FUNIT/funit_model.py at master · CALM-LMU/FUNIT · GitHub

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
159
160
161
162
"""
Copyright (C) 2019 NVIDIA Corporation.  All rights reserved.
Licensed under the CC BY-NC-SA 4.0 license
(https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).
"""
import copy

import torch
import torch.nn as nn

from networks import FewShotGen, GPPatchMcResDis
from debugUtils import printCheckpoint, Debugger
import torch.nn.functional as F
from globalConstants import GlobalConstants

import apex.amp as amp

PREFIX = "funit_model.py"

def recon_criterion(predict, target):
    if (target.shape[-1]!=predict.shape[-1]):
        print("Funit_model.py recon_criterion: SHAPE OF INPUT", target.shape, "AND OF PREDICTION", predict.shape, "AREN'T EQUAL!")
        predict = F.interpolate(predict, target.shape[-1])
    return torch.mean(torch.abs(predict - target))


class FUNITModel(nn.Module):
    def __init__(self, hp):
        super(FUNITModel, self).__init__()
        self.gen = FewShotGen(hp['gen'])
        self.dis = GPPatchMcResDis(hp['dis'])
        self.gen_test = copy.deepcopy(self.gen)

    def forward(self, co_data, cl_data, hp, mode):

        #debug = Debugger(self.forward.__name__, self.__class__.__name__, PREFIX) #Delete afterwards

        xa = co_data[0].cuda()
        la = co_data[1].cuda()
        xb = cl_data[0].cuda()
        lb = cl_data[1].cuda()
        if mode == 'gen_update':
            c_xa = self.gen.enc_content(xa)
            s_xa = self.gen.enc_class_model(xa)
            s_xb = self.gen.enc_class_model(xb)
            xt = self.gen.decode(c_xa, s_xb)  # translation
            xr = self.gen.decode(c_xa, s_xa)  # reconstruction
            #if (xt.shape[1]!=xa.shape[1]):
            #    print("SHAPE OF INPUT %d AND OF PREDICTION %d AREN'T EQUAL!" % (xa.shape, xt.shape))
            #    xt = F.interpolate(xt, xa.shape[1])
            l_adv_t, gacc_t, xt_gan_feat = self.dis.calc_gen_loss(xt, lb)
            l_adv_r, gacc_r, xr_gan_feat = self.dis.calc_gen_loss(xr, la)
            _, xb_gan_feat = self.dis(xb, lb)
            _, xa_gan_feat = self.dis(xa, la)
            l_c_rec = recon_criterion(xr_gan_feat.mean(3).mean(2),
                                      xa_gan_feat.mean(3).mean(2))
            l_m_rec = recon_criterion(xt_gan_feat.mean(3).mean(2),
                                      xb_gan_feat.mean(3).mean(2))
            l_x_rec = recon_criterion(xr, xa.float())
            l_adv = 0.5 * (l_adv_t + l_adv_r)
            acc = 0.5 * (gacc_t + gacc_r)
            l_total = (hp['gan_w'] * l_adv + hp['r_w'] * l_x_rec + hp[
                'fm_w'] * (l_c_rec + l_m_rec))
            if (GlobalConstants.usingApex):
                with amp.scale_loss(l_total, [self.gen_opt, self.dis_opt]) as scaled_loss:
                    scaled_loss.backward()
            else:
                l_total.backward()
            return l_total, l_adv, l_x_rec, l_c_rec, l_m_rec, acc
        elif mode == 'dis_update':
            xb.requires_grad_()
            l_real_pre, acc_r, resp_r = self.dis.calc_dis_real_loss(xb, lb)
            l_real = hp['gan_w'] * l_real_pre
            if (GlobalConstants.usingApex):
                with amp.scale_loss(l_real, [self.gen_opt, self.dis_opt]) as scaled_loss:
                    scaled_loss.backward(retain_graph=True)
            else:
                l_real.backward(retain_graph=True)
            l_reg_pre = self.dis.calc_grad2(resp_r, xb)
            l_reg = 10 * l_reg_pre
            if (GlobalConstants.usingApex):
                with amp.scale_loss(l_reg, [self.gen_opt, self.dis_opt]) as scaled_loss:
                    scaled_loss.backward()
            else:
                l_reg.backward()
            with torch.no_grad():
                c_xa = self.gen.enc_content(xa)
                s_xb = self.gen.enc_class_model(xb)
                xt = self.gen.decode(c_xa, s_xb)
            l_fake_p, acc_f, resp_f = self.dis.calc_dis_fake_loss(xt.detach(),
                                                                  lb)
            l_fake = hp['gan_w'] * l_fake_p
            if (GlobalConstants.usingApex):
                with amp.scale_loss(l_fake, [self.gen_opt, self.dis_opt]) as scaled_loss:
                    scaled_loss.backward()
            else:
                l_fake.backward()
            l_total = l_fake + l_real + l_reg
            acc = 0.5 * (acc_f + acc_r)
            return l_total, l_fake_p, l_real_pre, l_reg_pre, acc
        else:
            assert 0, 'Not support operation'

    def test(self, co_data, cl_data):
        self.eval()
        self.gen.eval()
        self.gen_test.eval()
        xa = co_data[0].cuda()
        xb = cl_data[0].cuda()
        c_xa_current = self.gen.enc_content(xa)
        s_xa_current = self.gen.enc_class_model(xa)
        s_xb_current = self.gen.enc_class_model(xb)
        xt_current = self.gen.decode(c_xa_current, s_xb_current)
        xr_current = self.gen.decode(c_xa_current, s_xa_current)
        c_xa = self.gen_test.enc_content(xa)
        s_xa = self.gen_test.enc_class_model(xa)
        s_xb = self.gen_test.enc_class_model(xb)
        xt = self.gen_test.decode(c_xa, s_xb)
        xr = self.gen_test.decode(c_xa, s_xa)
        self.train()
        return xa, xr_current, xt_current, xb, xr, xt

    def translate_k_shot(self, co_data, cl_data, k):
        self.eval()
        xa = co_data[0].cuda()
        xb = cl_data[0].cuda()
        c_xa_current = self.gen_test.enc_content(xa)
        if k == 1:
            c_xa_current = self.gen_test.enc_content(xa)
            s_xb_current = self.gen_test.enc_class_model(xb)
            xt_current = self.gen_test.decode(c_xa_current, s_xb_current)
        else:
            s_xb_current_before = self.gen_test.enc_class_model(xb)
            s_xb_current_after = s_xb_current_before.squeeze(-1).permute(1,
                                                                         2,
                                                                         0)
            s_xb_current_pool = torch.nn.functional.avg_pool1d(
                s_xb_current_after, k)
            s_xb_current = s_xb_current_pool.permute(2, 0, 1).unsqueeze(-1)
            xt_current = self.gen_test.decode(c_xa_current, s_xb_current)
        return xt_current

    def compute_k_style(self, style_batch, k):
        self.eval()
        style_batch = style_batch.cuda()
        s_xb_before = self.gen_test.enc_class_model(style_batch)
        s_xb_after = s_xb_before.squeeze(-1).permute(1, 2, 0)
        s_xb_pool = torch.nn.functional.avg_pool1d(s_xb_after, k)
        s_xb = s_xb_pool.permute(2, 0, 1).unsqueeze(-1)
        return s_xb

    def translate_simple(self, content_image, class_code):
        self.eval()
        xa = content_image.cuda()
        s_xb_current = class_code.cuda()
        c_xa_current = self.gen_test.enc_content(xa)
        xt_current = self.gen_test.decode(c_xa_current, s_xb_current)
        return xt_current

    def setOptimizersForApex(self, gen_opt, dis_opt):
        self.gen_opt = gen_opt
        self.dis_opt = dis_opt