in c2/lib/utils/model_helper.py [0:0]
def GetFlopsAndParams(model, gpu_id=0):
model_ops = model.net.Proto().op
master_gpu = 'gpu_{}'.format(gpu_id)
param_ops = []
for idx in range(len(model_ops)):
op_type = model.net.Proto().op[idx].type
op_input = model.net.Proto().op[idx].input[0]
if op_type in ['Conv', 'FC'] and op_input.find(master_gpu) >= 0:
param_ops.append(model.net.Proto().op[idx])
num_flops = 0
num_params = 0
num_interactions = 0
for idx in range(len(param_ops)):
op = param_ops[idx]
op_type = op.type
op_inputs = param_ops[idx].input
op_output = param_ops[idx].output[0]
layer_flops = 0
layer_params = 0
if op_type == 'Conv':
for op_input in op_inputs:
if '_w' in op_input and 'bias' not in op_input:
param_blob = op_input
param_shape = np.array(
workspace.FetchBlob(str(param_blob))).shape
layer_params = (
np.prod(param_shape)
)
output_shape = np.array(
workspace.FetchBlob(str(op_output))).shape
layer_flops = layer_params * np.prod(output_shape[2:])
layer_interactions = 0.5 * param_shape[0] * param_shape[1] * (param_shape[1] - 1)
# log.info('{} size {}x{}x{} FLOPs {} params {} inters {}'.format(
# str(param_blob),
# (param_shape[2] if len(param_shape) == 5 else 1),
# (param_shape[3] if len(param_shape) == 5 else param_shape[2]),
# (param_shape[4] if len(param_shape) == 5 else param_shape[3]),
# layer_flops,
# layer_params,
# layer_interactions))
elif op_type == 'FC':
for op_input in op_inputs:
if '_w' in op_input:
param_blob = op_input
param_shape = np.array(
workspace.FetchBlob(str(param_blob))).shape
layer_params = param_shape[0] * param_shape[1]
layer_flops = layer_params
layer_interactions = 0 # not count interactions on FC
layer_params /= 1000000
layer_flops /= 1000000000
layer_interactions /= 1000000000
num_flops += layer_flops
num_params += layer_params
num_interactions += layer_interactions
return num_flops, num_params, num_interactions