def __init__(self, model_path):
if not available:
msg = 'CaffeFunction is only supported on protobuf>=3 in Python3'
raise RuntimeError(msg)
super(CaffeFunction, self).__init__()
net = caffe_pb.NetParameter()
with open(model_path, 'rb') as model_file:
net.MergeFromString(model_file.read())
self.forwards = {}
self.split_map = {}
self.layers = []
if net.layer:
for layer in net.layer:
meth = _type_to_method.get(layer.type)
if meth:
meth(self, layer)
else:
warnings.warn(
'Skip the layer "%s", since CaffeFunction does not'
'support %s layer' % (layer.name, layer.type))
else: # v1 format
for layer in net.layers:
meth = _oldname_to_method.get(layer.type)
if meth:
meth(self, layer)
else:
warnings.warn(
'Skip the layer "%s", since CaffeFunction does not'
'support it' % layer.name)
def __call__(self, name, inputs, outputs):
dumped_list = _dump_graph(outputs)
f = None
net = None
if self.caffemodel is not None:
net = caffe_pb.NetParameter()
try:
if self.prototxt is not None:
f = open(self.prototxt, 'wt')
f.write('name: "{}"\n'.format(name))
assert len(inputs) == 1
f.write('layer {\n'
' name: "data"\n'
' type: "Input"\n'
' top: "data"\n'
' input_param { shape: {')
for i in inputs[0].shape:
f.write(' dim: ' + str(i))
f.write(' } }\n' '} \n')
for i in dumped_list:
self.dump_function_object(i, f, net)
finally:
if f is not None:
f.close()
if net is not None:
with open(self.caffemodel, 'wb') as f:
f.write(net.SerializeToString())
if self.debug:
import google.protobuf.text_format
with open(self.caffemodel + ".txt", 'w') as f:
f.write(google.protobuf.text_format.MessageToString(net))
def copy_ssd_norm_layer(caffe_ssd_path, pose_ssd_model):
net = caffe_pb.NetParameter()
with open(caffe_ssd_path, 'rb') as model_file:
net.MergeFromString(model_file.read())
def get_normalize_layer(net):
for layer in net.layer:
if layer.type == 'Normalize':
return layer
norm_layer = get_normalize_layer(net)
pose_ssd_model.ssd_net.norm4.scale.data = numpy.array(
norm_layer.blobs[0].data)
print("Copy layer: dst({}) / src({})".format("ssd_net.norm4",
"conv4_3_norm"))
def __init__(self, caffemodel): self.timer_hooks = [] net = caffe_pb.NetParameter() with open(caffemodel, 'rb') as model_file: net.MergeFromString(model_file.read()) if net.layer: for layer in net.layer: layer_impl_path = _layer_impl_path_dict.get(layer.type) if layer_impl_path: _setup_timer_by_layer(layer.type, layer_impl_path) else: print 'unimplemented layer timer: %s' % layer.type else: #v1 format for layer in net.layers: layer_type = _v1_to_new_name_dict[layer.type] layer_impl_path = _layer_impl_path_dict.get(layer_type) if layer_impl_path: self._setup_timer_by_layer(layer.name, layer_impl_path) else: print 'unimplemented layer timer: %s' % layer.type
def load_caffe_model(self, model_path):
net = caffe_pb.NetParameter()
with open(model_path, 'rb') as model_file:
net.MergeFromString(model_file.read())
print("Loaded")
if net.layer:
for layer in net.layer:
if layer.type == 'Convolution':
#continue
print(layer.name, layer.type)
obj = self.get_layer(layer.name)
blobs = layer.blobs
param = layer.convolution_param
num = _get_num(blobs[0])
channels = _get_channels(blobs[0])
#print(obj.out_channels)
#print(_get_num(blobs[0]))
assert obj.ksize == _get_ksize(param)
assert obj.stride[0] == _get_stride(param)
assert obj.pad[0] == _get_pad(param)
assert obj.W.data.shape[1] == _get_channels(blobs[0])
assert obj.out_channels == _get_num(blobs[0])
n_in = obj.W.data.shape[1]
n_out = obj.out_channels
part_size = len(blobs[0].data) // param.group
obj.W.data[...] = 0
for i in six.moves.range(param.group):
in_slice = slice(i * n_in // param.group,
(i + 1) * n_in // param.group)
out_slice = slice(i * n_out // param.group,
(i + 1) * n_out // param.group)
w = obj.W.data[out_slice, in_slice]
data = np.array(blobs[0].data[i * part_size:(i + 1) *
part_size])
w[:] = data.reshape(w.shape)
if param.bias_term:
obj.b.data[:] = blobs[1].data
elif layer.type == 'Deconvolution':
print(layer.name, layer.type)
obj = self.get_layer(layer.name)
blobs = layer.blobs
param = layer.convolution_param
num = _get_num(blobs[0])
channels = _get_channels(blobs[0])
assert obj.ksize == _get_ksize(param)
assert obj.stride[0] == _get_stride(param)
assert obj.pad[0] == _get_pad(param)
assert obj.W.data.shape[0] == num
assert obj.out_channels == channels # _get_channels(blobs[0])
part_size = len(blobs[0].data) #// param.group
obj.W.data[...] = 0
obj.W.data = np.array(blobs[0].data[0:part_size]).reshape(
obj.W.data.shape)
if param.bias_term:
obj.b.data[:] = blobs[1].data
