189 lines
8.7 KiB
Python
189 lines
8.7 KiB
Python
import math
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
import torchvision
|
|
import numpy as np
|
|
from models.archs.srflow_orig.RRDBNet_arch import RRDBNet
|
|
from models.archs.srflow_orig.FlowUpsamplerNet import FlowUpsamplerNet
|
|
import models.archs.srflow_orig.thops as thops
|
|
import models.archs.srflow_orig.flow as flow
|
|
from utils.util import opt_get
|
|
|
|
|
|
class SRFlowNet(nn.Module):
|
|
def __init__(self, in_nc, out_nc, nf, nb, gc=32, scale=4, K=None, opt=None, step=None):
|
|
super(SRFlowNet, self).__init__()
|
|
|
|
self.opt = opt
|
|
self.quant = 255 if opt_get(opt, ['datasets', 'train', 'quant']) is \
|
|
None else opt_get(opt, ['datasets', 'train', 'quant'])
|
|
initial_stride = opt_get(opt, ['networks', 'generator', 'initial_stride'], 1)
|
|
self.RRDB = RRDBNet(in_nc, out_nc, nf=nf, nb=nb, gc=gc, scale=scale, opt=opt, initial_conv_stride=initial_stride)
|
|
if 'pretrain_rrdb' in opt['networks']['generator'].keys():
|
|
rrdb_state_dict = torch.load(opt['networks']['generator']['pretrain_rrdb'])
|
|
self.RRDB.load_state_dict(rrdb_state_dict, strict=True)
|
|
|
|
hidden_channels = opt_get(opt, ['networks', 'generator','flow', 'hidden_channels'])
|
|
hidden_channels = hidden_channels or 64
|
|
self.RRDB_training = opt_get(self.opt, ['networks', 'generator','train_RRDB'], default=False)
|
|
self.flow_scale = opt_get(self.opt, ['networks', 'generator', 'flow_scale'], default=opt['scale']) # <!-- hack to enable RRDB to do 2x scaling while retaining the flow architecture of 4x.
|
|
|
|
self.patch_sz = opt_get(self.opt, ['networks', 'generator', 'flow', 'patch_size'], 160)
|
|
self.flowUpsamplerNet = \
|
|
FlowUpsamplerNet((self.patch_sz, self.patch_sz, 3), hidden_channels, K,
|
|
flow_coupling=opt['networks']['generator']['flow']['coupling'], opt=opt)
|
|
self.force_act_norm_init_until = opt_get(self.opt, ['networks', 'generator', 'flow', 'act_norm_start_step'])
|
|
self.act_norm_always_init = False
|
|
self.i = 0
|
|
self.dbg_logp = 0
|
|
self.dbg_logdet = 0
|
|
|
|
def get_random_z(self, heat, seed=None, batch_size=1, lr_shape=None, device='cuda'):
|
|
if seed: torch.manual_seed(seed)
|
|
if opt_get(self.opt, ['networks', 'generator', 'flow', 'split', 'enable']):
|
|
C = self.flowUpsamplerNet.C
|
|
H = int(self.flow_scale * lr_shape[0] // (self.flowUpsamplerNet.scaleH * self.flow_scale / self.RRDB.scale))
|
|
W = int(self.flow_scale * lr_shape[1] // (self.flowUpsamplerNet.scaleW * self.flow_scale / self.RRDB.scale))
|
|
|
|
size = (batch_size, C, H, W)
|
|
if heat == 0:
|
|
z = torch.zeros(size)
|
|
else:
|
|
z = torch.normal(mean=0, std=heat, size=size)
|
|
else:
|
|
L = opt_get(self.opt, ['networks', 'generator', 'flow', 'L']) or 3
|
|
fac = 2 ** (L - 3)
|
|
z_size = int(self.lr_size // (2 ** (L - 3)))
|
|
z = torch.normal(mean=0, std=heat, size=(batch_size, 3 * 8 * 8 * fac * fac, z_size, z_size))
|
|
return z.to(device)
|
|
|
|
def update_for_step(self, step, experiments_path='.'):
|
|
if self.act_norm_always_init and step > self.force_act_norm_init_until:
|
|
set_act_norm_always_init = True
|
|
set_value = False
|
|
self.act_norm_always_init = False
|
|
elif not self.act_norm_always_init and step < self.force_act_norm_init_until:
|
|
set_act_norm_always_init = True
|
|
set_value = True
|
|
self.act_norm_always_init = True
|
|
else:
|
|
set_act_norm_always_init = False
|
|
if set_act_norm_always_init:
|
|
for m in self.modules():
|
|
from models.archs.srflow_orig.FlowActNorms import _ActNorm
|
|
if isinstance(m, _ActNorm):
|
|
m.force_initialization = set_value
|
|
|
|
def forward(self, gt=None, lr=None, z=None, eps_std=None, reverse=False, epses=None, reverse_with_grad=False,
|
|
lr_enc=None,
|
|
add_gt_noise=True, step=None, y_label=None):
|
|
if not reverse:
|
|
return self.normal_flow(gt, lr, epses=epses, lr_enc=lr_enc, add_gt_noise=add_gt_noise, step=step,
|
|
y_onehot=y_label)
|
|
else:
|
|
assert lr.shape[1] == 3
|
|
if z is None:
|
|
# Synthesize it. Accommodate mismatches in LR scale and flow_scale, which are normally handled by the RRDB subnet.
|
|
lr_shape = [d * self.opt['scale'] / self.flow_scale for d in lr.shape[2:]]
|
|
z = self.get_random_z(eps_std, batch_size=lr.shape[0], lr_shape=lr_shape, device=lr.device)
|
|
if reverse_with_grad:
|
|
return self.reverse_flow(lr, z, y_onehot=y_label, eps_std=eps_std, epses=epses, lr_enc=lr_enc,
|
|
add_gt_noise=add_gt_noise)
|
|
else:
|
|
with torch.no_grad():
|
|
return self.reverse_flow(lr, z, y_onehot=y_label, eps_std=eps_std, epses=epses, lr_enc=lr_enc,
|
|
add_gt_noise=add_gt_noise)
|
|
|
|
def normal_flow(self, gt, lr, y_onehot=None, epses=None, lr_enc=None, add_gt_noise=True, step=None):
|
|
if lr_enc is None:
|
|
if self.RRDB_training:
|
|
lr_enc = self.rrdbPreprocessing(lr)
|
|
else:
|
|
with torch.no_grad():
|
|
lr_enc = self.rrdbPreprocessing(lr)
|
|
|
|
logdet = torch.zeros_like(gt[:, 0, 0, 0])
|
|
pixels = thops.pixels(gt)
|
|
|
|
z = gt
|
|
|
|
if add_gt_noise:
|
|
# Setup
|
|
noiseQuant = opt_get(self.opt, ['networks', 'generator','flow', 'augmentation', 'noiseQuant'], True)
|
|
if noiseQuant:
|
|
z = z + ((torch.rand(z.shape, device=z.device) - 0.5) / self.quant)
|
|
logdet = logdet + float(-np.log(self.quant) * pixels)
|
|
|
|
# Encode
|
|
epses, logdet = self.flowUpsamplerNet(rrdbResults=lr_enc, gt=z, logdet=logdet, reverse=False, epses=[],
|
|
y_onehot=y_onehot)
|
|
|
|
objective = logdet.clone()
|
|
|
|
if isinstance(epses, (list, tuple)):
|
|
z = epses[-1]
|
|
else:
|
|
z = epses
|
|
|
|
logp = flow.GaussianDiag.logp(None, None, z)
|
|
objective = objective + logp
|
|
|
|
nll = (-objective) / float(np.log(2.) * pixels)
|
|
self.dbg_logp = -logp.mean().item() / float(np.log(2.) * pixels)
|
|
self.dbg_logdet = -logdet.mean().item() / float(np.log(2.) * pixels)
|
|
|
|
if isinstance(epses, list):
|
|
return epses, nll, logdet
|
|
return z, nll, logdet
|
|
|
|
def get_debug_values(self, s, n):
|
|
return {"logp": self.dbg_logp, "logdet": self.dbg_logdet}
|
|
|
|
def rrdbPreprocessing(self, lr):
|
|
rrdbResults = self.RRDB(lr, get_steps=True)
|
|
block_idxs = opt_get(self.opt, ['networks', 'generator', 'flow', 'stackRRDB', 'blocks']) or []
|
|
if len(block_idxs) > 0:
|
|
concat = torch.cat([rrdbResults["block_{}".format(idx)] for idx in block_idxs], dim=1)
|
|
|
|
if opt_get(self.opt, ['networks', 'generator','flow', 'stackRRDB', 'concat']) or False:
|
|
keys = ['last_lr_fea', 'fea_up1', 'fea_up2', 'fea_up4']
|
|
if 'fea_up0' in rrdbResults.keys():
|
|
keys.append('fea_up0')
|
|
if 'fea_up-1' in rrdbResults.keys():
|
|
keys.append('fea_up-1')
|
|
if self.flow_scale >= 8:
|
|
keys.append('fea_up8')
|
|
if self.flow_scale == 16:
|
|
keys.append('fea_up16')
|
|
for k in keys:
|
|
h = rrdbResults[k].shape[2]
|
|
w = rrdbResults[k].shape[3]
|
|
rrdbResults[k] = torch.cat([rrdbResults[k], F.interpolate(concat, (h, w))], dim=1)
|
|
return rrdbResults
|
|
|
|
def get_score(self, disc_loss_sigma, z):
|
|
score_real = 0.5 * (1 - 1 / (disc_loss_sigma ** 2)) * thops.sum(z ** 2, dim=[1, 2, 3]) - \
|
|
z.shape[1] * z.shape[2] * z.shape[3] * math.log(disc_loss_sigma)
|
|
return -score_real
|
|
|
|
def reverse_flow(self, lr, z, y_onehot, eps_std, epses=None, lr_enc=None, add_gt_noise=True):
|
|
logdet = torch.zeros_like(lr[:, 0, 0, 0])
|
|
pixels = thops.pixels(lr) * self.flow_scale ** 2
|
|
|
|
if add_gt_noise:
|
|
logdet = logdet - float(-np.log(self.quant) * pixels)
|
|
|
|
if lr_enc is None:
|
|
if self.RRDB_training:
|
|
lr_enc = self.rrdbPreprocessing(lr)
|
|
else:
|
|
with torch.no_grad():
|
|
lr_enc = self.rrdbPreprocessing(lr)
|
|
|
|
x, logdet = self.flowUpsamplerNet(rrdbResults=lr_enc, z=z, eps_std=eps_std, reverse=True, epses=epses,
|
|
logdet=logdet)
|
|
|
|
return x, logdet
|