-
Notifications
You must be signed in to change notification settings - Fork 0
/
vqvae_eval.py
154 lines (139 loc) · 6.42 KB
/
vqvae_eval.py
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
import os
import torch
import numpy as np
from torch.utils.tensorboard import SummaryWriter
import json
import clip
import options.option_transformer as option_trans
import models.vqvae as vqvae
from models.vqvae_multi import VQVAE_MULTI
from models.vqvae_general import VQVAE_decode
import utils.utils_model as utils_model
import utils.eval_trans as eval_trans
from dataset import dataset_TM_eval
import models.t2m_trans as trans
import models.t2m_timesformer as trans_time
from options.get_eval_option import get_opt
import options.option_vq as option_vq
from models.evaluator_wrapper import EvaluatorModelWrapper
import warnings
warnings.filterwarnings('ignore')
from exit.utils import base_dir, init_save_folder
##### ---- Exp dirs ---- #####
args = option_vq.get_args_parser()
torch.manual_seed(args.seed)
args.out_dir = f'{args.out_dir}/eval'
os.makedirs(args.out_dir, exist_ok = True)
init_save_folder(args)
##### ---- Logger ---- #####
logger = utils_model.get_logger(args.out_dir)
writer = SummaryWriter(args.out_dir)
logger.info(json.dumps(vars(args), indent=4, sort_keys=True))
from utils.word_vectorizer import WordVectorizer
w_vectorizer = WordVectorizer('./glove', 'our_vab')
#TODO
val_loader = dataset_TM_eval.DATALoader(args.dataname, True, 32, w_vectorizer)
dataset_opt_path = 'checkpoints/kit/Comp_v6_KLD005/opt.txt' if args.dataname == 'kit' else 'checkpoints/t2m/Comp_v6_KLD005/opt.txt'
wrapper_opt = get_opt(dataset_opt_path, torch.device('cuda'))
eval_wrapper = EvaluatorModelWrapper(wrapper_opt)
##### ---- Network ---- #####
clip_model, clip_preprocess = clip.load("ViT-B/32", device=torch.device('cuda'), jit=False) # Must set jit=False for training
clip.model.convert_weights(clip_model) # Actually this line is unnecessary since clip by default already on float16
clip_model.eval()
for p in clip_model.parameters():
p.requires_grad = False
# https://github.com/openai/CLIP/issues/111
class TextCLIP(torch.nn.Module):
def __init__(self, model) :
super(TextCLIP, self).__init__()
self.model = model
def forward(self,text):
with torch.no_grad():
word_emb = self.model.token_embedding(text).type(self.model.dtype)
word_emb = word_emb + self.model.positional_embedding.type(self.model.dtype)
word_emb = word_emb.permute(1, 0, 2) # NLD -> LND
word_emb = self.model.transformer(word_emb)
word_emb = self.model.ln_final(word_emb).permute(1, 0, 2).float()
enctxt = self.model.encode_text(text).float()
return enctxt, word_emb
clip_model = TextCLIP(clip_model)
if args.teacher_pth:
net= VQVAE_MULTI(args, ## use args to define different parameters in different quantizers
args.nb_code,#8192
args.code_dim,#32
args.down_t,#2
args.stride_t,#2
args.width,#512
args.depth,#3
args.dilation_growth_rate,#3
args.vq_act,#'relu'
None,#None
{'mean': torch.from_numpy(val_loader.dataset.mean).cuda().float(),
'std': torch.from_numpy(val_loader.dataset.std).cuda().float()},
True)
logger.info('loading teacher checkpoint from {}'.format(args.teacher_pth))
ckpt = torch.load(args.teacher_pth, map_location='cpu')
net.load_state_dict(ckpt['net'], strict=True)
elif args.resume_pth:
teacher_net= VQVAE_MULTI(args, ## use args to define different parameters in different quantizers
args.nb_code,#8192
args.code_dim,#32
args.down_t,#2
args.stride_t,#2
args.width,#512
args.depth,#3
args.dilation_growth_rate,#3
args.vq_act,#'relu'
None,#None
{'mean': torch.from_numpy(val_loader.dataset.mean).cuda().float(),
'std': torch.from_numpy(val_loader.dataset.std).cuda().float()},
True)
net= VQVAE_decode(args, ## use args to define different parameters in different quantizers
teacher_net,
args.nb_code,#8192
args.code_dim,#32
args.down_t,#2
args.stride_t,#2
args.width,#512
args.depth,#3
args.dilation_growth_rate,#3
args.vq_act,#'relu'
None,#None
)
print ('loading checkpoint from {}'.format(args.resume_pth))
logger.info('loading checkpoint from {}'.format(args.resume_pth))
ckpt = torch.load(args.resume_pth, map_location='cpu')
net.load_state_dict(ckpt['net'], strict=True)
net.eval()
net.cuda()
fid = []
div = []
top1 = []
top2 = []
top3 = []
matching = []
repeat_time = 20
from tqdm import tqdm
for i in tqdm(range(repeat_time)):
best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_vqvae(args.out_dir, val_loader, net, logger, writer, 0, best_fid=1000, best_iter=0, best_div=100, best_top1=0, best_top2=0, best_top3=0, best_matching=100, eval_wrapper=eval_wrapper,wo_trajectory=args.wo_trajectory)
fid.append(best_fid)
div.append(best_div)
top1.append(best_top1)
top2.append(best_top2)
top3.append(best_top3)
matching.append(best_matching)
print('final result:')
print('fid: ', sum(fid)/repeat_time)
print('div: ', sum(div)/repeat_time)
print('top1: ', sum(top1)/repeat_time)
print('top2: ', sum(top2)/repeat_time)
print('top3: ', sum(top3)/repeat_time)
print('matching: ', sum(matching)/repeat_time)
fid = np.array(fid)
div = np.array(div)
top1 = np.array(top1)
top2 = np.array(top2)
top3 = np.array(top3)
matching = np.array(matching)
msg_final = f"FID. {np.mean(fid):.3f}, conf. {np.std(fid)*1.96/np.sqrt(repeat_time):.3f}, Diversity. {np.mean(div):.3f}, conf. {np.std(div)*1.96/np.sqrt(repeat_time):.3f}, TOP1. {np.mean(top1):.3f}, conf. {np.std(top1)*1.96/np.sqrt(repeat_time):.3f}, TOP2. {np.mean(top2):.3f}, conf. {np.std(top2)*1.96/np.sqrt(repeat_time):.3f}, TOP3. {np.mean(top3):.3f}, conf. {np.std(top3)*1.96/np.sqrt(repeat_time):.3f}, Matching. {np.mean(matching):.3f}, conf. {np.std(matching)*1.96/np.sqrt(repeat_time):.3f}"
logger.info(msg_final)