File size: 3,629 Bytes
b887ad8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# This file code from T2M(https://github.com/EricGuo5513/text-to-motion), licensed under the https://github.com/EricGuo5513/text-to-motion/blob/main/LICENSE.
# Copyright (c) 2022 Chuan Guo
import torch
from os.path import join as pjoin
import numpy as np
from .evaluator_modules import *


def build_models(opt):
    movement_enc = MovementConvEncoder(opt.dim_pose-4, opt.dim_movement_enc_hidden, opt.dim_movement_latent)
    text_enc = TextEncoderBiGRUCo(word_size=opt.dim_word,
                                  pos_size=opt.dim_pos_ohot,
                                  hidden_size=opt.dim_text_hidden,
                                  output_size=opt.dim_coemb_hidden,
                                  device=opt.device)

    motion_enc = MotionEncoderBiGRUCo(input_size=opt.dim_movement_latent,
                                      hidden_size=opt.dim_motion_hidden,
                                      output_size=opt.dim_coemb_hidden,
                                      device=opt.device)
    checkpoint = torch.load(pjoin(opt.evaluator_dir, opt.dataset_name, 'text_mot_match', 'model', 'finest.tar'),
                            map_location=opt.device)
    movement_enc.load_state_dict(checkpoint['movement_encoder'])
    text_enc.load_state_dict(checkpoint['text_encoder'])
    motion_enc.load_state_dict(checkpoint['motion_encoder'])
    print('\nLoading Evaluation Model Wrapper (Epoch %d) Completed!!' % (checkpoint['epoch']))
    return text_enc, motion_enc, movement_enc

class EvaluatorModelWrapper(object):

    def __init__(self, opt):
        self.text_encoder, self.motion_encoder, self.movement_encoder = build_models(opt)
        self.opt = opt
        self.device = opt.device

        self.text_encoder.to(opt.device)
        self.motion_encoder.to(opt.device)
        self.movement_encoder.to(opt.device)

        self.text_encoder.eval()
        self.motion_encoder.eval()
        self.movement_encoder.eval()

    # Please note that the results does not following the order of inputs
    def get_co_embeddings(self, word_embs, pos_ohot, cap_lens, motions, m_lens):
        with torch.no_grad():
            word_embs = word_embs.detach().to(self.device).float()
            pos_ohot = pos_ohot.detach().to(self.device).float()
            motions = motions.detach().to(self.device).float()

            align_idx = np.argsort(m_lens.data.tolist())[::-1].copy()
            motions = motions[align_idx]
            m_lens = m_lens[align_idx]

            '''Movement Encoding'''
            movements = self.movement_encoder(motions[..., :-4]).detach()
            m_lens = torch.div(m_lens, self.opt.unit_length, rounding_mode='trunc') 
            motion_embedding = self.motion_encoder(movements, m_lens)

            '''Text Encoding'''
            text_embedding = self.text_encoder(word_embs, pos_ohot, cap_lens)
            text_embedding = text_embedding[align_idx]
        return text_embedding, motion_embedding

    # Please note that the results does not following the order of inputs
    def get_motion_embeddings(self, motions, m_lens):
        with torch.no_grad():
            motions = motions.detach().to(self.device).float()

            align_idx = np.argsort(m_lens.data.tolist())[::-1].copy()
            motions = motions[align_idx]
            m_lens = m_lens[align_idx]

            '''Movement Encoding'''
            movements = self.movement_encoder(motions[..., :-4]).detach()
            m_lens = torch.div(m_lens, self.opt.unit_length, rounding_mode='trunc') 
            motion_embedding = self.motion_encoder(movements, m_lens)
        return motion_embedding