singan_model.py 12.9 KB
Newer Older
B
BrilliantYuKaimin 已提交
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import os
import cv2
import math
import warnings
from collections import OrderedDict
from sklearn.cluster import KMeans

import paddle
import paddle.nn.functional as F
import paddle.vision.transforms as T

from .base_model import BaseModel
from .builder import MODELS
from .generators.builder import build_generator
from .criterions.builder import build_criterion
from .discriminators.builder import build_discriminator
from ..solver import build_lr_scheduler, build_optimizer
warnings.filterwarnings('ignore', category=DeprecationWarning)
warnings.filterwarnings('ignore', category=UserWarning)


def pad_shape(shape, pad_size):
    shape[-2] += 2 * pad_size
    shape[-1] += 2 * pad_size
    return shape

def quant(x, num):
    n, c, h, w = x.shape
    kmeans = KMeans(num, random_state=0).fit(x.transpose([0, 2, 3, 1]).reshape([-1, c]))
    centers = kmeans.cluster_centers_
    x = centers[kmeans.labels_].reshape([n, h, w, c]).transpose([0, 3, 1, 2])
    return paddle.to_tensor(x, 'float32'), centers

def quant_to_centers(x, centers):
    n, c, h, w = x.shape
    num = centers.shape[0]
    kmeans = KMeans(num, init=centers, n_init=1).fit(x.transpose([0, 2, 3, 1]).reshape([-1, c]))
    x = centers[kmeans.labels_].reshape([n, h, w, c]).transpose([0, 3, 1, 2])
    return paddle.to_tensor(x, 'float32')

@MODELS.register()
class SinGANModel(BaseModel):
    def __init__(self,
                 generator,
                 discriminator,
                 gan_criterion=None,
                 recon_criterion=None,
                 gp_criterion=None,
                 train_image=None,
                 scale_factor=0.75,
                 min_size=25,
                 is_finetune=False,
                 finetune_scale=1,
                 color_num=5,
                 gen_iters=3,
                 disc_iters=3,
                 noise_amp_init=0.1):
        super(SinGANModel, self).__init__()
        
        # setup config
        self.gen_iters = gen_iters
        self.disc_iters = disc_iters
        self.min_size = min_size
        self.is_finetune = is_finetune
        self.noise_amp_init = noise_amp_init
        self.train_image = T.Compose([
            T.Transpose(),
            T.Normalize(127.5, 127.5)
        ])(cv2.cvtColor(cv2.imread(train_image, cv2.IMREAD_COLOR), cv2.COLOR_BGR2RGB))
        self.train_image = paddle.to_tensor(self.train_image).unsqueeze(0)
        self.scale_num = math.ceil(math.log(
            self.min_size / min(self.train_image.shape[-2:]), 
            scale_factor)) + 1
        self.scale_factor = math.pow(
            self.min_size / min(self.train_image.shape[-2:]), 
            1 / (self.scale_num - 1))
        self.reals = [
            F.interpolate(self.train_image, None, self.scale_factor ** i, 'bicubic')
        for i in range(self.scale_num - 1, -1, -1)]

        # build generator
        generator['scale_num'] = self.scale_num
        generator['coarsest_shape'] =self.reals[0].shape
        self.nets['netG'] = build_generator(generator)
        self.niose_pad_size = 0 if generator.get('noise_zero_pad', True) \
                                else self.nets['netG']._pad_size
        self.nets['netG'].scale_factor = paddle.to_tensor(self.scale_factor, 'float32')

        # build discriminator
        nfc_init = discriminator.pop('nfc_init', 32)
        min_nfc_init = discriminator.pop('min_nfc_init', 32)
        for i in range(self.scale_num):
            discriminator['nfc'] = min(nfc_init * pow(2, math.floor(i / 4)), 128)
            discriminator['min_nfc'] = min(min_nfc_init * pow(2, math.floor(i / 4)), 128)
            self.nets[f'netD{i}'] = build_discriminator(discriminator)

        # build criterion
        self.gan_criterion = build_criterion(gan_criterion)
        self.recon_criterion = build_criterion(recon_criterion)
        self.gp_criterion = build_criterion(gp_criterion)
        
        if self.is_finetune:
            self.finetune_scale = finetune_scale
            self.quant_real, self.quant_centers = quant(self.reals[finetune_scale], color_num)


        # setup training config
        self.lr_schedulers = OrderedDict()
        self.current_scale = (finetune_scale if self.is_finetune else 0) - 1
        self.current_iter = 0

    def set_total_iter(self, total_iter):
        super().set_total_iter(total_iter)
        if self.is_finetune:
            self.scale_iters = total_iter
        else:
            self.scale_iters = math.ceil(total_iter / self.scale_num)

    def setup_lr_schedulers(self, cfg):
        for i in range(self.scale_num):
            self.lr_schedulers[f"lr{i}"] = build_lr_scheduler(cfg)
        return self.lr_schedulers

    def setup_optimizers(self, lr_schedulers, cfg):
        for i in range(self.scale_num):
            self.optimizers[f'optim_netG{i}'] = build_optimizer(
                cfg['optimizer_G'], lr_schedulers[f"lr{i}"], self.nets[f'netG'].generators[i].parameters())
            self.optimizers[f'optim_netD{i}'] = build_optimizer(
                cfg['optimizer_D'], lr_schedulers[f"lr{i}"], self.nets[f'netD{i}'].parameters())
        return self.optimizers

    def setup_input(self, input):
        pass

    def backward_D(self):
        self.loss_D_real = self.gan_criterion(self.pred_real, True, True)
        self.loss_D_fake = self.gan_criterion(self.pred_fake, False, True)
        self.loss_D_gp = self.gp_criterion(self.nets[f'netD{self.current_scale}'], 
                                           self.real_img, 
                                           self.fake_img)
        self.loss_D = self.loss_D_real + self.loss_D_fake + self.loss_D_gp
        self.loss_D.backward()

        self.losses[f'scale{self.current_scale}/D_total_loss'] = self.loss_D
        self.losses[f'scale{self.current_scale}/D_real_loss'] = self.loss_D_real
        self.losses[f'scale{self.current_scale}/D_fake_loss'] = self.loss_D_fake
        self.losses[f'scale{self.current_scale}/D_gradient_penalty'] = self.loss_D_gp
        
    def backward_G(self):
        self.loss_G_gan = self.gan_criterion(self.pred_fake, True, False)
        self.loss_G_recon = self.recon_criterion(self.recon_img, self.real_img)
        self.loss_G = self.loss_G_gan + self.loss_G_recon
        self.loss_G.backward()

        self.losses[f'scale{self.current_scale}/G_adv_loss'] = self.loss_G_gan
        self.losses[f'scale{self.current_scale}/G_recon_loss'] = self.loss_G_recon

    def scale_prepare(self):
        self.real_img = self.reals[self.current_scale]
        self.lr_scheduler = self.lr_schedulers[f"lr{self.current_scale}"]
        for i in range(self.current_scale):
            self.optimizers.pop(f'optim_netG{i}', None)
            self.optimizers.pop(f'optim_netD{i}', None)
        self.losses.clear()
        self.visual_items.clear()
        self.visual_items[f'real_img_scale{self.current_scale}'] = self.real_img
        if self.is_finetune:
            self.visual_items['quant_real'] = self.quant_real
        
        self.recon_prev = paddle.zeros_like(self.reals[0])
        if self.current_scale > 0:
            z_pyramid = []
            for i in range(self.current_scale):
                if i == 0:
                    z = self.nets['netG'].z_fixed
                else:
                    z = paddle.zeros(
                        pad_shape(
                            self.reals[i].shape, self.niose_pad_size))
                z_pyramid.append(z)
            self.recon_prev = self.nets['netG'](
                z_pyramid, self.recon_prev, 
                self.current_scale - 1, 0).detach()
            self.recon_prev = F.interpolate(
                self.recon_prev, self.real_img.shape[-2:], None, 'bicubic')
            if self.is_finetune:
                self.recon_prev = quant_to_centers(self.recon_prev, self.quant_centers)
            self.nets['netG'].sigma[self.current_scale] = F.mse_loss(
                self.real_img, self.recon_prev
            ).sqrt() * self.noise_amp_init

        for i in range(self.scale_num):
            self.set_requires_grad(self.nets['netG'].generators[i], i == self.current_scale)
        
    def forward(self):
        if not self.is_finetune:
            self.fake_img = self.nets['netG'](
                self.z_pyramid, 
                paddle.zeros(
                    pad_shape(self.z_pyramid[0].shape, -self.niose_pad_size)), 
                self.current_scale, 0)
        else:
            x_prev = self.nets['netG'](
                self.z_pyramid[:self.finetune_scale], 
                paddle.zeros(
                    pad_shape(self.z_pyramid[0].shape, -self.niose_pad_size)), 
                self.finetune_scale - 1, 0)
            x_prev = F.interpolate(x_prev, self.z_pyramid[self.finetune_scale].shape[-2:], None, 'bicubic')
            x_prev_quant = quant_to_centers(x_prev, self.quant_centers)
            self.fake_img = self.nets['netG'](
                self.z_pyramid[self.finetune_scale:], 
                x_prev_quant, 
                self.current_scale, self.finetune_scale)

        self.recon_img = self.nets['netG'](
            [(paddle.randn if self.current_scale == 0 else paddle.zeros)(
                pad_shape(self.real_img.shape, self.niose_pad_size))],
            self.recon_prev, 
            self.current_scale, 
            self.current_scale)

        self.pred_real = self.nets[f'netD{self.current_scale}'](self.real_img)
        self.pred_fake = self.nets[f'netD{self.current_scale}'](
            self.fake_img.detach() if self.update_D else self.fake_img)

        self.visual_items[f'fake_img_scale{self.current_scale}'] = self.fake_img
        self.visual_items[f'recon_img_scale{self.current_scale}'] = self.recon_img
        if self.is_finetune:
            self.visual_items[f'prev_img_scale{self.current_scale}'] = x_prev
            self.visual_items[f'quant_prev_img_scale{self.current_scale}'] = x_prev_quant

    def train_iter(self, optimizers=None):
        if self.current_iter % self.scale_iters == 0:
            self.current_scale += 1
            self.scale_prepare()

        self.z_pyramid = [paddle.randn(
            pad_shape(self.reals[i].shape, self.niose_pad_size))
        for i in range(self.current_scale + 1)]

        self.update_D = (self.current_iter % (self.disc_iters + self.gen_iters) < self.disc_iters)
        self.set_requires_grad(self.nets[f'netD{self.current_scale}'], self.update_D)
        self.forward()
        if self.update_D:
            optimizers[f'optim_netD{self.current_scale}'].clear_grad()
            self.backward_D()
            optimizers[f'optim_netD{self.current_scale}'].step()
        else:
            optimizers[f'optim_netG{self.current_scale}'].clear_grad()
            self.backward_G()
            optimizers[f'optim_netG{self.current_scale}'].step()

        self.current_iter += 1

    def test_iter(self, metrics=None):
        z_pyramid = [paddle.randn(
            pad_shape(self.reals[i].shape, self.niose_pad_size))
        for i in range(self.scale_num)]
        self.nets['netG'].eval()
        fake_img = self.nets['netG'](
            z_pyramid, 
            paddle.zeros(pad_shape(z_pyramid[0].shape, -self.niose_pad_size)), 
            self.scale_num - 1, 0)
        self.visual_items['fake_img_test'] = fake_img
        with paddle.no_grad():
            if metrics is not None:
                for metric in metrics.values():
                    metric.update(fake_img, self.train_image)
        self.nets['netG'].train()
        
    class InferGenerator(paddle.nn.Layer):
        def set_config(self, generator, noise_shapes, scale_num):
            self.generator = generator
            self.noise_shapes = noise_shapes
            self.scale_num = scale_num

        def forward(self, x):
            coarsest_shape = self.generator._coarsest_shape
            z_pyramid = [paddle.randn(shp) for shp in self.noise_shapes]
            x_init = paddle.zeros(coarsest_shape)
            out = self.generator(z_pyramid, x_init, self.scale_num - 1, 0)
            return out

    def export_model(self,
                     export_model=None,
                     output_dir=None,
                     inputs_size=None,
                     export_serving_model=False):
        noise_shapes = [pad_shape(x.shape, self.niose_pad_size) for x in self.reals]
        infer_generator = self.InferGenerator()
        infer_generator.set_config(self.nets['netG'], noise_shapes, self.scale_num)
        paddle.jit.save(infer_generator,
                        os.path.join(output_dir, "singan_random_sample"),
                        input_spec=[1])