def __init__(self):
self.layers = {} # name of the layers
self.detail_layers = {} # detail of the net, #of each kind of layer
self.detail_blobs = {}
self._blobs = Blob_LOG() # Record the blobs
self._blob_data = []
self.cnet = caffe_net.Caffemodel('') # the caffenet
self.debug = False
self.pytorch_layer_name = None
def __init__(self):
"""
doing init() with inputs Variable before using it
"""
self.layers = {}
self._blobs = {}
self._blobs_data = []
self.cnet = caffe_net.Caffemodel('')
self.debug = False
def __init__(self):
"""
doing init() with inputs Variable before using it
"""
self.layers = {}
self.detail_layers = {}
self.detail_blobs = {}
self._blobs = Blob_LOG()
self._blobs_data = []
self.cnet = caffe_net.Caffemodel()
self.debug = True
def remove_layer_by_name(json_dict):
prevOPName = json_dict["prevOPName"]
nextOPName = json_dict["nextOPName"]
net = caffe_net.Caffemodel(caffemodel_file)
prevlayer = net.get_layer_by_name(prevOPName)
prevOPTop = prevlayer.top[0]
net.remove_layer_by_name(currOPName)
if nextOPName != "":
layer = net.get_layer_by_name(nextOPName)
layer.bottom[0] = prevOPTop
return net
def main():
net = caffe_net.Caffemodel(caffemodel)
net.save_prototxt(prototxt)
def set_layer_data(json_dict):
weight = np.array(json_dict["weight"]).astype(np.float32)
net = caffe_net.Caffemodel(caffemodel_file)
net.set_layer_data(name, weight)
return net
def add_new_layer(json_dict):
currOPTop = json_dict["currOPTop"]
layerType = json_dict["layerType"]
prevOPName = json_dict["prevOPName"]
nextOPName = json_dict["nextOPName"]
net = caffe_net.Caffemodel(caffemodel_file)
layer = net.get_layer_by_name(prevOPName)
prevOPBottom = layer.top[0]
layer_params = layer_param.Layer_param(currOPName,
layerType,
currOPTop,
prevOPBottom,
phase=1)
if layerType.endswith("Convolution"):
ke_h = json_dict["kernel_h"]
ke_w = json_dict["kernel_w"]
num_out = json_dict["num_output"]
str_h = json_dict["stride_h"]
str_w = json_dict["stride_w"]
group = json_dict["group"]
bias = json_dict["bias_term"]
dila = json_dict["dilation"]
p_h = json_dict["pad_h"]
pad_w = json_dict["pad_w"]
weight = np.array(json_dict["weight"]).astype(np.float32)
# 卷积层后面必须是BN层
layerOP = net.get_layer_by_name(nextOPName)
assert layerOP.type.endswith("BatchNorm")
for blob in layerOP.blobs:
shape = blob.shape.dim
assert num_out == shape[0]
break
layer_params.conv_param(kernel_h=ke_h,
kernel_w=ke_w,
num_output=num_out,
stride_h=str_h,
stride_w=str_w)
net.add_layer_with_data(layer_params,
datas=weight,
bottom=prevOPBottom)
elif layerType.endswith("Softmax"):
axis = json_dict["num_axis"]
weight = json_dict["weight"]
layer_params.softmax_param(num_axis=axis)
net.add_layer_with_data(layer_params,
datas=weight,
bottom=prevOPBottom)
elif layerType.endswith("Permute"):
weight = json_dict["weight"]
layer_params.permute_param()
net.add_layer_with_data(layer_params,
datas=weight,
bottom=prevOPBottom)
elif layerType.endswith("Reshape"):
weight = json_dict["weight"]
shape = json_dict["shape"]
layer_params.reshape_param(shape=shape)
net.add_layer_with_data(layer_params,
datas=weight,
bottom=prevOPBottom)
if nextOPName != "":
layer = net.get_layer_by_name(nextOPName)
layer.bottom[0] = currOPTop
return net
