Hugging Face's logo

JayceAnova
/
Benchmark-Single

Model card Files
xet
Community
Benchmark-Single
File size: 4,598 Bytes 
811e03d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
 
import os
import collections
import json
import logging
import argparse
import numpy as np
import pandas as pd
import torch
from time import time
from torch import optim
from tqdm import tqdm
from torch.utils.data import DataLoader

from rq_llama import *
from index.datasets import EmbDataset

def if_collided(all_indices_str):
    tot_item = len(all_indices_str)
    tot_indice = len(set(all_indices_str.tolist()))
    return tot_item == tot_indice

def get_indices_count(all_indices_str):
    indices_count = collections.defaultdict(int)
    for index in all_indices_str:
        indices_count[index] += 1
    return indices_count

def get_collision_item(all_indices_str):
    index2id = {}
    for i, index in enumerate(all_indices_str):
        if index not in index2id:
            index2id[index] = []
        index2id[index].append(i)
    collision_item_groups = []
    for index in index2id:
        if len(index2id[index]) > 1:
            collision_item_groups.append(index2id[index])
    return collision_item_groups

def parse_args():
    parser = argparse.ArgumentParser(description = "Index")
    parser.add_argument("--ckpt_path", type = str, default = "", help = "")
    parser.add_argument("--data_path", type = str, default = "", help = "")
    parser.add_argument("--save_path", type = str, default = "", help = "")
    parser.add_argument("--device_map", type = str, default = "1", help = "gpu or cpu")
    return parser.parse_args()

args = parse_args()
print(args)

data = EmbDataset(args.data_path)
data_loader = DataLoader(data, num_workers = 4, batch_size = 64, shuffle = False, pin_memory = True)
device_map = {'': int(args.device_map)}
MODEL = LlamaWithRQ.from_pretrained(args.ckpt_path, torch_dtype = torch.float16, low_cpu_mem_usage = True, device_map = device_map)
MODEL.eval()
device = MODEL.device
rqvae = MODEL.rqvae
prefix = MODEL.prefix

postfix = '<p-{}>'
index_table = {}

all_indices = []
all_indices_str = []
with torch.no_grad():
    for x in tqdm(data_loader):
        indices = rqvae.get_indices(x.to(device), False)
        indices = indices.view(-1, indices.shape[-1]).cpu().numpy()
        for index in indices:
            code = []
            for i, ind in enumerate(index):
                code.append(prefix[i].format(int(ind)))

            if str(code) in index_table:
                index_table[str(code)] += 1
            else:
                index_table[str(code)] = 0
            code.append(postfix.format(index_table[str(code)]))

            all_indices.append(code)
            all_indices_str.append(str(code))

all_indices = np.array(all_indices)
all_indices_str = np.array(all_indices_str)

for vq in rqvae.rq.vq_layers[:-1]:
    vq.sk_epsilon=0.0
if rqvae.rq.vq_layers[-1].sk_epsilon == 0.0:
    rqvae.rq.vq_layers[-1].sk_epsilon = 0.003

'''

tt = 0

while True:

    if tt >= 20 or if_collided(all_indices_str):

        break



    collision_item_groups = get_collision_item(all_indices_str)

    # print(collision_item_groups)

    print(len(collision_item_groups))

    with torch.no_grad():

        for collision_items in collision_item_groups:

            indices = rqvae.get_indices(data[collision_items].to(device), True)

            indices = indices.view(-1, indices.shape[-1]).cpu().numpy()

            for item, index in zip(collision_items, indices):

                code = []

                for i, ind in enumerate(index):

                    code.append(prefix[i].format(int(ind)))



                all_indices[item] = code

                all_indices_str[item] = str(code)

        tt += 1

'''

print("All indices number: ", len(all_indices))
print("Max number of conflicts: ", max(get_indices_count(all_indices_str).values()))
print('Re-index number:', max(index_table.values()))

# tot_item = len(all_indices_str)
# tot_indice = len(set(all_indices_str.tolist()))
# print("Collision Rate", (tot_item - tot_indice) / tot_item)

all_indices_dict = {}
for item, indices in enumerate(all_indices.tolist()):
    all_indices_dict[item] = list(indices)

reindex_dict = {'reindex': max(index_table.values())}

json_path = os.path.join(args.save_path,'indices.json')
with open(json_path, 'w',encoding = 'utf-8') as f:
    json.dump(all_indices_dict, f)

reindex_path = os.path.join(args.save_path,'reindex.json')
with open(reindex_path, 'w',encoding = 'utf-8') as f:
    json.dump(reindex_dict, f)

# with open(args.save_path, 'w',encoding = 'utf-8') as f:
#     json.dump(all_indices_dict, f)