Skip to content

Instantly share code, notes, and snippets.

@bearpelican

bearpelican/00_cyclegan_fastai-v1_w_metrics.ipynb

Created May 21, 2020 19:13
Show Gist options
  • Save bearpelican/d2833b3e65134e66b59904f0aef11666 to your computer and use it in GitHub Desktop.
Save bearpelican/d2833b3e65134e66b59904f0aef11666 to your computer and use it in GitHub Desktop.
Download ZIP
Display the source blob
Display the rendered blob
Raw
00_cyclegan_fastai-v1_w_metrics.ipynb
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"from fastai.vision import *"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Data"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"One-time download, uncomment the next cells to get the data."
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"#path = Config().data_path()"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"#! wget https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/horse2zebra.zip -P {path}\n",
"#! unzip -q -n {path}/horse2zebra.zip -d {path}\n",
"#! rm {path}/horse2zebra.zip"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"data_path = Path('../../../../mnt/wamri/WAMRI-LevensonLab/datasets/')\n",
"muse2he_path = data_path/'muse2he_urothelial_carcinoma'\n",
"muse2he_path.ls()\n",
"\n",
"muse_path = muse2he_path/'trainA'\n",
"he_path = muse2he_path/'trainB'\n",
"\n",
"torch.cuda.set_device(0) #set GPU id"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"# path = Config().data_path()/'horse2zebra'\n",
"# path.ls()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"See [this tutorial](https://docs.fast.ai/tutorial.itemlist.html) for a detailed walkthrough of how/why this custom `ItemList` was created."
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [],
"source": [
"class ImageTuple(ItemBase):\n",
" def __init__(self, img1, img2):\n",
" self.img1,self.img2 = img1,img2\n",
" self.obj,self.data = (img1,img2),[-1+2*img1.data,-1+2*img2.data]\n",
" \n",
" def apply_tfms(self, tfms, **kwargs):\n",
" img1 = self.img1.apply_tfms(tfms, **kwargs)\n",
" img2 = self.img2.apply_tfms(tfms, **kwargs)\n",
" return ImageTuple(img1, img2)\n",
" \n",
" def to_one(self): return Image(0.5+torch.cat(self.data,2)/2)"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"class TargetTupleList(ItemList):\n",
" def reconstruct(self, t:Tensor): \n",
" if len(t.size()) == 0: return t\n",
" return ImageTuple(Image(t[0]/2+0.5),Image(t[1]/2+0.5))"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"class ImageTupleList(ImageList):\n",
" _label_cls=TargetTupleList\n",
" def __init__(self, items, itemsB=None, **kwargs):\n",
" self.itemsB = itemsB\n",
" super().__init__(items, **kwargs)\n",
" \n",
" def new(self, items, **kwargs):\n",
" return super().new(items, itemsB=self.itemsB, **kwargs)\n",
" \n",
" def get(self, i):\n",
" img1 = super().get(i)\n",
" fn = self.itemsB[random.randint(0, len(self.itemsB)-1)]\n",
" return ImageTuple(img1, open_image(fn))\n",
" \n",
" def reconstruct(self, t:Tensor): \n",
" return ImageTuple(Image(t[0]/2+0.5),Image(t[1]/2+0.5))\n",
" \n",
" @classmethod\n",
" def from_folders(cls, path, folderA, folderB, **kwargs):\n",
" itemsB = ImageList.from_folder(path/folderB).items\n",
" res = super().from_folder(path/folderA, itemsB=itemsB, **kwargs)\n",
" res.path = path\n",
" return res\n",
" \n",
" def show_xys(self, xs, ys, figsize:Tuple[int,int]=(12,6), **kwargs):\n",
" \"Show the `xs` and `ys` on a figure of `figsize`. `kwargs` are passed to the show method.\"\n",
" rows = int(math.sqrt(len(xs)))\n",
" fig, axs = plt.subplots(rows,rows,figsize=figsize)\n",
" for i, ax in enumerate(axs.flatten() if rows > 1 else [axs]):\n",
" xs[i].to_one().show(ax=ax, **kwargs)\n",
" plt.tight_layout()\n",
"\n",
" def show_xyzs(self, xs, ys, zs, figsize:Tuple[int,int]=None, **kwargs):\n",
" \"\"\"Show `xs` (inputs), `ys` (targets) and `zs` (predictions) on a figure of `figsize`.\n",
" `kwargs` are passed to the show method.\"\"\"\n",
" figsize = ifnone(figsize, (12,3*len(xs)))\n",
" fig,axs = plt.subplots(len(xs), 2, figsize=figsize)\n",
" fig.suptitle('Ground truth / Predictions', weight='bold', size=14)\n",
" for i,(x,z) in enumerate(zip(xs,zs)):\n",
" x.to_one().show(ax=axs[i,0], **kwargs)\n",
" z.to_one().show(ax=axs[i,1], **kwargs)"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
"# src = ImageTupleList.from_folders(muse2he_path, 'trainA', 'trainB').split_none().label_empty()"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [],
"source": [
"data = (ImageTupleList.from_folders(muse2he_path, 'trainA', 'trainB')\n",
" .split_none()\n",
" .label_empty()\n",
" .transform(2*[[crop(size=256,row_pct=0,col_pct=0),flip_lr(p=0.5)]],size=512,resize_method=ResizeMethod.SQUISH)\n",
" .databunch(bs=4,num_workers=2))\n",
"data.valid_dl = data.train_dl # a hack for proper evaluation of loss and metrics at end of training"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"data.show_batch(rows=2)"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(torch.Size([4, 3, 256, 256]), torch.Size([4, 3, 256, 256]))"
]
},
"execution_count": 12,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"xb, yb = data.one_batch()\n",
"xb[0].shape, xb[1].shape"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"train_image = next(iter(data.train_dl))[0][0][0]\n",
"plt.imshow(((train_image.permute(1,2,0)+1)/2*255).cpu().to(torch.int))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Models"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"We use the models that were introduced in the [cycleGAN paper](https://arxiv.org/abs/1703.10593)."
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {},
"outputs": [],
"source": [
"def convT_norm_relu(ch_in:int, ch_out:int, norm_layer:nn.Module, ks:int=3, stride:int=2, bias:bool=True):\n",
" return [nn.ConvTranspose2d(ch_in, ch_out, kernel_size=ks, stride=stride, padding=1, output_padding=1, bias=bias),\n",
" norm_layer(ch_out), nn.ReLU(True)]"
]
},
{
"cell_type": "code",
"execution_count": 15,
"metadata": {},
"outputs": [],
"source": [
"def pad_conv_norm_relu(ch_in:int, ch_out:int, pad_mode:str, norm_layer:nn.Module, ks:int=3, bias:bool=True, \n",
" pad=1, stride:int=1, activ:bool=True, init:Callable=nn.init.kaiming_normal_, init_gain:int=0.02)->List[nn.Module]:\n",
" layers = []\n",
" if pad_mode == 'reflection': layers.append(nn.ReflectionPad2d(pad))\n",
" elif pad_mode == 'border': layers.append(nn.ReplicationPad2d(pad))\n",
" p = pad if pad_mode == 'zeros' else 0\n",
" conv = nn.Conv2d(ch_in, ch_out, kernel_size=ks, padding=p, stride=stride, bias=bias)\n",
" if init:\n",
" if init == nn.init.normal_:\n",
" init(conv.weight, 0.0, init_gain)\n",
" else:\n",
" init(conv.weight)\n",
" if hasattr(conv, 'bias') and hasattr(conv.bias, 'data'): conv.bias.data.fill_(0.)\n",
" layers += [conv, norm_layer(ch_out)]\n",
" if activ: layers.append(nn.ReLU(inplace=True))\n",
" return layers"
]
},
{
"cell_type": "code",
"execution_count": 16,
"metadata": {},
"outputs": [],
"source": [
"class ResnetBlock(nn.Module):\n",
" def __init__(self, dim:int, pad_mode:str='reflection', norm_layer:nn.Module=None, dropout:float=0., bias:bool=True):\n",
" super().__init__()\n",
" assert pad_mode in ['zeros', 'reflection', 'border'], f'padding {pad_mode} not implemented.'\n",
" norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)\n",
" layers = pad_conv_norm_relu(dim, dim, pad_mode, norm_layer, bias=bias)\n",
" if dropout != 0: layers.append(nn.Dropout(dropout))\n",
" layers += pad_conv_norm_relu(dim, dim, pad_mode, norm_layer, bias=bias, activ=False)\n",
" self.conv_block = nn.Sequential(*layers)\n",
"\n",
" def forward(self, x): return x + self.conv_block(x)"
]
},
{
"cell_type": "code",
"execution_count": 17,
"metadata": {},
"outputs": [],
"source": [
"def resnet_generator(ch_in:int, ch_out:int, n_ftrs:int=64, norm_layer:nn.Module=None, \n",
" dropout:float=0., n_blocks:int=9, pad_mode:str='reflection')->nn.Module:\n",
" norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)\n",
" bias = (norm_layer == nn.InstanceNorm2d)\n",
" layers = pad_conv_norm_relu(ch_in, n_ftrs, 'reflection', norm_layer, pad=3, ks=7, bias=bias)\n",
" for i in range(2):\n",
" layers += pad_conv_norm_relu(n_ftrs, n_ftrs *2, 'zeros', norm_layer, stride=2, bias=bias)\n",
" n_ftrs *= 2\n",
" layers += [ResnetBlock(n_ftrs, pad_mode, norm_layer, dropout, bias) for _ in range(n_blocks)]\n",
" for i in range(2):\n",
" layers += convT_norm_relu(n_ftrs, n_ftrs//2, norm_layer, bias=bias)\n",
" n_ftrs //= 2\n",
" layers += [nn.ReflectionPad2d(3), nn.Conv2d(n_ftrs, ch_out, kernel_size=7, padding=0), nn.Tanh()]\n",
" return nn.Sequential(*layers)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"resnet_generator(3, 3)"
]
},
{
"cell_type": "code",
"execution_count": 19,
"metadata": {},
"outputs": [],
"source": [
"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, \n",
" activ:bool=True, slope:float=0.2, init:Callable=nn.init.normal_, init_gain:int=0.02)->List[nn.Module]:\n",
" conv = nn.Conv2d(ch_in, ch_out, kernel_size=ks, padding=pad, stride=stride, bias=bias)\n",
" if init:\n",
" if init == nn.init.normal_:\n",
" init(conv.weight, 0.0, init_gain)\n",
" else:\n",
" init(conv.weight)\n",
" if hasattr(conv, 'bias') and hasattr(conv.bias, 'data'): conv.bias.data.fill_(0.)\n",
" layers = [conv]\n",
" if norm_layer is not None: layers.append(norm_layer(ch_out))\n",
" if activ: layers.append(nn.LeakyReLU(slope, inplace=True))\n",
" return layers"
]
},
{
"cell_type": "code",
"execution_count": 20,
"metadata": {},
"outputs": [],
"source": [
"def discriminator(ch_in:int, n_ftrs:int=64, n_layers:int=3, norm_layer:nn.Module=None, sigmoid:bool=False)->nn.Module:\n",
" norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)\n",
" bias = (norm_layer == nn.InstanceNorm2d)\n",
" layers = conv_norm_lr(ch_in, n_ftrs, ks=4, stride=2, pad=1)\n",
" for i in range(n_layers-1):\n",
" new_ftrs = 2*n_ftrs if i <= 3 else n_ftrs\n",
" layers += conv_norm_lr(n_ftrs, new_ftrs, norm_layer, ks=4, stride=2, pad=1, bias=bias)\n",
" n_ftrs = new_ftrs\n",
" new_ftrs = 2*n_ftrs if n_layers <=3 else n_ftrs\n",
" layers += conv_norm_lr(n_ftrs, new_ftrs, norm_layer, ks=4, stride=1, pad=1, bias=bias)\n",
" layers.append(nn.Conv2d(new_ftrs, 1, kernel_size=4, stride=1, padding=1))\n",
" if sigmoid: layers.append(nn.Sigmoid())\n",
" return nn.Sequential(*layers)"
]
},
{
"cell_type": "code",
"execution_count": 21,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"Sequential(\n",
" (0): Conv2d(3, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))\n",
" (1): LeakyReLU(negative_slope=0.2, inplace=True)\n",
" (2): Conv2d(64, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))\n",
" (3): InstanceNorm2d(128, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)\n",
" (4): LeakyReLU(negative_slope=0.2, inplace=True)\n",
" (5): Conv2d(128, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1))\n",
" (6): InstanceNorm2d(256, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)\n",
" (7): LeakyReLU(negative_slope=0.2, inplace=True)\n",
" (8): Conv2d(256, 512, kernel_size=(4, 4), stride=(1, 1), padding=(1, 1))\n",
" (9): InstanceNorm2d(512, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)\n",
" (10): LeakyReLU(negative_slope=0.2, inplace=True)\n",
" (11): Conv2d(512, 1, kernel_size=(4, 4), stride=(1, 1), padding=(1, 1))\n",
")"
]
},
"execution_count": 21,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"discriminator(3)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"We group two discriminators and two generators in a single model, then a `Callback` will take care of training them properly."
]
},
{
"cell_type": "code",
"execution_count": 22,
"metadata": {},
"outputs": [],
"source": [
"class CycleGAN(nn.Module):\n",
" \n",
" def __init__(self, ch_in:int, ch_out:int, n_features:int=64, disc_layers:int=3, gen_blocks:int=6, lsgan:bool=True, \n",
" drop:float=0., norm_layer:nn.Module=None):\n",
" super().__init__()\n",
" self.D_A = discriminator(ch_in, n_features, disc_layers, norm_layer, sigmoid=not lsgan)\n",
" self.D_B = discriminator(ch_in, n_features, disc_layers, norm_layer, sigmoid=not lsgan)\n",
" self.G_A = resnet_generator(ch_in, ch_out, n_features, norm_layer, drop, gen_blocks)\n",
" self.G_B = resnet_generator(ch_in, ch_out, n_features, norm_layer, drop, gen_blocks)\n",
" #G_A: takes real input B and generates fake input A\n",
" #G_B: takes real input A and generates fake input B\n",
" #D_A: trained to make the difference between real input A and fake input A\n",
" #D_B: trained to make the difference between real input B and fake input B\n",
" \n",
" def forward(self, real_A, real_B):\n",
" fake_A, fake_B = self.G_A(real_B), self.G_B(real_A)\n",
" idt_A, idt_B = self.G_A(real_A), self.G_B(real_B) #Needed for the identity loss during training.\n",
" if not self.training: return torch.cat([fake_A[:,None], fake_B[:,None], idt_A[:,None], idt_B[:,None]],1)\n",
" return [fake_A, fake_B, idt_A, idt_B]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"`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."
]
},
{
"cell_type": "code",
"execution_count": 23,
"metadata": {},
"outputs": [],
"source": [
"class AdaptiveLoss(nn.Module):\n",
" def __init__(self, crit):\n",
" super().__init__()\n",
" self.crit = crit\n",
" \n",
" def forward(self, output, target:bool, **kwargs):\n",
" targ = output.new_ones(*output.size()) if target else output.new_zeros(*output.size())\n",
" return self.crit(output, targ, **kwargs)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The main loss used to train the generators. It has three parts:\n",
"- the classic GAN loss: they must make the critics believe their images are real\n",
"- identity loss: if they are given an image from the set they are trying to imitate, they should return the same thing\n",
"- cycle loss: if an image from A goes through the generator that imitates B then through the generator that imitates A, it should be the same as the initial image. Same for B and switching the generators"
]
},
{
"cell_type": "code",
"execution_count": 24,
"metadata": {},
"outputs": [],
"source": [
"class CycleGanLoss(nn.Module):\n",
" \n",
" def __init__(self, cgan:nn.Module, lambda_A:float=10., lambda_B:float=10, lambda_idt:float=0.5, lsgan:bool=True):\n",
" super().__init__()\n",
" self.cgan,self.l_A,self.l_B,self.l_idt = cgan,lambda_A,lambda_B,lambda_idt\n",
" self.crit = AdaptiveLoss(F.mse_loss if lsgan else F.binary_cross_entropy)\n",
" \n",
" def set_training(self,training):\n",
" self.training = training\n",
"\n",
" def set_input(self, input):\n",
" self.real_A,self.real_B = input\n",
"\n",
" def forward(self, output, target):\n",
" if self.training:\n",
" fake_A, fake_B, idt_A, idt_B = output\n",
" else:\n",
" fake_A, fake_B, idt_A, idt_B = output[:,0,:,:,:], output[:,1,:,:,:], output[:,2,:,:,:], output[:,3,:,:,:]\n",
"\n",
" #Generators should return identity on the datasets they try to convert to\n",
" 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))\n",
" #Generators are trained to trick the discriminators so the following should be ones\n",
" self.gen_loss = self.crit(self.cgan.D_A(fake_A), True) + self.crit(self.cgan.D_B(fake_B), True)\n",
" #Cycle loss\n",
" self.cyc_loss = self.l_A * F.l1_loss(self.cgan.G_A(fake_B), self.real_A)\n",
" self.cyc_loss += self.l_B * F.l1_loss(self.cgan.G_B(fake_A), self.real_B)\n",
" return self.id_loss+self.gen_loss+self.cyc_loss"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"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`."
]
},
{
"cell_type": "code",
"execution_count": 25,
"metadata": {},
"outputs": [],
"source": [
"class CycleGANTrainer(LearnerCallback):\n",
" _order = -20 #Need to run before the Recorder\n",
" \n",
" def _set_trainable(self, D_A=False, D_B=False):\n",
" gen = (not D_A) and (not D_B)\n",
" requires_grad(self.learn.model.G_A, gen)\n",
" requires_grad(self.learn.model.G_B, gen)\n",
" requires_grad(self.learn.model.D_A, D_A)\n",
" requires_grad(self.learn.model.D_B, D_B)\n",
" if not gen:\n",
" self.opt_D_A.lr, self.opt_D_A.mom = self.learn.opt.lr, self.learn.opt.mom\n",
" self.opt_D_A.wd, self.opt_D_A.beta = self.learn.opt.wd, self.learn.opt.beta\n",
" self.opt_D_B.lr, self.opt_D_B.mom = self.learn.opt.lr, self.learn.opt.mom\n",
" self.opt_D_B.wd, self.opt_D_B.beta = self.learn.opt.wd, self.learn.opt.beta\n",
" \n",
" def on_train_begin(self, metrics_names, **kwargs):\n",
" self.G_A,self.G_B = self.learn.model.G_A,self.learn.model.G_B\n",
" self.D_A,self.D_B = self.learn.model.D_A,self.learn.model.D_B\n",
" self.crit = self.learn.loss_func.crit\n",
" if not getattr(self,'opt_G',None):\n",
" self.opt_G = self.learn.opt.new([nn.Sequential(*flatten_model(self.G_A), *flatten_model(self.G_B))])\n",
" else: \n",
" self.opt_G.lr,self.opt_G.wd = self.opt.lr,self.opt.wd\n",
" self.opt_G.mom,self.opt_G.beta = self.opt.mom,self.opt.beta\n",
" if not getattr(self,'opt_D_A',None):\n",
" self.opt_D_A = self.learn.opt.new([nn.Sequential(*flatten_model(self.D_A))])\n",
" if not getattr(self,'opt_D_B',None):\n",
" self.opt_D_B = self.learn.opt.new([nn.Sequential(*flatten_model(self.D_B))])\n",
" self.learn.opt.opt = self.opt_G.opt\n",
" self._set_trainable()\n",
" self.id_smter,self.gen_smter,self.cyc_smter = SmoothenValue(0.98),SmoothenValue(0.98),SmoothenValue(0.98)\n",
" self.da_smter,self.db_smter = SmoothenValue(0.98),SmoothenValue(0.98)\n",
" self.recorder.add_metric_names(['id_loss', 'gen_loss', 'cyc_loss', 'D_A_loss', 'D_B_loss'])\n",
" \n",
" def on_epoch_begin(self, **kwargs):\n",
" torch.cuda.empty_cache()\n",
" \n",
" def on_batch_begin(self, last_input, **kwargs):\n",
" self.training = self.learn.model.training\n",
" self.learn.loss_func.set_training(self.training)\n",
" self.learn.loss_func.set_input(last_input)\n",
" \n",
" \n",
" def on_backward_begin(self, **kwargs):\n",
" self.id_smter.add_value(self.loss_func.id_loss.detach().cpu())\n",
" self.gen_smter.add_value(self.loss_func.gen_loss.detach().cpu())\n",
" self.cyc_smter.add_value(self.loss_func.cyc_loss.detach().cpu())\n",
" \n",
" def on_batch_end(self, last_input, last_output, **kwargs):\n",
" self.G_A.zero_grad(); self.G_B.zero_grad()\n",
" fake_A, fake_B = last_output[0].detach(), last_output[1].detach()\n",
" real_A, real_B = last_input\n",
" self._set_trainable(D_A=True)\n",
" self.D_A.zero_grad()\n",
" loss_D_A = 0.5 * (self.crit(self.D_A(real_A), True) + self.crit(self.D_A(fake_A), False))\n",
" self.da_smter.add_value(loss_D_A.detach().cpu())\n",
" if self.training:\n",
" loss_D_A.backward()\n",
" self.opt_D_A.step()\n",
" self._set_trainable(D_B=True)\n",
" self.D_B.zero_grad()\n",
" loss_D_B = 0.5 * (self.crit(self.D_B(real_B), True) + self.crit(self.D_B(fake_B), False))\n",
" self.db_smter.add_value(loss_D_B.detach().cpu())\n",
"\n",
" if self.training:\n",
" loss_D_B.backward()\n",
" self.opt_D_B.step()\n",
" self._set_trainable()\n",
" \n",
" def on_epoch_end(self, last_metrics, **kwargs):\n",
" return add_metrics(last_metrics, [s.smooth for s in [self.id_smter,self.gen_smter,self.cyc_smter,\n",
" self.da_smter,self.db_smter]])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Training"
]
},
{
"cell_type": "code",
"execution_count": 26,
"metadata": {},
"outputs": [],
"source": [
"# Taken from https://github.com/fastai/fastai/blob/master/fastai/callbacks/flat_cos_anneal.py\n",
"# Work of Zach Mueller and Mikhail Grankin\n",
"from fastai.callback import *\n",
"from fastai.callbacks import *\n",
"def FlatAnnealScheduler(learn, lr:float=2e-4, n_epochs:int=100, n_epochs_decay:int=100, curve:str='linear'):\n",
" tot_epochs = n_epochs + n_epochs_decay\n",
" start_pct = n_epochs/tot_epochs\n",
" n = len(learn.data.train_dl)\n",
" anneal_start = int(n * tot_epochs * start_pct)\n",
" batch_finish = ((n * tot_epochs) - anneal_start)\n",
" if curve==\"cosine\": curve_type=annealing_cos\n",
" elif curve==\"linear\": curve_type=annealing_linear\n",
" elif curve==\"exponential\": curve_type=annealing_exp\n",
" else: raise ValueError(f\"annealing type not supported {curve}\")\n",
" phase0 = TrainingPhase(anneal_start).schedule_hp('lr', lr)\n",
" phase1 = TrainingPhase(batch_finish).schedule_hp('lr', lr, anneal=curve_type)\n",
" phases = [phase0, phase1]\n",
" return GeneralScheduler(learn, phases)"
]
},
{
"cell_type": "code",
"execution_count": 27,
"metadata": {},
"outputs": [],
"source": [
"def fit_fa(learn:Learner, n_epochs:int=100, n_epochs_decay:int=100, lr:float=2e-4, curve:str='linear',\n",
" wd:float=None, callbacks:Optional[CallbackList]=None)->None:\n",
" \"Fit a model with Flat Cosine Annealing\"\n",
" max_lr = learn.lr_range(lr)\n",
" callbacks = listify(callbacks)\n",
" callbacks.append(FlatAnnealScheduler(learn, lr, n_epochs, n_epochs_decay, curve))\n",
" learn.fit(n_epochs+n_epochs_decay, max_lr, wd=wd, callbacks=callbacks)\n",
" \n",
"Learner.fit_fa = fit_fa"
]
},
{
"cell_type": "code",
"execution_count": 28,
"metadata": {},
"outputs": [],
"source": [
"class AverageMetric(LearnerCallback):\n",
" \"Wrap a `func` in a callback for metrics computation.\"\n",
" def __init__(self, func):\n",
" # If func has a __name__ use this one else it should be a partial\n",
" name = func.__name__ if hasattr(func, '__name__') else func.func.__name__\n",
" self.func, self.name = func, name\n",
" self.world = num_distrib()\n",
"\n",
" def on_epoch_begin(self, **kwargs):\n",
" \"Set the inner value to 0.\"\n",
" self.val, self.count = 0.,0\n",
"\n",
" def on_batch_begin(self, last_input, **kwargs):\n",
" self.last_input = last_input\n",
" \n",
" def on_batch_end(self, last_output, last_target, **kwargs):\n",
" \"Update metric computation with `last_output` and `last_target`.\"\n",
" if not is_listy(last_target): last_target=[last_target]\n",
" self.count += first_el(last_target).size(0)\n",
" val = self.func(self.last_input, last_output, *last_target)\n",
" if self.world:\n",
" val = val.clone()\n",
" dist.all_reduce(val, op=dist.ReduceOp.SUM)\n",
" val /= self.world\n",
" self.val += first_el(last_target).size(0) * val.detach().cpu()\n",
"\n",
" def on_epoch_end(self, last_metrics, **kwargs):\n",
" \"Set the final result in `last_metrics`.\"\n",
" return add_metrics(last_metrics, self.val/self.count)\n"
]
},
{
"cell_type": "code",
"execution_count": 29,
"metadata": {},
"outputs": [],
"source": [
"from metrics import *"
]
},
{
"cell_type": "code",
"execution_count": 30,
"metadata": {},
"outputs": [],
"source": [
"def ssim_fastai(xb,yb,_):\n",
" real_A, real_B = xb\n",
" fake_A, fake_B = yb[:,0,:,:,:], yb[:,1,:,:,:]\n",
" real_A = (real_A/2 + 0.5)*255\n",
" fake_B = (fake_B/2 + 0.5)*255\n",
" return ssim(real_A,fake_B)"
]
},
{
"cell_type": "code",
"execution_count": 31,
"metadata": {},
"outputs": [],
"source": [
"def psnr_fastai(xb, yb, _):\n",
" real_A, real_B = xb\n",
" fake_A, fake_B = yb[:,0,:,:,:], yb[:,1,:,:,:]\n",
" real_A = (real_A/2 + 0.5)*255\n",
" fake_B = (fake_B/2 + 0.5)*255\n",
" return psnr(real_A,fake_B)"
]
},
{
"cell_type": "code",
"execution_count": 32,
"metadata": {},
"outputs": [],
"source": [
"cycle_gan = CycleGAN(3,3, gen_blocks=9)\n",
"learn = Learner(data, cycle_gan, loss_func=CycleGanLoss(cycle_gan), opt_func=partial(optim.Adam, betas=(0.5,0.999)),\n",
" callback_fns=[CycleGANTrainer],metrics=[AverageMetric(ssim_fastai),AverageMetric(psnr_fastai)])"
]
},
{
"cell_type": "code",
"execution_count": 33,
"metadata": {},
"outputs": [],
"source": [
"#learn.lr_find()"
]
},
{
"cell_type": "code",
"execution_count": 34,
"metadata": {},
"outputs": [],
"source": [
"#learn.recorder.plot()"
]
},
{
"cell_type": "code",
"execution_count": 35,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: left;\">\n",
" <th>epoch</th>\n",
" <th>train_loss</th>\n",
" <th>valid_loss</th>\n",
" <th>ssim_fastai</th>\n",
" <th>psnr_fastai</th>\n",
" <th>id_loss</th>\n",
" <th>gen_loss</th>\n",
" <th>cyc_loss</th>\n",
" <th>D_A_loss</th>\n",
" <th>D_B_loss</th>\n",
" <th>time</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <td>0</td>\n",
" <td>6.561883</td>\n",
" <td>5.135588</td>\n",
" <td>0.473679</td>\n",
" <td>9.387200</td>\n",
" <td>1.884625</td>\n",
" <td>0.796862</td>\n",
" <td>3.880396</td>\n",
" <td>0.324226</td>\n",
" <td>0.321400</td>\n",
" <td>00:27</td>\n",
" </tr>\n",
" <tr>\n",
" <td>1</td>\n",
" <td>5.647283</td>\n",
" <td>4.936441</td>\n",
" <td>0.465910</td>\n",
" <td>10.266727</td>\n",
" <td>1.607140</td>\n",
" <td>0.716910</td>\n",
" <td>3.323231</td>\n",
" <td>0.280169</td>\n",
" <td>0.277287</td>\n",
" <td>00:27</td>\n",
" </tr>\n",
" <tr>\n",
" <td>2</td>\n",
" <td>5.211876</td>\n",
" <td>4.717415</td>\n",
" <td>0.565417</td>\n",
" <td>9.830503</td>\n",
" <td>1.462108</td>\n",
" <td>0.719012</td>\n",
" <td>3.030756</td>\n",
" <td>0.264041</td>\n",
" <td>0.256112</td>\n",
" <td>00:27</td>\n",
" </tr>\n",
" <tr>\n",
" <td>3</td>\n",
" <td>4.929071</td>\n",
" <td>7.111959</td>\n",
" <td>0.502945</td>\n",
" <td>8.873150</td>\n",
" <td>1.369080</td>\n",
" <td>0.718374</td>\n",
" <td>2.841614</td>\n",
" <td>0.286366</td>\n",
" <td>0.260715</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>4</td>\n",
" <td>4.971952</td>\n",
" <td>4.743098</td>\n",
" <td>0.542775</td>\n",
" <td>9.094391</td>\n",
" <td>1.351116</td>\n",
" <td>0.769224</td>\n",
" <td>2.851610</td>\n",
" <td>0.250340</td>\n",
" <td>0.247486</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>5</td>\n",
" <td>4.926872</td>\n",
" <td>5.141806</td>\n",
" <td>0.515856</td>\n",
" <td>9.646178</td>\n",
" <td>1.288682</td>\n",
" <td>0.873950</td>\n",
" <td>2.764238</td>\n",
" <td>0.273015</td>\n",
" <td>0.241806</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>6</td>\n",
" <td>4.840660</td>\n",
" <td>4.774198</td>\n",
" <td>0.530930</td>\n",
" <td>8.995791</td>\n",
" <td>1.263278</td>\n",
" <td>0.897851</td>\n",
" <td>2.679529</td>\n",
" <td>0.246634</td>\n",
" <td>0.240409</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>7</td>\n",
" <td>4.715111</td>\n",
" <td>5.307607</td>\n",
" <td>0.485308</td>\n",
" <td>8.678684</td>\n",
" <td>1.200815</td>\n",
" <td>0.959975</td>\n",
" <td>2.554320</td>\n",
" <td>0.208429</td>\n",
" <td>0.203911</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>8</td>\n",
" <td>4.717269</td>\n",
" <td>5.407540</td>\n",
" <td>0.478808</td>\n",
" <td>8.845898</td>\n",
" <td>1.175990</td>\n",
" <td>1.029106</td>\n",
" <td>2.512172</td>\n",
" <td>0.190207</td>\n",
" <td>0.200359</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>9</td>\n",
" <td>4.771541</td>\n",
" <td>4.586983</td>\n",
" <td>0.521542</td>\n",
" <td>9.332252</td>\n",
" <td>1.168054</td>\n",
" <td>1.083328</td>\n",
" <td>2.520158</td>\n",
" <td>0.220892</td>\n",
" <td>0.198271</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>10</td>\n",
" <td>4.653002</td>\n",
" <td>5.640585</td>\n",
" <td>0.528441</td>\n",
" <td>9.376346</td>\n",
" <td>1.133995</td>\n",
" <td>1.054805</td>\n",
" <td>2.464202</td>\n",
" <td>0.203103</td>\n",
" <td>0.278971</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>11</td>\n",
" <td>4.518406</td>\n",
" <td>4.460074</td>\n",
" <td>0.426281</td>\n",
" <td>8.360112</td>\n",
" <td>1.089192</td>\n",
" <td>1.063930</td>\n",
" <td>2.365285</td>\n",
" <td>0.186630</td>\n",
" <td>0.224311</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>12</td>\n",
" <td>4.445597</td>\n",
" <td>4.410108</td>\n",
" <td>0.473550</td>\n",
" <td>8.817196</td>\n",
" <td>1.074073</td>\n",
" <td>1.043954</td>\n",
" <td>2.327569</td>\n",
" <td>0.207671</td>\n",
" <td>0.253871</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>13</td>\n",
" <td>4.358193</td>\n",
" <td>4.004363</td>\n",
" <td>0.533533</td>\n",
" <td>9.450353</td>\n",
" <td>1.067293</td>\n",
" <td>0.979354</td>\n",
" <td>2.311545</td>\n",
" <td>0.214134</td>\n",
" <td>0.234769</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>14</td>\n",
" <td>4.156375</td>\n",
" <td>3.787990</td>\n",
" <td>0.524728</td>\n",
" <td>9.441799</td>\n",
" <td>1.031719</td>\n",
" <td>0.918310</td>\n",
" <td>2.206345</td>\n",
" <td>0.236098</td>\n",
" <td>0.212539</td>\n",
" <td>00:29</td>\n",
" </tr>\n",
" <tr>\n",
" <td>15</td>\n",
" <td>4.053391</td>\n",
" <td>3.995172</td>\n",
" <td>0.527359</td>\n",
" <td>10.342698</td>\n",
" <td>1.005355</td>\n",
" <td>0.885043</td>\n",
" <td>2.162991</td>\n",
" <td>0.240513</td>\n",
" <td>0.207699</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>16</td>\n",
" <td>4.067924</td>\n",
" <td>3.625066</td>\n",
" <td>0.502665</td>\n",
" <td>9.238791</td>\n",
" <td>1.003777</td>\n",
" <td>0.911058</td>\n",
" <td>2.153088</td>\n",
" <td>0.246938</td>\n",
" <td>0.191794</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>17</td>\n",
" <td>4.037714</td>\n",
" <td>3.711069</td>\n",
" <td>0.464412</td>\n",
" <td>7.298216</td>\n",
" <td>1.000791</td>\n",
" <td>0.896303</td>\n",
" <td>2.140619</td>\n",
" <td>0.249679</td>\n",
" <td>0.187362</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>18</td>\n",
" <td>3.857249</td>\n",
" <td>4.345922</td>\n",
" <td>0.544652</td>\n",
" <td>9.388030</td>\n",
" <td>0.948343</td>\n",
" <td>0.891913</td>\n",
" <td>2.016993</td>\n",
" <td>0.257431</td>\n",
" <td>0.198414</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>19</td>\n",
" <td>3.762412</td>\n",
" <td>4.109817</td>\n",
" <td>0.490418</td>\n",
" <td>9.542933</td>\n",
" <td>0.926382</td>\n",
" <td>0.875323</td>\n",
" <td>1.960707</td>\n",
" <td>0.261851</td>\n",
" <td>0.210432</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>20</td>\n",
" <td>3.702829</td>\n",
" <td>3.526021</td>\n",
" <td>0.464647</td>\n",
" <td>8.307538</td>\n",
" <td>0.898579</td>\n",
" <td>0.908769</td>\n",
" <td>1.895481</td>\n",
" <td>0.261681</td>\n",
" <td>0.217534</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>21</td>\n",
" <td>3.622834</td>\n",
" <td>3.706043</td>\n",
" <td>0.502107</td>\n",
" <td>10.158941</td>\n",
" <td>0.875865</td>\n",
" <td>0.904829</td>\n",
" <td>1.842140</td>\n",
" <td>0.240720</td>\n",
" <td>0.208291</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>22</td>\n",
" <td>3.538983</td>\n",
" <td>3.435079</td>\n",
" <td>0.546980</td>\n",
" <td>9.465912</td>\n",
" <td>0.851039</td>\n",
" <td>0.919320</td>\n",
" <td>1.768624</td>\n",
" <td>0.258890</td>\n",
" <td>0.202507</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>23</td>\n",
" <td>3.513905</td>\n",
" <td>3.234569</td>\n",
" <td>0.508788</td>\n",
" <td>8.967988</td>\n",
" <td>0.841775</td>\n",
" <td>0.934261</td>\n",
" <td>1.737869</td>\n",
" <td>0.246404</td>\n",
" <td>0.231526</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>24</td>\n",
" <td>3.515435</td>\n",
" <td>3.423039</td>\n",
" <td>0.495962</td>\n",
" <td>8.955706</td>\n",
" <td>0.829034</td>\n",
" <td>0.963347</td>\n",
" <td>1.723054</td>\n",
" <td>0.217231</td>\n",
" <td>0.240800</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>25</td>\n",
" <td>3.495776</td>\n",
" <td>3.450833</td>\n",
" <td>0.489034</td>\n",
" <td>9.838599</td>\n",
" <td>0.812738</td>\n",
" <td>1.014734</td>\n",
" <td>1.668304</td>\n",
" <td>0.194369</td>\n",
" <td>0.248123</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>26</td>\n",
" <td>3.449779</td>\n",
" <td>3.365834</td>\n",
" <td>0.487408</td>\n",
" <td>9.713398</td>\n",
" <td>0.795532</td>\n",
" <td>1.049908</td>\n",
" <td>1.604338</td>\n",
" <td>0.181287</td>\n",
" <td>0.248181</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>27</td>\n",
" <td>3.393342</td>\n",
" <td>3.238819</td>\n",
" <td>0.534515</td>\n",
" <td>9.728602</td>\n",
" <td>0.766979</td>\n",
" <td>1.090326</td>\n",
" <td>1.536036</td>\n",
" <td>0.172006</td>\n",
" <td>0.235865</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>28</td>\n",
" <td>3.330395</td>\n",
" <td>3.262840</td>\n",
" <td>0.526032</td>\n",
" <td>9.696510</td>\n",
" <td>0.745259</td>\n",
" <td>1.104066</td>\n",
" <td>1.481069</td>\n",
" <td>0.159359</td>\n",
" <td>0.235580</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" <tr>\n",
" <td>29</td>\n",
" <td>3.290842</td>\n",
" <td>3.225084</td>\n",
" <td>0.528845</td>\n",
" <td>9.593530</td>\n",
" <td>0.723691</td>\n",
" <td>1.135712</td>\n",
" <td>1.431439</td>\n",
" <td>0.151975</td>\n",
" <td>0.231465</td>\n",
" <td>00:28</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"learn.fit_fa(lr=2e-4,n_epochs=15,n_epochs_decay=15)"
]
},
{
"cell_type": "code",
"execution_count": 36,
"metadata": {},
"outputs": [],
"source": [
"learn.save('30fit')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's look at some results using `Learner.show_results`."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"learn.show_results(ds_type=DatasetType.Train, rows=2)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"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."
]
},
{
"cell_type": "code",
"execution_count": 38,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(100, 344)"
]
},
"execution_count": 38,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"len(learn.data.train_ds.items),len(learn.data.train_ds.itemsB)"
]
},
{
"cell_type": "code",
"execution_count": 39,
"metadata": {},
"outputs": [],
"source": [
"def get_batch(filenames, tfms, **kwargs):\n",
" samples = [open_image(fn) for fn in filenames]\n",
" for s in samples: s = s.apply_tfms(tfms, **kwargs)\n",
" batch = torch.stack([s.data for s in samples], 0).cuda()\n",
" return 2. * (batch - 0.5)"
]
},
{
"cell_type": "code",
"execution_count": 40,
"metadata": {},
"outputs": [],
"source": [
"fnames = learn.data.train_ds.items[:8]"
]
},
{
"cell_type": "code",
"execution_count": 41,
"metadata": {},
"outputs": [],
"source": [
"x = get_batch(fnames, get_transforms()[1], size=128)"
]
},
{
"cell_type": "code",
"execution_count": 42,
"metadata": {},
"outputs": [],
"source": [
"learn.model.eval()\n",
"tfms = get_transforms()[1]\n",
"bs = 16"
]
},
{
"cell_type": "code",
"execution_count": 43,
"metadata": {},
"outputs": [],
"source": [
"def get_losses(fnames, gen, crit, bs=16):\n",
" losses_in,losses_out = [],[]\n",
" with torch.no_grad():\n",
" for i in progress_bar(range(0, len(fnames), bs)):\n",
" xb = get_batch(fnames[i:i+bs], tfms, size=128)\n",
" fakes = gen(xb)\n",
" preds_in,preds_out = crit(xb),crit(fakes)\n",
" loss_in = learn.loss_func.crit(preds_in, True,reduction='none')\n",
" loss_out = learn.loss_func.crit(preds_out,True,reduction='none')\n",
" losses_in.append(loss_in.view(loss_in.size(0),-1).mean(1))\n",
" losses_out.append(loss_out.view(loss_out.size(0),-1).mean(1))\n",
" return torch.cat(losses_in),torch.cat(losses_out)"
]
},
{
"cell_type": "code",
"execution_count": 44,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"\n",
" <div>\n",
" <style>\n",
" /* Turns off some styling */\n",
" progress {\n",
" /* gets rid of default border in Firefox and Opera. */\n",
" border: none;\n",
" /* Needs to be in here for Safari polyfill so background images work as expected. */\n",
" background-size: auto;\n",
" }\n",
" .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {\n",
" background: #F44336;\n",
" }\n",
" </style>\n",
" <progress value='7' class='' max='7', style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
" 100.00% [7/7 00:03<00:00]\n",
" </div>\n",
" "
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"losses_A = get_losses(data.train_ds.x.items, learn.model.G_B, learn.model.D_B)"
]
},
{
"cell_type": "code",
"execution_count": 45,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"\n",
" <div>\n",
" <style>\n",
" /* Turns off some styling */\n",
" progress {\n",
" /* gets rid of default border in Firefox and Opera. */\n",
" border: none;\n",
" /* Needs to be in here for Safari polyfill so background images work as expected. */\n",
" background-size: auto;\n",
" }\n",
" .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {\n",
" background: #F44336;\n",
" }\n",
" </style>\n",
" <progress value='22' class='' max='22', style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
" 100.00% [22/22 00:13<00:00]\n",
" </div>\n",
" "
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"losses_B = get_losses(data.train_ds.x.itemsB, learn.model.G_A, learn.model.D_A)"
]
},
{
"cell_type": "code",
"execution_count": 46,
"metadata": {},
"outputs": [],
"source": [
"def show_best(fnames, losses, gen, n=8):\n",
" sort_idx = losses.argsort()\n",
" _,axs = plt.subplots(n//2, 4, figsize=(12,2*n))\n",
" xb = get_batch(fnames[sort_idx][:n], tfms, size=128)\n",
" with torch.no_grad():\n",
" fakes = gen(xb)\n",
" xb,fakes = (1+xb)/2,(1+fakes)/2\n",
" for i in range(n):\n",
" axs.flatten()[2*i].imshow(xb[i].permute(1,2,0).cpu())\n",
" axs.flatten()[2*i].axis('off')\n",
" axs.flatten()[2*i+1].imshow(fakes[i].permute(1,2,0).cpu())\n",
" axs.flatten()[2*i+1].set_title(losses[sort_idx][i].item())\n",
" axs.flatten()[2*i+1].axis('off')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"show_best(data.train_ds.x.items, losses_A[1].cpu(), learn.model.G_B)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"show_best(data.train_ds.x.itemsB, losses_B[1].cpu(), learn.model.G_A)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Save Image Predictions (Fake images)"
]
},
{
"cell_type": "code",
"execution_count": 49,
"metadata": {},
"outputs": [],
"source": [
"from torch.utils.data import Dataset, DataLoader\n",
"import torchvision\n",
"import glob\n",
"\n",
"class FolderDataset(Dataset):\n",
" def __init__(self, path,transforms=None):\n",
" self.files = glob.glob(path+'/*')\n",
" self.totensor = torchvision.transforms.ToTensor()\n",
" if transforms:\n",
" self.transform = torchvision.transforms.Compose(transforms)\n",
" else:\n",
" self.transform = lambda x: x\n",
" \n",
" def __len__(self):\n",
" return len(self.files)\n",
"\n",
" def __getitem__(self, idx):\n",
" image = PIL.Image.open(self.files[idx % len(self.files)])\n",
" image = self.totensor(image)\n",
" image = self.transform(image)\n",
" return self.files[idx], image\n",
"\n",
"def load_dataset(test_path):\n",
" dataset = FolderDataset(\n",
" path=test_path,\n",
" #transforms=[torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]\n",
" ) \n",
" loader = torch.utils.data.DataLoader(\n",
" dataset,\n",
" batch_size=2,\n",
" num_workers=4,\n",
" shuffle=True\n",
" )\n",
" return loader"
]
},
{
"cell_type": "code",
"execution_count": 50,
"metadata": {},
"outputs": [],
"source": [
"import tqdm\n",
"def get_preds_cyclegan(learn,test_path,pred_path,suffix='png'):\n",
" \n",
" assert os.path.exists(test_path)\n",
" \n",
" if not os.path.exists(pred_path):\n",
" os.mkdir(pred_path)\n",
" \n",
" model = learn.model.G_A\n",
" \n",
" test_dl = load_dataset(test_path)\n",
" \n",
" for i, xb in tqdm.tqdm(enumerate(test_dl),total=len(test_dl)):\n",
" fn, im = xb\n",
" preds = (learn.model.G_B(im.cuda())/2 + 0.5)\n",
" for i in range(len(fn)):\n",
" new_fn = os.path.join(pred_path,'.'.join([os.path.basename(fn[i]).split('.')[0]+'_fakeB',suffix])) \n",
" torchvision.utils.save_image(preds[i],new_fn)"
]
},
{
"cell_type": "code",
"execution_count": 51,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"100%|██████████| 50/50 [00:12<00:00, 3.86it/s]\n"
]
}
],
"source": [
"get_preds_cyclegan(learn,str(muse2he_path/'testA'),'./preds')"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.6"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment