From 61a86a3c1e52b62c130a90b7c8b428d1447ebd03 Mon Sep 17 00:00:00 2001 From: James Betker Date: Thu, 7 Jan 2021 10:20:15 -0700 Subject: [PATCH] VQVAE --- codes/models/vqvae/__init__.py | 0 codes/models/vqvae/vqvae.py | 249 +++++++++++++++++++++++++ codes/scripts/extract_square_images.py | 8 +- codes/train.py | 2 +- 4 files changed, 254 insertions(+), 5 deletions(-) create mode 100644 codes/models/vqvae/__init__.py create mode 100644 codes/models/vqvae/vqvae.py diff --git a/codes/models/vqvae/__init__.py b/codes/models/vqvae/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/codes/models/vqvae/vqvae.py b/codes/models/vqvae/vqvae.py new file mode 100644 index 00000000..d1418d1d --- /dev/null +++ b/codes/models/vqvae/vqvae.py @@ -0,0 +1,249 @@ +# Copyright 2018 The Sonnet Authors. All Rights Reserved. +# +# 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. +# ============================================================================ + + +# Borrowed from https://github.com/rosinality/vq-vae-2-pytorch +# Which was itself orrowed from https://github.com/deepmind/sonnet + + +import torch +from torch import nn +from torch.nn import functional as F + +import torch.distributed as distributed + +from trainer.networks import register_model +from utils.util import checkpoint, opt_get + + +class Quantize(nn.Module): + def __init__(self, dim, n_embed, decay=0.99, eps=1e-5): + super().__init__() + + self.dim = dim + self.n_embed = n_embed + self.decay = decay + self.eps = eps + + embed = torch.randn(dim, n_embed) + self.register_buffer("embed", embed) + self.register_buffer("cluster_size", torch.zeros(n_embed)) + self.register_buffer("embed_avg", embed.clone()) + + def forward(self, input): + flatten = input.reshape(-1, self.dim) + dist = ( + flatten.pow(2).sum(1, keepdim=True) + - 2 * flatten @ self.embed + + self.embed.pow(2).sum(0, keepdim=True) + ) + _, embed_ind = (-dist).max(1) + embed_onehot = F.one_hot(embed_ind, self.n_embed).type(flatten.dtype) + embed_ind = embed_ind.view(*input.shape[:-1]) + quantize = self.embed_code(embed_ind) + + if self.training: + embed_onehot_sum = embed_onehot.sum(0) + embed_sum = flatten.transpose(0, 1) @ embed_onehot + + if distributed.is_initialized() and distributed.get_world_size() > 1: + distributed.all_reduce(embed_onehot_sum) + distributed.all_reduce(embed_sum) + + self.cluster_size.data.mul_(self.decay).add_( + embed_onehot_sum, alpha=1 - self.decay + ) + self.embed_avg.data.mul_(self.decay).add_(embed_sum, alpha=1 - self.decay) + n = self.cluster_size.sum() + cluster_size = ( + (self.cluster_size + self.eps) / (n + self.n_embed * self.eps) * n + ) + embed_normalized = self.embed_avg / cluster_size.unsqueeze(0) + self.embed.data.copy_(embed_normalized) + + diff = (quantize.detach() - input).pow(2).mean() + quantize = input + (quantize - input).detach() + + return quantize, diff, embed_ind + + def embed_code(self, embed_id): + return F.embedding(embed_id, self.embed.transpose(0, 1)) + + +class ResBlock(nn.Module): + def __init__(self, in_channel, channel): + super().__init__() + + self.conv = nn.Sequential( + nn.ReLU(inplace=True), + nn.Conv2d(in_channel, channel, 3, padding=1), + nn.ReLU(inplace=True), + nn.Conv2d(channel, in_channel, 1), + ) + + def forward(self, input): + out = self.conv(input) + out += input + + return out + + +class Encoder(nn.Module): + def __init__(self, in_channel, channel, n_res_block, n_res_channel, stride): + super().__init__() + + if stride == 4: + blocks = [ + nn.Conv2d(in_channel, channel // 2, 4, stride=2, padding=1), + nn.ReLU(inplace=True), + nn.Conv2d(channel // 2, channel, 4, stride=2, padding=1), + nn.ReLU(inplace=True), + nn.Conv2d(channel, channel, 3, padding=1), + ] + + elif stride == 2: + blocks = [ + nn.Conv2d(in_channel, channel // 2, 4, stride=2, padding=1), + nn.ReLU(inplace=True), + nn.Conv2d(channel // 2, channel, 3, padding=1), + ] + + for i in range(n_res_block): + blocks.append(ResBlock(channel, n_res_channel)) + + blocks.append(nn.ReLU(inplace=True)) + + self.blocks = nn.Sequential(*blocks) + + def forward(self, input): + return self.blocks(input) + + +class Decoder(nn.Module): + def __init__( + self, in_channel, out_channel, channel, n_res_block, n_res_channel, stride + ): + super().__init__() + + blocks = [nn.Conv2d(in_channel, channel, 3, padding=1)] + + for i in range(n_res_block): + blocks.append(ResBlock(channel, n_res_channel)) + + blocks.append(nn.ReLU(inplace=True)) + + if stride == 4: + blocks.extend( + [ + nn.ConvTranspose2d(channel, channel // 2, 4, stride=2, padding=1), + nn.ReLU(inplace=True), + nn.ConvTranspose2d( + channel // 2, out_channel, 4, stride=2, padding=1 + ), + ] + ) + + elif stride == 2: + blocks.append( + nn.ConvTranspose2d(channel, out_channel, 4, stride=2, padding=1) + ) + + self.blocks = nn.Sequential(*blocks) + + def forward(self, input): + return self.blocks(input) + + +class VQVAE(nn.Module): + def __init__( + self, + in_channel=3, + channel=128, + n_res_block=2, + n_res_channel=32, + codebook_dim=64, + codebook_size=512, + decay=0.99, + ): + super().__init__() + + self.enc_b = Encoder(in_channel, channel, n_res_block, n_res_channel, stride=4) + self.enc_t = Encoder(channel, channel, n_res_block, n_res_channel, stride=2) + self.quantize_conv_t = nn.Conv2d(channel, codebook_dim, 1) + self.quantize_t = Quantize(codebook_dim, codebook_size) + self.dec_t = Decoder( + codebook_dim, codebook_dim, channel, n_res_block, n_res_channel, stride=2 + ) + self.quantize_conv_b = nn.Conv2d(codebook_dim + channel, codebook_dim, 1) + self.quantize_b = Quantize(codebook_dim, codebook_size) + self.upsample_t = nn.ConvTranspose2d( + codebook_dim, codebook_dim, 4, stride=2, padding=1 + ) + self.dec = Decoder( + codebook_dim + codebook_dim, + in_channel, + channel, + n_res_block, + n_res_channel, + stride=4, + ) + + def forward(self, input): + quant_t, quant_b, diff, _, _ = self.encode(input) + dec = self.decode(quant_t, quant_b) + + return dec, diff + + def encode(self, input): + enc_b = checkpoint(self.enc_b, input) + enc_t = checkpoint(self.enc_t, enc_b) + + quant_t = self.quantize_conv_t(enc_t).permute(0, 2, 3, 1) + quant_t, diff_t, id_t = self.quantize_t(quant_t) + quant_t = quant_t.permute(0, 3, 1, 2) + diff_t = diff_t.unsqueeze(0) + + dec_t = checkpoint(self.dec_t, quant_t) + enc_b = torch.cat([dec_t, enc_b], 1) + + quant_b = checkpoint(self.quantize_conv_b, enc_b).permute(0, 2, 3, 1) + quant_b, diff_b, id_b = self.quantize_b(quant_b) + quant_b = quant_b.permute(0, 3, 1, 2) + diff_b = diff_b.unsqueeze(0) + + return quant_t, quant_b, diff_t + diff_b, id_t, id_b + + def decode(self, quant_t, quant_b): + upsample_t = self.upsample_t(quant_t) + quant = torch.cat([upsample_t, quant_b], 1) + dec = checkpoint(self.dec, quant) + + return dec + + def decode_code(self, code_t, code_b): + quant_t = self.quantize_t.embed_code(code_t) + quant_t = quant_t.permute(0, 3, 1, 2) + quant_b = self.quantize_b.embed_code(code_b) + quant_b = quant_b.permute(0, 3, 1, 2) + + dec = self.decode(quant_t, quant_b) + + return dec + + +@register_model +def register_vqvae(opt_net, opt): + kw = opt_get(opt_net, ['kwargs'], {}) + return VQVAE(**kw) diff --git a/codes/scripts/extract_square_images.py b/codes/scripts/extract_square_images.py index bf0777bf..fff68aee 100644 --- a/codes/scripts/extract_square_images.py +++ b/codes/scripts/extract_square_images.py @@ -14,14 +14,14 @@ def main(): split_img = False opt = {} opt['n_thread'] = 4 - opt['compression_level'] = 90 # JPEG compression quality rating. + opt['compression_level'] = 98 # JPEG compression quality rating. # CV_IMWRITE_PNG_COMPRESSION from 0 to 9. A higher value means a smaller size and longer # compression time. If read raw images during training, use 0 for faster IO speed. opt['dest'] = 'file' - opt['input_folder'] = ['F:\\4k6k\\datasets\\ns_images\\vixen\\vix_cropped'] - opt['save_folder'] = 'F:\\4k6k\\datasets\\ns_images\\video_512_cropped' - opt['imgsize'] = 512 + opt['input_folder'] = ['F:\\4k6k\\datasets\\ns_images\\imagesets\\imageset_1024_square_with_new'] + opt['save_folder'] = 'F:\\4k6k\\datasets\\ns_images\\imagesets\\imageset_256_full' + opt['imgsize'] = 256 #opt['bottom_crop'] = 120 save_folder = opt['save_folder'] diff --git a/codes/train.py b/codes/train.py index 2cef4362..e57e4152 100644 --- a/codes/train.py +++ b/codes/train.py @@ -295,7 +295,7 @@ class Trainer: if __name__ == '__main__': parser = argparse.ArgumentParser() - parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_pixpro_resnet.yml') + parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_imgset_stylesr.yml') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args()