CLOTH_WEB_V1 / cal_compatibility.py
SIUBIU's picture
Update cal_compatibility.py
de57da4 verified
raw
history blame
1.5 kB
import vgg16
import numpy as np
def cosine_similarity(vector1, vector2):
# # 计算向量的点积
# dot_product = np.dot(vector1, vector2)
# # 计算向量的模长
# magnitude_vector1 = np.linalg.norm(vector1)
# magnitude_vector2 = np.linalg.norm(vector2)
# # 计算余弦相似度
# return dot_product / (magnitude_vector1 * magnitude_vector2)
flat_vector1 = vector1.reshape(-1)
flat_vector2 = vector2.reshape(-1)
dot_product = np.dot(flat_vector1, flat_vector2)
norm_vector1 = np.linalg.norm(flat_vector1)
norm_vector2 = np.linalg.norm(flat_vector2)
return dot_product / (norm_vector1 * norm_vector2)
def cal_compatibility():
n = 4096
access_feature = []
cloth_feature = []
for item_id in range(1, 9):
access_feature.append(vgg16.extract_features('downloads/access_' + '%s.jpg' % item_id)[0])
for item_id in range(1, 7):
cloth_feature.append(vgg16.extract_features('downloads/cloth_' + '%s.jpeg' % item_id)[0])
best_score = float('-inf')
best_cloth = 0
best_access = 0
for i in range(1, 9):
for j in range(1, 7):
score = cosine_similarity(access_feature[i - 1], cloth_feature[j - 1])
if score > best_score:
best_score = score
best_cloth = j
best_access = i
print(best_cloth, best_access)
picture = [f"downloads/cloth_{best_cloth}.jpeg", f"downloads/access_{best_access}.jpg"]
return picture