/
baselines.py
158 lines (156 loc) · 8.27 KB
/
baselines.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
155
156
157
158
import random
import json
import numpy as np
from collections import defaultdict
import argparse
import os
from tqdm import tqdm
import torch
from transformers import AutoTokenizer, AutoModel
from sentence_transformers import SentenceTransformer
import models.vqvae as vqvae
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--vq-dir")
parser.add_argument("--output-dir")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--params-path")
parser.add_argument("--mean-std-path")
parser.add_argument("--train-segments-path")
parser.add_argument("--val-segments-path")
parser.add_argument("--max-motion-length", type=int)
parser.add_argument("--history-size", type=int)
parser.add_argument("--nearest-neighbor", action="store_true")
parser.add_argument("--normalize", action="store_true")
parser.add_argument("--embedding-model-name")
parser.add_argument("--batch-size", type=int, default=8)
parser.add_argument("--fps", type=int, default=30)
parser.add_argument("--static-face", action="store_true")
parser.add_argument("--random-train-select", action="store_true")
args = parser.parse_args()
os.system('mkdir '+args.output_dir)
random.seed(args.seed)
np.random.seed(seed=args.seed)
if args.mean_std_path is not None:
mean = np.load(os.path.join(args.mean_std_path, 'mean.npy'))
std = np.load(os.path.join(args.mean_std_path, 'std.npy'))
with open(args.params_path) as f:
params = json.load(f)
for key in params:
if not hasattr(args, key):
setattr(args, key, params[key])
if args.nearest_neighbor or args.random_train_select:
segments = torch.load(args.train_segments_path, map_location='cpu')
text_to_motion = {}
fps = args.fps
text_to_file_id = {}
for seg in segments:
if seg['split_end_frame']-seg['split_start_frame'] < args.max_motion_length:
continue
for start in range(seg['split_start_frame'], seg['split_end_frame']-args.max_motion_length+1):
words = [word for word in seg['before_words']+seg['during_words'] if word['end']*fps >= start-args.history_size*fps and word['end']*fps < start+args.max_motion_length]
text = ' '.join([word['text'] for word in words])
if text not in text_to_motion:
text_to_motion[text] = torch.cat((seg['p0_exp'][start-seg['split_start_frame']:start-seg['split_start_frame']+args.max_motion_length,:], seg['p0_pose'][start-seg['split_start_frame']:start-seg['split_start_frame']+args.max_motion_length,:]), dim=1).numpy()
text_to_file_id[text] = seg['fname']+'_'+str(start)
# text_to_motion[text] = torch.cat((seg['p0_exp'], seg['p0_pose']), dim=1)
all_texts = []
text_id_to_motion = {}
file_ids = []
for i, text in enumerate(text_to_motion):
all_texts.append(text)
text_id_to_motion[i] = text_to_motion[text]
file_ids.append(text_to_file_id[text])
if args.nearest_neighbor:
text_embeddings = []
tokenizer = AutoTokenizer.from_pretrained(args.embedding_model_name)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = SentenceTransformer(args.embedding_model_name).eval()
if torch.cuda.is_available():
model = model.cuda()
# bos_token = ''
for i in tqdm(range(0, len(all_texts), args.batch_size)):
sentence_embeddings = torch.from_numpy(model.encode(all_texts[i:i+args.batch_size]))
for j in range(sentence_embeddings.shape[0]):
text_embeddings.append(sentence_embeddings[j:j+1,:].cpu())
text_embeddings = torch.cat(text_embeddings, dim=0)
assert text_embeddings.shape[0] == len(all_texts), str(text_embeddings.shape)+', '+str(len(all_texts))
print(text_embeddings.shape)
print(text_embeddings[:2,:])
if torch.cuda.is_available():
text_embeddings = text_embeddings.cuda()
if args.normalize:
# text_embeddings = text_embeddings / torch.linalg.norm(text_embeddings, dim=-1, keepdim=True)
text_embeddings = torch.nn.functional.normalize(text_embeddings, p=2, dim=1)
val_segments = torch.load(args.val_segments_path, map_location="cpu")
val_segments_dict = {}
for seg in val_segments:
for i in range(seg['split_start_frame'], seg['split_end_frame']):
val_segments_dict[seg['fname'].split('/')[-1]+'_'+str(i)] = seg
# print(val_segments_dict.keys())
else:
net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
args.nb_code,
args.code_dim,
args.output_emb_width,
args.down_t,
args.stride_t,
args.width,
args.depth,
args.dilation_growth_rate)
ckpt = torch.load(args.resume_pth, map_location='cpu')
net.load_state_dict(ckpt['net'], strict=True)
net.eval()
if torch.cuda.is_available():
net.cuda()
chosen = defaultdict(int)
count = 0
for root, _, files in tqdm(os.walk(args.vq_dir)):
for fname in files:
if fname[-4:] == ".npy":
count += 1
gt_vq = np.load(os.path.join(root, fname))
num_frames = gt_vq.reshape(-1).shape[0]*(2**args.down_t)
if args.nearest_neighbor:
seg = val_segments_dict[fname.split('.npy')[0]]
start_frame = int(fname.split('.npy')[0].split('_')[-1])
words = [word['text'] for word in seg['before_words']+seg['during_words'] if word['end']*fps >= start_frame-args.history_size*fps and word['end']*fps < start_frame+num_frames]
text = ' '.join(words)
embedding = torch.from_numpy(model.encode([text])).view(1, -1)
if torch.cuda.is_available():
embedding = embedding.to('cuda:0')
if args.normalize:
# text_embeddings = text_embeddings / torch.linalg.norm(text_embeddings, dim=-1, keepdim=True)
embedding = torch.nn.functional.normalize(embedding, p=2, dim=1)
best_index = (text_embeddings @ embedding.t()).view(-1).argmax().item()
# if count > 180:
print(fname, text, '|||', all_texts[best_index])
chosen[best_index] += 1
motion = text_id_to_motion[best_index][:num_frames,:]
elif args.random_train_select:
index = random.choice(list(range(len(text_id_to_motion))))
motion = text_id_to_motion[index][:num_frames,:]
elif args.static_face:
motion = np.expand_dims(mean, axis=0)
if len(gt_vq.shape) == 3:
motion = np.repeat(np.expand_dims(motion, axis=0), num_frames, axis=1)
else:
motion = np.repeat(motion, num_frames, axis=0)
else:
random_pred = np.random.randint(low=0, high=args.nb_code, size=gt_vq.shape)
inp = torch.from_numpy(random_pred).view(1, -1)
if torch.cuda.is_available():
inp = inp.cuda()
with torch.no_grad():
decoded = net.forward_decoder(inp)
motion = decoded.cpu().view(-1, 56).numpy()
motion = (motion*std.reshape(1, -1))+mean.reshape(1, -1)
path_parts = os.path.join(root, fname).replace('.npy', '_pred.npy').split('/')
new_path = os.path.join(args.output_dir, *path_parts[-4:])
path_parts = new_path.split('/')
# print(path_parts)
for j in range(len(path_parts)-1):
if not os.path.exists('/'.join(path_parts[:j+1])):
os.system('mkdir '+'/'.join(path_parts[:j+1]))
np.save(new_path, motion)