr 2, decay 0)
2 -> (lr 1, decay 1; lr 2, decay 0)
3 -> (lr 10, decay 1; lr 20, decay 0)
4 -> (lr 1, decay 1; lr 1, decay 1)
"""
conv_layer = caffe_pb2.LayerParameter()
conv_layer.name = name
conv_layer.type = 'Convolution'
conv_layer.bottom.append(bottom)
conv_layer.top.append(top)
conv_layer.convolution_param.num_output = num_output
conv_layer.convolution_param.pad.append(pad)
conv_layer.convolution_param.kernel_size.append(kernel_size)
conv_layer.convolution_param.weight_filler.type = "gaussian"
conv_layer.convolution_param.weight_filler.std = std
conv_layer.convolution_param.bias_filler.type = "constant"
conv_layer.convolution_param.bias_filler.value = bias
conv_layer.convolution_param.dilation.append(dilation)
conv_layer.ClearField('param')
conv_layer.param.extend([caffe_pb2.ParamSpec()] * 2)
if param_type == 1:
conv_layer.param[0].lr_mult = 1.0
conv_layer.param[0].decay_mult = 0.0
conv_layer.param[1].lr_mult = 2.0
conv_layer.param[1].decay_mult = 0.0
elif param_type == 2:
conv_layer.param[0].lr_mult = 1.0
conv_layer.param[0].decay_mult = 1.0
conv_layer.param[1].lr_mult = 2.0
conv_layer.param[1].decay_mult = 0.0
elif param_type == 3:
conv_layer.param[0].lr_mult = 10.0
conv_layer.param[0].decay_mult = 1.0
conv_layer.param[1].lr_mult = 20.0
conv_layer.param[1].decay_mult = 0.0
elif param_type == 4:
conv_layer.param[0].lr_mult = 1.0
conv_layer.param[0].decay_mult = 1.0
conv_layer.param[1].lr_mult = 2.0
conv_layer.param[1].decay_mult = 1.0
return conv_layer
def add_conv_layer(net_msg,name,bottom,num_output,pad,kernel_size,stride,bias_term=True):
conv_layer = net_msg.layer.add()
conv_layer.name = name
conv_layer.type = 'Convolution'
conv_layer.bottom._values.append(bottom)
conv_layer.top._values.append(conv_layer.name)
# param info for weight and bias
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 1
conv_layer.param._values.append(lr_param)
if bias_term:
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 2
conv_layer.param._values.append(lr_param)
# conv parameters
conv_layer.convolution_param.num_output = num_output
conv_layer.convolution_param.pad._values.append(pad)
conv_layer.convolution_param.kernel_size._values.append(kernel_size)
conv_layer.convolution_param.stride._values.append(stride)
conv_layer.convolution_param.weight_filler.type = 'msra'
conv_layer.convolution_param.bias_term = bias_term
if bias_term:
conv_layer.convolution_param.bias_filler.type = 'constant'
def create_inner_product_layer(input, out_name, output_number):
inner_product_layer = caffe_pb2.LayerParameter(
name=out_name,
type="InnerProduct",
bottom=[input],
top=[out_name],
param=[
caffe_pb2.ParamSpec(
lr_mult=1.0,
decay_mult=1.0,
),
caffe_pb2.ParamSpec(
lr_mult=2.0,
decay_mult=0.0,
)
],
inner_product_param=caffe_pb2.InnerProductParameter(
num_output=output_number,
bias_term=False,
weight_filler=caffe_pb2.FillerParameter(type="xavier"),
bias_filler=caffe_pb2.FillerParameter(type="constant", value=0.0)))
net = caffe_pb2.NetParameter()
net.layer.extend([inner_product_layer])
return net
def deconvolution(torch_layer):
log.info("do deconvolution")
# log.info(torch_layer)
# print_all(torch_layer["output"])
# print_all(torch_layer["gradInput"])
# get output and cal num_output, need forward torch
output = torch_layer["output"]
size = len(output)
assert size > 0
# print(len(output[0]))
# gradinput = torch_layer["gradInput"]
# print(len(output))
# print(len(gradinput))
# print(output[0].size[1])
# print(gradinput[0].size[1])
# assert torch_layer is None
layer = pb2.LayerParameter()
layer.type = "Deconvolution"
factor = int(torch_layer["scale_factor"])
layer.convolution_param.num_output = len(output[0])
layer.convolution_param.group = len(output[0])
# layer.convolution_param.stride = factor
# layer.convolution_param.kernel_size = (2 * factor - factor % 2)
layer.convolution_param.stride_w = factor
layer.convolution_param.stride_h = factor
layer.convolution_param.kernel_w = (2 * factor - factor % 2)
layer.convolution_param.kernel_h = (2 * factor - factor % 2)
layer.convolution_param.pad.append(int(np.ceil((factor - 1) / 2.)))
layer.convolution_param.bias_term = False
layer.convolution_param.weight_filler.type = 'bilinear'
param_spec = pb2.ParamSpec()
param_spec.lr_mult = 0
param_spec.decay_mult = 0
layer.param.extend([param_spec])
# weight = torch_layer["weight"]
# bias = torch_layer["bias"]
# assert len(weight.shape) == 4, weight.shape
# (nOutputPlane, nInputPlane, kH_, kW_) = weight.shape
#
# (kW, kH, dW, dH, padW, padH) = [
# int(torch_layer.get(f, 0))
# for f in ["kW", "kH", "dW", "dH", "padW", "padH"]]
# assert kH_ == kH
# assert kW_ == kW
# layer.convolution_param.num_output = nOutputPlane
# layer.convolution_param.kernel_w = kW
# layer.convolution_param.stride_w = dW
# layer.convolution_param.pad_w = padW
# layer.convolution_param.kernel_h = kH
# layer.convolution_param.stride_h = dH
# layer.convolution_param.pad_h = padH
#
# if "bias" in torch_layer:
# bias = torch_layer["bias"]
# layer.blobs.extend([as_blob(weight), as_blob(bias)])
# else:
# layer.convolution_param.bias_term = False
# layer.blobs.extend([as_blob(weight), as_blob(bias)])
return layer
def _get_primal_dual_param(param_name, lr_mult):
param = caffe_pb2.ParamSpec()
param.lr_mult = lr_mult
param.decay_mult = 0
param.name = param_name
return [param]
def main(args):
# Set default output file names
if args.output_model is None:
file_name = osp.splitext(args.model)[0]
args.output_model = file_name + '_inference.prototxt'
if args.output_weights is None:
file_name = osp.splitext(args.weights)[0]
args.output_weights = file_name + '_inference.caffemodel'
with open(args.model) as f:
model = caffe_pb2.NetParameter()
pb.text_format.Parse(f.read(), model)
# Determine the BN layers to be absorbed or replaced
# Create the new layers
new_layers = []
absorbed = {}
replaced = {}
for i, layer in enumerate(model.layer):
if layer.type != 'BN':
new_layers.append(layer)
continue
assert len(layer.bottom) == 1
assert len(layer.top) == 1
bottom_blob = layer.bottom[0]
top_blob = layer.top[0]
# Check if can be absorbed. As there could be some inplace layers,
# for example, conv -> relu -> bn. In such case, the BN cannot be
# absorbed.
can_be_absorbed = False
for j in xrange(i - 1, -1, -1):
if bottom_blob in model.layer[j].top:
if model.layer[j].type not in ['Convolution', 'InnerProduct']:
can_be_absorbed = False
break
else:
can_be_absorbed = True
bottom_layer = model.layer[j]
if can_be_absorbed:
# Rename the blob in the top layers
for j in xrange(i + 1, len(model.layer)):
update_blob_name(model.layer[j].bottom, top_blob, bottom_blob)
update_blob_name(model.layer[j].top, top_blob, bottom_blob)
if bottom_layer.type == 'Convolution':
bottom_layer.convolution_param.bias_term = True
elif bottom_layer.type == 'InnerProduct':
bottom_layer.inner_product_param.bias_term = True
absorbed[layer.name] = bottom_layer.name
elif args.replace_by == 'affine':
# Replace by an scale bias layer
new_layer = caffe_pb2.LayerParameter()
new_layer.name = layer.name + '_affine'
new_layer.type = 'Scale'
new_layer.bottom.extend([bottom_blob])
new_layer.top.extend([top_blob])
new_layer.scale_param.bias_term = True
replaced[layer.name] = new_layer.name
new_layers.append(new_layer)
elif args.replace_by == 'frozen':
# Freeze the BN layer
layer.bn_param.frozen = True
del layer.param[:]
param = caffe_pb2.ParamSpec()
param.lr_mult = 0
param.decay_mult = 0
layer.param.extend([param] * 2)
new_layers.append(layer)
# Save the prototxt
output_model = caffe_pb2.NetParameter()
output_model.CopyFrom(model)
del output_model.layer[:]
output_model.layer.extend(new_layers)
with open(args.output_model, 'w') as f:
f.write(pb.text_format.MessageToString(output_model))
# Copy the parameters
weights = caffe.Net(args.model, args.weights, caffe.TEST)
output_weights = caffe.Net(args.output_model, caffe.TEST)
for name in np.intersect1d(weights.params.keys(),
output_weights.params.keys()):
# Some original conv / inner product layers do not have bias_term
for i in xrange(
min(len(weights.params[name]),
len(output_weights.params[name]))):
output_weights.params[name][i].data[...] = \
weights.params[name][i].data.copy()
# Absorb the BN parameters
for old, new in absorbed.iteritems():
scale, bias, mean, tmp = [p.data.ravel() for p in weights.params[old]]
invstd = tmp if args.bn_style == 'invstd' else \
np.power(tmp + args.epsilon, -0.5)
W, b = output_weights.params[new]
assert W.data.ndim == 4 or W.data.ndim == 2
assert b.data.ndim == 1
if W.data.ndim == 4:
W.data[...] = (W.data * scale[:, None, None, None] *
invstd[:, None, None, None])
elif W.data.ndim == 2:
W.data[...] = W.data * scale[:, None] * invstd[:, None]
b.data[...] = (b.data[...] - mean) * scale * invstd + bias
# Fill up the affine layers
for old, new in replaced.iteritems():
scale, bias, mean, tmp = [p.data.ravel() for p in weights.params[old]]
invstd = tmp if args.bn_style == 'invstd' else \
np.power(tmp + args.epsilon, -0.5)
W, b = output_weights.params[new]
assert W.data.ndim == 1
assert b.data.ndim == 1
W.data[...] = scale * invstd
b.data[...] = bias - scale * mean * invstd
# Check if the conversion is correct
check(weights, output_weights)
# Save the caffemodel
output_weights.save(args.output_weights)
conv_positions = []
position_idx = 0
for cur_layer in loop_layers:
if 'Convolution' == cur_layer.type:
print 'generating:', cur_layer.name, cur_layer.type
conv_positions.append(position_idx)
#p conv layer
conv_p = net_msg.layer.add()
conv_p.CopyFrom(
cur_layer) #get a copy to make sure other parameters are consist
conv_p.name = conv_p.name + 'p'
conv_p.bottom._values[0] = cur_layer.bottom._values[0]
conv_p.top._values[0] = conv_p.name
if len(conv_p.param) == 0:
lr_param = caffe_pb2.ParamSpec()
lr_param.lr_mult = 0
lr_param.decay_mult = 0
conv_p.param._values.append(lr_param)
else:
conv_p.param._values[0].lr_mult = 0
conv_p.param._values[0].decay_mult = 0
try:
#conv_p.param._values[1].lr_mult = 0
#conv_p.param._values[1].decay_mult = 0
del conv_p.param._values[1]
except:
print "Failed to operate param field: {}".format(conv_p.name)
conv_p.convolution_param.num_output = m_all[layer_idx]
