47 lines
1.0 KiB
Python
47 lines
1.0 KiB
Python
import json
|
|
|
|
import numpy as np
|
|
|
|
#datum = np.load('dump.npz.npy')
|
|
datum = np.load('dump_so400m.npy')
|
|
|
|
|
|
with open('dump.json','r') as rr:
|
|
js = json.load(rr)
|
|
|
|
|
|
emb_dict = dict()
|
|
for embed in js[1]:
|
|
fr = embed['frame_number']
|
|
vec = embed['vector']
|
|
emb_dict[fr] = np.asarray(vec)
|
|
|
|
|
|
|
|
|
|
|
|
def cosine_sim(emb_in_1, emb_in_2):
|
|
emb_in_1 = emb_in_1.astype(np.float32)
|
|
emb_in_2 = emb_in_2.astype(np.float32)
|
|
emb1_norm = np.linalg.norm(emb_in_1)
|
|
emb2_norm = np.linalg.norm(emb_in_2)
|
|
dot_prod = np.dot(emb_in_1, emb_in_2)
|
|
similarity = dot_prod/(emb1_norm * emb2_norm)
|
|
|
|
return np.round(emb1_norm,5),np.round( emb2_norm,5), np.round( dot_prod,5),np.round( similarity,5)
|
|
|
|
|
|
arr_in_deepstream = np.asarray([y for _,y in emb_dict.items()])
|
|
normed = np.divide(datum.T, np.linalg.norm(datum, axis=1)).T
|
|
print('_________________________')
|
|
|
|
|
|
print(len(emb_dict))
|
|
print(len(datum))
|
|
for fr, emb in emb_dict.items():
|
|
emb1 = np.linalg.norm(emb)
|
|
emb2 = np.linalg.norm(datum[fr])
|
|
print( cosine_sim(emb, datum[fr]))
|
|
|
|
|