# 结果余弦相似度对比
import numpy as np
import pdb
def count_difference(groundtruth, inputs):
statistical_method = {
'cosine_similarity':
lambda X1, X2: np.sum(X1 * X2) /
(np.sqrt(np.sum(X1**2)) * np.sqrt(np.sum(X2**2))),
'maximum_absolute_error':
lambda X1, X2: np.max(np.abs(X1 - X2)).tolist(),
'accumulated_relative_error':
lambda X1, X2: np.sum(np.abs(np.nan_to_num((X1 - X2) / X1))),
'relative_euclidean_distance':
lambda X1, X2: np.sqrt(
np.sum((X1 / np.sqrt(np.sum(X1**2)) - X2 / np.sqrt(np.sum(X2**2)))
**2)).tolist(),
'kullback_leibler_divergence':
lambda X1, X2: np.sum(X1 * np.nan_to_num(np.log(X1 / X2))),
'standard_deviation':
lambda *X: [(np.mean(x).tolist(), np.std(x).tolist()) for x in X]
}
reports = {}
for input_key in groundtruth.keys():
if input_key in inputs.keys():
reports[input_key] = {}
gt_input = groundtruth[input_key].reshape(-1, 1)
compare_input = inputs[input_key].reshape(-1, 1)
assert gt_input.size == compare_input.size# 要求对比的两张图尺寸一致
for key, value in statistical_method.items():
reports[input_key][key] = value(gt_input, compare_input)
return reports
def main():
a = np.fromfile("/home/wangmaolin/for_test/tofile/conv_82_memory", dtype=np.float32)
print(a.shape)
print(a.dtype)
inputs = {"data": a}
c = np.fromfile("/home/wangmaolin/for_test/onnx_output/onnx_output_conv_82", dtype=np.float32)
print(c.shape)
print(c.dtype)
gt_inputs = {"data": c}
report = count_difference(gt_inputs, inputs)
print(report)
if __name__ == '__main__':
main()