def parse_caffemodel(model_path):
caffemodel = caffe_pb2.NetParameter()
with open(model_path, 'rb') as f:
caffemodel.MergeFromString(f.read())
# Check trainable layers
print(*set([(layer.type, len(layer.blobs)) for layer in caffemodel.layer]), sep='\n')
params = OrderedDict()
previous_layer_type = None
for layer in caffemodel.layer:
print('{} ({}): {}'.format(layer.name, layer.type, len(layer.blobs)))
# Skip the shared branch
if 'res075' in layer.name or 'res05' in layer.name:
continue
# Convolution or Dilated Convolution
if 'Convolution' in layer.type:
params[layer.name] = {}
params[layer.name]['kernel_size'] = layer.convolution_param.kernel_size[0]
params[layer.name]['weight'] = list(layer.blobs[0].data)
if len(layer.blobs) == 2:
params[layer.name]['bias'] = list(layer.blobs[1].data)
if len(layer.convolution_param.stride) == 1: # or []
params[layer.name]['stride'] = layer.convolution_param.stride[0]
else:
params[layer.name]['stride'] = 1
if len(layer.convolution_param.pad) == 1: # or []
params[layer.name]['padding'] = layer.convolution_param.pad[0]
else:
params[layer.name]['padding'] = 0
if isinstance(layer.convolution_param.dilation, int):
params[layer.name]['dilation'] = layer.convolution_param.dilation
elif len(layer.convolution_param.dilation) == 1:
params[layer.name]['dilation'] = layer.convolution_param.dilation[0]
else:
params[layer.name]['dilation'] = 1
# Batch Normalization
elif 'BatchNorm' in layer.type:
params[layer.name] = {}
params[layer.name]['running_mean'] = np.array(layer.blobs[0].data) / layer.blobs[2].data[0] # NOQA
params[layer.name]['running_var'] = np.array(layer.blobs[1].data) / layer.blobs[2].data[0] # NOQA
params[layer.name]['eps'] = layer.batch_norm_param.eps
params[layer.name]['momentum'] = layer.batch_norm_param.moving_average_fraction
batch_norm_layer = layer.name
# Scale
elif 'Scale' in layer.type:
assert (previous_layer_type == 'BatchNorm')
params[batch_norm_layer]['weight'] = list(layer.blobs[0].data)
params[batch_norm_layer]['bias'] = list(layer.blobs[1].data)
previous_layer_type = layer.type
return params
def parse_caffemodel(model_path):
caffemodel = caffe_pb2.NetParameter()
with open(model_path, 'rb') as f:
caffemodel.MergeFromString(f.read())
# Check trainable layers
print(set([(layer.type, len(layer.blobs)) for layer in caffemodel.layer]))
params = OrderedDict()
for layer in caffemodel.layer:
print('{} ({}): {}'.format(layer.name, layer.type, len(layer.blobs)))
# Convolution or Dilated Convolution
if 'Convolution' in layer.type:
params[layer.name] = {}
params[layer.name][
'kernel_size'] = layer.convolution_param.kernel_size[0]
params[layer.name]['stride'] = layer.convolution_param.stride[0]
params[layer.name]['weight'] = list(layer.blobs[0].data)
if len(layer.blobs) == 2:
params[layer.name]['bias'] = list(layer.blobs[1].data)
if len(layer.convolution_param.pad) == 1: # or []
params[layer.name]['padding'] = layer.convolution_param.pad[0]
else:
params[layer.name]['padding'] = 0
if isinstance(layer.convolution_param.dilation, int): # or []
params[
layer.name]['dilation'] = layer.convolution_param.dilation
else:
params[layer.name]['dilation'] = 1
# Batch Normalization
elif 'BN' in layer.type:
params[layer.name] = {}
params[layer.name]['weight'] = list(layer.blobs[0].data)
params[layer.name]['bias'] = list(layer.blobs[1].data)
params[layer.name]['running_mean'] = list(layer.blobs[2].data)
params[layer.name]['running_var'] = list(layer.blobs[3].data)
params[layer.name]['eps'] = layer.bn_param.eps
params[layer.name]['momentum'] = layer.bn_param.momentum
return params
def parse_caffemodel(model_path):
caffemodel = caffe_pb2.NetParameter()
with open(model_path, "rb") as f:
caffemodel.MergeFromString(f.read())
# Check trainable layers
print(set([(layer.type, len(layer.blobs)) for layer in caffemodel.layer]))
params = OrderedDict()
for layer in caffemodel.layer:
print("{} ({}): {}".format(layer.name, layer.type, len(layer.blobs)))
# Convolution or Dilated Convolution
if "Convolution" in layer.type:
params[layer.name] = {}
params[layer.name][
"kernel_size"] = layer.convolution_param.kernel_size[0]
params[layer.name]["stride"] = layer.convolution_param.stride[0]
params[layer.name]["weight"] = list(layer.blobs[0].data)
if len(layer.blobs) == 2:
params[layer.name]["bias"] = list(layer.blobs[1].data)
if len(layer.convolution_param.pad) == 1: # or []
params[layer.name]["padding"] = layer.convolution_param.pad[0]
else:
params[layer.name]["padding"] = 0
if isinstance(layer.convolution_param.dilation, int): # or []
params[
layer.name]["dilation"] = layer.convolution_param.dilation
else:
params[layer.name]["dilation"] = 1
# Batch Normalization
elif "BN" in layer.type:
params[layer.name] = {}
params[layer.name]["weight"] = list(layer.blobs[0].data)
params[layer.name]["bias"] = list(layer.blobs[1].data)
params[layer.name]["running_mean"] = list(layer.blobs[2].data)
params[layer.name]["running_var"] = list(layer.blobs[3].data)
params[layer.name]["eps"] = layer.bn_param.eps
params[layer.name]["momentum"] = layer.bn_param.momentum
return params
def parse_caffemodel(model_path):
caffemodel = caffe_pb2.NetParameter()
with open(model_path, "rb") as f:
caffemodel.MergeFromString(f.read())
# Check trainable layers
print(
*Counter([(layer.type, len(layer.blobs))
for layer in caffemodel.layer]).most_common(),
sep="\n",
)
params = OrderedDict()
previous_layer_type = None
for layer in caffemodel.layer:
# Skip the shared branch
if "res075" in layer.name or "res05" in layer.name:
continue
print(
"\033[34m[Caffe]\033[00m",
"{} ({}): {}".format(layer.name, layer.type, len(layer.blobs)),
)
# Convolution or Dilated Convolution
if "Convolution" in layer.type:
params[layer.name] = {}
params[layer.name][
"kernel_size"] = layer.convolution_param.kernel_size[0]
params[layer.name]["weight"] = list(layer.blobs[0].data)
if len(layer.blobs) == 2:
params[layer.name]["bias"] = list(layer.blobs[1].data)
if len(layer.convolution_param.stride) == 1: # or []
params[
layer.name]["stride"] = layer.convolution_param.stride[0]
else:
params[layer.name]["stride"] = 1
if len(layer.convolution_param.pad) == 1: # or []
params[layer.name]["padding"] = layer.convolution_param.pad[0]
else:
params[layer.name]["padding"] = 0
if isinstance(layer.convolution_param.dilation, int):
params[
layer.name]["dilation"] = layer.convolution_param.dilation
elif len(layer.convolution_param.dilation) == 1:
params[layer.
name]["dilation"] = layer.convolution_param.dilation[0]
else:
params[layer.name]["dilation"] = 1
# Fully-connected
elif "InnerProduct" in layer.type:
params[layer.name] = {}
params[layer.name]["weight"] = list(layer.blobs[0].data)
if len(layer.blobs) == 2:
params[layer.name]["bias"] = list(layer.blobs[1].data)
# Batch Normalization
elif "BatchNorm" in layer.type:
params[layer.name] = {}
params[layer.name]["running_mean"] = (
np.array(layer.blobs[0].data) / layer.blobs[2].data[0])
params[layer.name]["running_var"] = (
np.array(layer.blobs[1].data) / layer.blobs[2].data[0])
params[layer.name]["eps"] = layer.batch_norm_param.eps
params[layer.name][
"momentum"] = layer.batch_norm_param.moving_average_fraction
params[layer.name]["num_batches_tracked"] = np.array(0)
batch_norm_layer = layer.name
# Scale
elif "Scale" in layer.type:
assert previous_layer_type == "BatchNorm"
params[batch_norm_layer]["weight"] = list(layer.blobs[0].data)
params[batch_norm_layer]["bias"] = list(layer.blobs[1].data)
elif "Pooling" in layer.type:
params[layer.name] = {}
params[layer.name]["kernel_size"] = layer.pooling_param.kernel_size
params[layer.name]["stride"] = layer.pooling_param.stride
params[layer.name]["padding"] = layer.pooling_param.pad
previous_layer_type = layer.type
return params
