from fastai.vision import *
One-time download, uncomment the next cells to get the data.
#path = Config().data_path()
#! wget https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/horse2zebra.zip -P {path}
#! unzip -q -n {path}/horse2zebra.zip -d {path}
#! rm {path}/horse2zebra.zip
path = Config().data_path()/'horse2zebra'
path.ls()
See this tutorial for a detailed walkthrough of how/why this custom ItemList
was created.
class ImageTuple(ItemBase):
def __init__(self, img1, img2):
self.img1,self.img2 = img1,img2
self.obj,self.data = (img1,img2),[-1+2*img1.data,-1+2*img2.data]
def apply_tfms(self, tfms, **kwargs):
self.img1 = self.img1.apply_tfms(tfms, **kwargs)
self.img2 = self.img2.apply_tfms(tfms, **kwargs)
return self
def to_one(self): return Image(0.5+torch.cat(self.data,2)/2)
class TargetTupleList(ItemList):
def reconstruct(self, t:Tensor):
if len(t.size()) == 0: return t
return ImageTuple(Image(t[0]/2+0.5),Image(t[1]/2+0.5))
class ImageTupleList(ImageList):
_label_cls=TargetTupleList
def __init__(self, items, itemsB=None, **kwargs):
self.itemsB = itemsB
super().__init__(items, **kwargs)
def new(self, items, **kwargs):
return super().new(items, itemsB=self.itemsB, **kwargs)
def get(self, i):
img1 = super().get(i)
fn = self.itemsB[random.randint(0, len(self.itemsB)-1)]
return ImageTuple(img1, open_image(fn))
def reconstruct(self, t:Tensor):
return ImageTuple(Image(t[0]/2+0.5),Image(t[1]/2+0.5))
@classmethod
def from_folders(cls, path, folderA, folderB, **kwargs):
itemsB = ImageList.from_folder(path/folderB).items
res = super().from_folder(path/folderA, itemsB=itemsB, **kwargs)
res.path = path
return res
def show_xys(self, xs, ys, figsize:Tuple[int,int]=(12,6), **kwargs):
"Show the `xs` and `ys` on a figure of `figsize`. `kwargs` are passed to the show method."
rows = int(math.sqrt(len(xs)))
fig, axs = plt.subplots(rows,rows,figsize=figsize)
for i, ax in enumerate(axs.flatten() if rows > 1 else [axs]):
xs[i].to_one().show(ax=ax, **kwargs)
plt.tight_layout()
def show_xyzs(self, xs, ys, zs, figsize:Tuple[int,int]=None, **kwargs):
"""Show `xs` (inputs), `ys` (targets) and `zs` (predictions) on a figure of `figsize`.
`kwargs` are passed to the show method."""
figsize = ifnone(figsize, (12,3*len(xs)))
fig,axs = plt.subplots(len(xs), 2, figsize=figsize)
fig.suptitle('Ground truth / Predictions', weight='bold', size=14)
for i,(x,z) in enumerate(zip(xs,zs)):
x.to_one().show(ax=axs[i,0], **kwargs)
z.to_one().show(ax=axs[i,1], **kwargs)
data = (ImageTupleList.from_folders(path, 'trainA', 'trainB')
.split_none()
.label_empty()
.transform(get_transforms(), size=128)
.databunch(bs=4))
data.show_batch(rows=2)
We use the models that were introduced in the cycleGAN paper.
def convT_norm_relu(ch_in:int, ch_out:int, norm_layer:nn.Module, ks:int=3, stride:int=2, bias:bool=True):
return [nn.ConvTranspose2d(ch_in, ch_out, kernel_size=ks, stride=stride, padding=1, output_padding=1, bias=bias),
norm_layer(ch_out), nn.ReLU(True)]
def pad_conv_norm_relu(ch_in:int, ch_out:int, pad_mode:str, norm_layer:nn.Module, ks:int=3, bias:bool=True,
pad=1, stride:int=1, activ:bool=True, init:Callable=nn.init.kaiming_normal_)->List[nn.Module]:
layers = []
if pad_mode == 'reflection': layers.append(nn.ReflectionPad2d(pad))
elif pad_mode == 'border': layers.append(nn.ReplicationPad2d(pad))
p = pad if pad_mode == 'zeros' else 0
conv = nn.Conv2d(ch_in, ch_out, kernel_size=ks, padding=p, stride=stride, bias=bias)
if init:
init(conv.weight)
if hasattr(conv, 'bias') and hasattr(conv.bias, 'data'): conv.bias.data.fill_(0.)
layers += [conv, norm_layer(ch_out)]
if activ: layers.append(nn.ReLU(inplace=True))
return layers
class ResnetBlock(nn.Module):
def __init__(self, dim:int, pad_mode:str='reflection', norm_layer:nn.Module=None, dropout:float=0., bias:bool=True):
super().__init__()
assert pad_mode in ['zeros', 'reflection', 'border'], f'padding {pad_mode} not implemented.'
norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)
layers = pad_conv_norm_relu(dim, dim, pad_mode, norm_layer, bias=bias)
if dropout != 0: layers.append(nn.Dropout(dropout))
layers += pad_conv_norm_relu(dim, dim, pad_mode, norm_layer, bias=bias, activ=False)
self.conv_block = nn.Sequential(*layers)
def forward(self, x): return x + self.conv_block(x)
def resnet_generator(ch_in:int, ch_out:int, n_ftrs:int=64, norm_layer:nn.Module=None,
dropout:float=0., n_blocks:int=6, pad_mode:str='reflection')->nn.Module:
norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)
bias = (norm_layer == nn.InstanceNorm2d)
layers = pad_conv_norm_relu(ch_in, n_ftrs, 'reflection', norm_layer, pad=3, ks=7, bias=bias)
for i in range(2):
layers += pad_conv_norm_relu(n_ftrs, n_ftrs *2, 'zeros', norm_layer, stride=2, bias=bias)
n_ftrs *= 2
layers += [ResnetBlock(n_ftrs, pad_mode, norm_layer, dropout, bias) for _ in range(n_blocks)]
for i in range(2):
layers += convT_norm_relu(n_ftrs, n_ftrs//2, norm_layer, bias=bias)
n_ftrs //= 2
layers += [nn.ReflectionPad2d(3), nn.Conv2d(n_ftrs, ch_out, kernel_size=7, padding=0), nn.Tanh()]
return nn.Sequential(*layers)
resnet_generator(3, 3)
def conv_norm_lr(ch_in:int, ch_out:int, norm_layer:nn.Module=None, ks:int=3, bias:bool=True, pad:int=1, stride:int=1,
activ:bool=True, slope:float=0.2, init:Callable=nn.init.kaiming_normal_)->List[nn.Module]:
conv = nn.Conv2d(ch_in, ch_out, kernel_size=ks, padding=pad, stride=stride, bias=bias)
if init:
init(conv.weight)
if hasattr(conv, 'bias') and hasattr(conv.bias, 'data'): conv.bias.data.fill_(0.)
layers = [conv]
if norm_layer is not None: layers.append(norm_layer(ch_out))
if activ: layers.append(nn.LeakyReLU(slope, inplace=True))
return layers
def discriminator(ch_in:int, n_ftrs:int=64, n_layers:int=3, norm_layer:nn.Module=None, sigmoid:bool=False)->nn.Module:
norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)
bias = (norm_layer == nn.InstanceNorm2d)
layers = conv_norm_lr(ch_in, n_ftrs, ks=4, stride=2, pad=1)
for i in range(n_layers-1):
new_ftrs = 2*n_ftrs if i <= 3 else n_ftrs
layers += conv_norm_lr(n_ftrs, new_ftrs, norm_layer, ks=4, stride=2, pad=1, bias=bias)
n_ftrs = new_ftrs
new_ftrs = 2*n_ftrs if n_layers <=3 else n_ftrs
layers += conv_norm_lr(n_ftrs, new_ftrs, norm_layer, ks=4, stride=1, pad=1, bias=bias)
layers.append(nn.Conv2d(new_ftrs, 1, kernel_size=4, stride=1, padding=1))
if sigmoid: layers.append(nn.Sigmoid())
return nn.Sequential(*layers)
discriminator(3)
We group two discriminators and two generators in a single model, then a Callback
will take care of training them properly.
class CycleGAN(nn.Module):
def __init__(self, ch_in:int, ch_out:int, n_features:int=64, disc_layers:int=3, gen_blocks:int=6, lsgan:bool=True,
drop:float=0., norm_layer:nn.Module=None):
super().__init__()
self.D_A = discriminator(ch_in, n_features, disc_layers, norm_layer, sigmoid=not lsgan)
self.D_B = discriminator(ch_in, n_features, disc_layers, norm_layer, sigmoid=not lsgan)
self.G_A = resnet_generator(ch_in, ch_out, n_features, norm_layer, drop, gen_blocks)
self.G_B = resnet_generator(ch_in, ch_out, n_features, norm_layer, drop, gen_blocks)
#G_A: takes real input B and generates fake input A
#G_B: takes real input A and generates fake input B
#D_A: trained to make the difference between real input A and fake input A
#D_B: trained to make the difference between real input B and fake input B
def forward(self, real_A, real_B):
fake_A, fake_B = self.G_A(real_B), self.G_B(real_A)
if not self.training: return torch.cat([fake_A[:,None],fake_B[:,None]], 1)
idt_A, idt_B = self.G_A(real_A), self.G_B(real_B) #Needed for the identity loss during training.
return [fake_A, fake_B, idt_A, idt_B]
AdaptiveLoss
is a wrapper around a PyTorch loss function to compare an output of any size with a single number (0. or 1.). It will generate a target with the same shape as the output. A discriminator returns a feature map, and we want it to predict zeros (or ones) for each feature.
class AdaptiveLoss(nn.Module):
def __init__(self, crit):
super().__init__()
self.crit = crit
def forward(self, output, target:bool, **kwargs):
targ = output.new_ones(*output.size()) if target else output.new_zeros(*output.size())
return self.crit(output, targ, **kwargs)
The main loss used to train the generators. It has three parts:
class CycleGanLoss(nn.Module):
def __init__(self, cgan:nn.Module, lambda_A:float=10., lambda_B:float=10, lambda_idt:float=0.5, lsgan:bool=True):
super().__init__()
self.cgan,self.l_A,self.l_B,self.l_idt = cgan,lambda_A,lambda_B,lambda_idt
self.crit = AdaptiveLoss(F.mse_loss if lsgan else F.binary_cross_entropy)
def set_input(self, input):
self.real_A,self.real_B = input
def forward(self, output, target):
fake_A, fake_B, idt_A, idt_B = output
#Generators should return identity on the datasets they try to convert to
self.id_loss = self.l_idt * (self.l_A * F.l1_loss(idt_A, self.real_A) + self.l_B * F.l1_loss(idt_B, self.real_B))
#Generators are trained to trick the discriminators so the following should be ones
self.gen_loss = self.crit(self.cgan.D_A(fake_A), True) + self.crit(self.cgan.D_B(fake_B), True)
#Cycle loss
self.cyc_loss = self.l_A * F.l1_loss(self.cgan.G_A(fake_B), self.real_A)
self.cyc_loss += self.l_B * F.l1_loss(self.cgan.G_B(fake_A), self.real_B)
return self.id_loss+self.gen_loss+self.cyc_loss
The main callback to train a cycle GAN. The training loop will train the generators (so learn.opt
is given those parameters) while the critics are trained by the callback during on_batch_end
.
class CycleGANTrainer(LearnerCallback):
_order = -20 #Need to run before the Recorder
def _set_trainable(self, D_A=False, D_B=False):
gen = (not D_A) and (not D_B)
requires_grad(self.learn.model.G_A, gen)
requires_grad(self.learn.model.G_B, gen)
requires_grad(self.learn.model.D_A, D_A)
requires_grad(self.learn.model.D_B, D_B)
if not gen:
self.opt_D_A.lr, self.opt_D_A.mom = self.learn.opt.lr, self.learn.opt.mom
self.opt_D_A.wd, self.opt_D_A.beta = self.learn.opt.wd, self.learn.opt.beta
self.opt_D_B.lr, self.opt_D_B.mom = self.learn.opt.lr, self.learn.opt.mom
self.opt_D_B.wd, self.opt_D_B.beta = self.learn.opt.wd, self.learn.opt.beta
def on_train_begin(self, **kwargs):
self.G_A,self.G_B = self.learn.model.G_A,self.learn.model.G_B
self.D_A,self.D_B = self.learn.model.D_A,self.learn.model.D_B
self.crit = self.learn.loss_func.crit
if not getattr(self,'opt_G',None):
self.opt_G = self.learn.opt.new([nn.Sequential(*flatten_model(self.G_A), *flatten_model(self.G_B))])
else:
self.opt_G.lr,self.opt_G.wd = self.opt.lr,self.opt.wd
self.opt_G.mom,self.opt_G.beta = self.opt.mom,self.opt.beta
if not getattr(self,'opt_D_A',None):
self.opt_D_A = self.learn.opt.new([nn.Sequential(*flatten_model(self.D_A))])
if not getattr(self,'opt_D_B',None):
self.opt_D_B = self.learn.opt.new([nn.Sequential(*flatten_model(self.D_B))])
self.learn.opt.opt = self.opt_G.opt
self._set_trainable()
self.id_smter,self.gen_smter,self.cyc_smter = SmoothenValue(0.98),SmoothenValue(0.98),SmoothenValue(0.98)
self.da_smter,self.db_smter = SmoothenValue(0.98),SmoothenValue(0.98)
self.recorder.add_metric_names(['id_loss', 'gen_loss', 'cyc_loss', 'D_A_loss', 'D_B_loss'])
def on_batch_begin(self, last_input, **kwargs):
self.learn.loss_func.set_input(last_input)
def on_backward_begin(self, **kwargs):
self.id_smter.add_value(self.loss_func.id_loss.detach().cpu())
self.gen_smter.add_value(self.loss_func.gen_loss.detach().cpu())
self.cyc_smter.add_value(self.loss_func.cyc_loss.detach().cpu())
def on_batch_end(self, last_input, last_output, **kwargs):
self.G_A.zero_grad(); self.G_B.zero_grad()
fake_A, fake_B = last_output[0].detach(), last_output[1].detach()
real_A, real_B = last_input
self._set_trainable(D_A=True)
self.D_A.zero_grad()
loss_D_A = 0.5 * (self.crit(self.D_A(real_A), True) + self.crit(self.D_A(fake_A), False))
self.da_smter.add_value(loss_D_A.detach().cpu())
loss_D_A.backward()
self.opt_D_A.step()
self._set_trainable(D_B=True)
self.D_B.zero_grad()
loss_D_B = 0.5 * (self.crit(self.D_B(real_B), True) + self.crit(self.D_B(fake_B), False))
self.db_smter.add_value(loss_D_B.detach().cpu())
loss_D_B.backward()
self.opt_D_B.step()
self._set_trainable()
def on_epoch_end(self, last_metrics, **kwargs):
return add_metrics(last_metrics, [s.smooth for s in [self.id_smter,self.gen_smter,self.cyc_smter,
self.da_smter,self.db_smter]])
cycle_gan = CycleGAN(3,3, gen_blocks=9)
learn = Learner(data, cycle_gan, loss_func=CycleGanLoss(cycle_gan), opt_func=partial(optim.Adam, betas=(0.5,0.99)),
callback_fns=[CycleGANTrainer])
learn.lr_find()
learn.recorder.plot()
learn.fit(100, 1e-4)
learn.save('100fit')
learn = learn.load('100fit')
Let's look at some results using Learner.show_results
.
learn.show_results(ds_type=DatasetType.Train, rows=2)
learn.show_results(ds_type=DatasetType.Train, rows=2)
Now let's go through all the images of the training set and find the ones that are the best converted (according to our critics) or the worst converted.
len(learn.data.train_ds.items),len(learn.data.train_ds.itemsB)
def get_batch(filenames, tfms, **kwargs):
samples = [open_image(fn) for fn in filenames]
for s in samples: s = s.apply_tfms(tfms, **kwargs)
batch = torch.stack([s.data for s in samples], 0).cuda()
return 2. * (batch - 0.5)
fnames = learn.data.train_ds.items[:8]
x = get_batch(fnames, get_transforms()[1], size=128)
learn.model.eval()
tfms = get_transforms()[1]
bs = 16
def get_losses(fnames, gen, crit, bs=16):
losses_in,losses_out = [],[]
with torch.no_grad():
for i in progress_bar(range(0, len(fnames), bs)):
xb = get_batch(fnames[i:i+bs], tfms, size=128)
fakes = gen(xb)
preds_in,preds_out = crit(xb),crit(fakes)
loss_in = learn.loss_func.crit(preds_in, True,reduction='none')
loss_out = learn.loss_func.crit(preds_out,True,reduction='none')
losses_in.append(loss_in.view(loss_in.size(0),-1).mean(1))
losses_out.append(loss_out.view(loss_out.size(0),-1).mean(1))
return torch.cat(losses_in),torch.cat(losses_out)
losses_A = get_losses(data.train_ds.x.items, learn.model.G_B, learn.model.D_B)
losses_B = get_losses(data.train_ds.x.itemsB, learn.model.G_A, learn.model.D_A)
def show_best(fnames, losses, gen, n=8):
sort_idx = losses.argsort().cpu()
_,axs = plt.subplots(n//2, 4, figsize=(12,2*n))
xb = get_batch(fnames[sort_idx][:n], tfms, size=128)
with torch.no_grad():
fakes = gen(xb)
xb,fakes = (1+xb.cpu())/2,(1+fakes.cpu())/2
for i in range(n):
axs.flatten()[2*i].imshow(xb[i].permute(1,2,0))
axs.flatten()[2*i].axis('off')
axs.flatten()[2*i+1].imshow(fakes[i].permute(1,2,0))
axs.flatten()[2*i+1].set_title(losses[sort_idx][i].item())
axs.flatten()[2*i+1].axis('off')
show_best(data.train_ds.x.items, losses_A[1], learn.model.G_B)
show_best(data.train_ds.x.itemsB, losses_B[1], learn.model.G_A)