a proposal for small sample problem

docs/proposals/algorithms/[OSPP]GAN and Self-taught Learning/[OSPP]A Proposal of Integrating GAN and Self-taught Learning to ianvs.md

@@ -0,0 +1,33 @@
+# Integrate GAN and Self-taught Learning into ianvs Lifelong Learning to Handle Unknown Tasks
+
+## Motivation
+
+In the process of ianvs lifelong learning, there would be a chance to confront unknown tasks, whose data are always heterogeneous small sample. Generate Adversarial Networks(GAN) is the start-of-art generative model and GAN can generate fake data according to the distribution of the real data. Naturally, we try to utilize GAN to handle small sample problem. Self-taught learning is an approach to improve classfication performance using sparse coding to construct higher-level features with the unlabeled data. Hence, we combine GAN and self-taught learning to help ianvs lifelong learning handle unknown tasks.
+
+### Goals
+
+* Handle unknown tasks
+* Implement of a lightweight GAN to solve small sample problem
+* Utilize self-taught learning to solve heterogeneous problem
+
+## Proposal
+We focus on the process of handling unknown tasks.     
+
+The overview is as follows:
+
+![](images/overview.png)
+
+The process is illustrated as below:    
+1. GAN exploits the unknown task sample to generate more fake sample. 
+2. Self-taught learning unit utilize the fake sample and orginal unknown task sample and its label to train a classifier.
+3. A well trained classifier is output.
+
+### GAN Design
+We use the networks design by [TOWARDS FASTER AND STABILIZED GAN TRAINING FOR HIGH-FIDELITY FEW-SHOT IMAGE SYNTHESIS](https://openreview.net/forum?id=1Fqg133qRaI). The design is aimed for small training data and pour computing devices. Therefore, it is perfectly suitable for handling unkwnon tasks of ianvs lifelong learning. The networks is shown below.    
+
+![](images/GAN.png)
+
+### Self-taught Learing Design
+Self-taught learning uses unlabeled data to find the latent feature of data and then makes every labeled data a represention using the latent feature and uses the represention and label corresponding to train classifier.    
+
+![](images/self-taught%20learning.png)

examples/GANwithSelf-taughtLearning/GAN/diffaug.py

@@ -0,0 +1,76 @@
+# Differentiable Augmentation for Data-Efficient GAN Training
+# Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han
+# https://arxiv.org/pdf/2006.10738
+
+import torch
+import torch.nn.functional as F
+
+
+def DiffAugment(x, policy='', channels_first=True):
+    if policy:
+        if not channels_first:
+            x = x.permute(0, 3, 1, 2)
+        for p in policy.split(','):
+            for f in AUGMENT_FNS[p]:
+                x = f(x)
+        if not channels_first:
+            x = x.permute(0, 2, 3, 1)
+        x = x.contiguous()
+    return x
+
+
+def rand_brightness(x):
+    x = x + (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) - 0.5)
+    return x
+
+
+def rand_saturation(x):
+    x_mean = x.mean(dim=1, keepdim=True)
+    x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) * 2) + x_mean
+    return x
+
+
+def rand_contrast(x):
+    x_mean = x.mean(dim=[1, 2, 3], keepdim=True)
+    x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) + 0.5) + x_mean
+    return x
+
+
+def rand_translation(x, ratio=0.125):
+    shift_x, shift_y = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
+    translation_x = torch.randint(-shift_x, shift_x + 1, size=[x.size(0), 1, 1], device=x.device)
+    translation_y = torch.randint(-shift_y, shift_y + 1, size=[x.size(0), 1, 1], device=x.device)
+    grid_batch, grid_x, grid_y = torch.meshgrid(
+        torch.arange(x.size(0), dtype=torch.long, device=x.device),
+        torch.arange(x.size(2), dtype=torch.long, device=x.device),
+        torch.arange(x.size(3), dtype=torch.long, device=x.device),
+    )
+    grid_x = torch.clamp(grid_x + translation_x + 1, 0, x.size(2) + 1)
+    grid_y = torch.clamp(grid_y + translation_y + 1, 0, x.size(3) + 1)
+    x_pad = F.pad(x, [1, 1, 1, 1, 0, 0, 0, 0])
+    x = x_pad.permute(0, 2, 3, 1).contiguous()[grid_batch, grid_x, grid_y].permute(0, 3, 1, 2)
+    return x
+
+
+def rand_cutout(x, ratio=0.5):
+    cutout_size = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
+    offset_x = torch.randint(0, x.size(2) + (1 - cutout_size[0] % 2), size=[x.size(0), 1, 1], device=x.device)
+    offset_y = torch.randint(0, x.size(3) + (1 - cutout_size[1] % 2), size=[x.size(0), 1, 1], device=x.device)
+    grid_batch, grid_x, grid_y = torch.meshgrid(
+        torch.arange(x.size(0), dtype=torch.long, device=x.device),
+        torch.arange(cutout_size[0], dtype=torch.long, device=x.device),
+        torch.arange(cutout_size[1], dtype=torch.long, device=x.device),
+    )
+    grid_x = torch.clamp(grid_x + offset_x - cutout_size[0] // 2, min=0, max=x.size(2) - 1)
+    grid_y = torch.clamp(grid_y + offset_y - cutout_size[1] // 2, min=0, max=x.size(3) - 1)
+    mask = torch.ones(x.size(0), x.size(2), x.size(3), dtype=x.dtype, device=x.device)
+    mask[grid_batch, grid_x, grid_y] = 0
+    x = x * mask.unsqueeze(1)
+    return x
+
+
+AUGMENT_FNS = {
+    'color': [rand_brightness, rand_saturation, rand_contrast],
+    'translation': [rand_translation],
+    'cutout': [rand_cutout],
+}

examples/GANwithSelf-taughtLearning/GAN/generate_fake_imgs.py

@@ -0,0 +1,46 @@
+import torch
+
+from models import Generator, weights_init
+
+import matplotlib.pyplot as plt
+
+import os
+
+from collections import OrderedDict
+
+import numpy as np
+
+from skimage import io
+
+# print(os.getcwd())
+
+device = 'cuda'
+
+ngf = 64
+nz = 256
+im_size = 1024
+netG = Generator(ngf=ngf, nz=nz, im_size=im_size).to(device)
+weights_init(netG)
+weights = torch.load(os.getcwd() + '/train_results/test1/models/50000.pth')
+# print(weights['g'])
+netG_weights = OrderedDict()
+for name, weight in weights['g'].items():
+    name = name.split('.')[1:]
+    name = '.'.join(name)
+    netG_weights[name] = weight
+netG.load_state_dict(netG_weights)
+current_batch_size = 1
+
+
+index = 1
+while index <= 3000:
+    noise = torch.Tensor(current_batch_size, nz).normal_(0, 1).to(device)
+    fake_images = netG(noise)[0]
+    for fake_image in fake_images:
+        fake_image = fake_image.detach().cpu().numpy().transpose(1, 2, 0)
+        fake_image = fake_image * np.array([0.5, 0.5, 0.5])
+        fake_image = fake_image + np.array([0.5, 0.5, 0.5])
+        fake_image = (fake_image * 255).astype(np.uint8)
+        io.imsave('../data/fake_imgs1/' + str(index) + '.png', fake_image)
+        print('figure {} done'.format(index))
+        index += 1

examples/GANwithSelf-taughtLearning/GAN/lpips/__init__.py

@@ -0,0 +1,168 @@
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import skimage
+import torch
+from torch.autograd import Variable
+
+from lpips import dist_model
+
+
+if skimage.__version__ == '0.14.3':
+    from skimage.measure import compare_ssim
+else:
+    from skimage.metrics import structural_similarity as compare_ssim
+
+
+
+class PerceptualLoss(torch.nn.Module):
+    def __init__(self, model='net-lin', net='alex', colorspace='rgb', spatial=False, use_gpu=True, gpu_ids=[0]): # VGG using our perceptually-learned weights (LPIPS metric)
+    # def __init__(self, model='net', net='vgg', use_gpu=True): # "default" way of using VGG as a perceptual loss
+        super(PerceptualLoss, self).__init__()
+        print('Setting up Perceptual loss...')
+        self.use_gpu = use_gpu
+        self.spatial = spatial
+        self.gpu_ids = gpu_ids
+        self.model = dist_model.DistModel()
+        self.model.initialize(model=model, net=net, use_gpu=use_gpu, colorspace=colorspace, spatial=self.spatial, gpu_ids=gpu_ids)
+        print('...[%s] initialized'%self.model.name())
+        print('...Done')
+
+    def forward(self, pred, target, normalize=False):
+        """
+        Pred and target are Variables.
+        If normalize is True, assumes the images are between [0,1] and then scales them between [-1,+1]
+        If normalize is False, assumes the images are already between [-1,+1]
+
+        Inputs pred and target are Nx3xHxW
+        Output pytorch Variable N long
+        """
+
+        if normalize:
+            target = 2 * target  - 1
+            pred = 2 * pred  - 1
+
+        return self.model.forward(target, pred)
+
+def normalize_tensor(in_feat,eps=1e-10):
+    norm_factor = torch.sqrt(torch.sum(in_feat**2,dim=1,keepdim=True))
+    return in_feat/(norm_factor+eps)
+
+def l2(p0, p1, range=255.):
+    return .5*np.mean((p0 / range - p1 / range)**2)
+
+def psnr(p0, p1, peak=255.):
+    return 10*np.log10(peak**2/np.mean((1.*p0-1.*p1)**2))
+
+def dssim(p0, p1, range=255.):
+    return (1 - compare_ssim(p0, p1, data_range=range, multichannel=True)) / 2.
+
+def rgb2lab(in_img,mean_cent=False):
+    from skimage import color
+    img_lab = color.rgb2lab(in_img)
+    if(mean_cent):
+        img_lab[:,:,0] = img_lab[:,:,0]-50
+    return img_lab
+
+def tensor2np(tensor_obj):
+    # change dimension of a tensor object into a numpy array
+    return tensor_obj[0].cpu().float().numpy().transpose((1,2,0))
+
+def np2tensor(np_obj):
+     # change dimenion of np array into tensor array
+    return torch.Tensor(np_obj[:, :, :, np.newaxis].transpose((3, 2, 0, 1)))
+
+def tensor2tensorlab(image_tensor,to_norm=True,mc_only=False):
+    # image tensor to lab tensor
+    from skimage import color
+
+    img = tensor2im(image_tensor)
+    img_lab = color.rgb2lab(img)
+    if(mc_only):
+        img_lab[:,:,0] = img_lab[:,:,0]-50
+    if(to_norm and not mc_only):
+        img_lab[:,:,0] = img_lab[:,:,0]-50
+        img_lab = img_lab/100.
+
+    return np2tensor(img_lab)
+
+def tensorlab2tensor(lab_tensor,return_inbnd=False):
+    from skimage import color
+    import warnings
+    warnings.filterwarnings("ignore")
+
+    lab = tensor2np(lab_tensor)*100.
+    lab[:,:,0] = lab[:,:,0]+50
+
+    rgb_back = 255.*np.clip(color.lab2rgb(lab.astype('float')),0,1)
+    if(return_inbnd):
+        # convert back to lab, see if we match
+        lab_back = color.rgb2lab(rgb_back.astype('uint8'))
+        mask = 1.*np.isclose(lab_back,lab,atol=2.)
+        mask = np2tensor(np.prod(mask,axis=2)[:,:,np.newaxis])
+        return (im2tensor(rgb_back),mask)
+    else:
+        return im2tensor(rgb_back)
+
+def rgb2lab(input):
+    from skimage import color
+    return color.rgb2lab(input / 255.)
+
+def tensor2im(image_tensor, imtype=np.uint8, cent=1., factor=255./2.):
+    image_numpy = image_tensor[0].cpu().float().numpy()
+    image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + cent) * factor
+    return image_numpy.astype(imtype)
+
+def im2tensor(image, imtype=np.uint8, cent=1., factor=255./2.):
+    return torch.Tensor((image / factor - cent)
+                        [:, :, :, np.newaxis].transpose((3, 2, 0, 1)))
+
+def tensor2vec(vector_tensor):
+    return vector_tensor.data.cpu().numpy()[:, :, 0, 0]
+
+def voc_ap(rec, prec, use_07_metric=False):
+    """ ap = voc_ap(rec, prec, [use_07_metric])
+    Compute VOC AP given precision and recall.
+    If use_07_metric is true, uses the
+    VOC 07 11 point method (default:False).
+    """
+    if use_07_metric:
+        # 11 point metric
+        ap = 0.
+        for t in np.arange(0., 1.1, 0.1):
+            if np.sum(rec >= t) == 0:
+                p = 0
+            else:
+                p = np.max(prec[rec >= t])
+            ap = ap + p / 11.
+    else:
+        # correct AP calculation
+        # first append sentinel values at the end
+        mrec = np.concatenate(([0.], rec, [1.]))
+        mpre = np.concatenate(([0.], prec, [0.]))
+
+        # compute the precision envelope
+        for i in range(mpre.size - 1, 0, -1):
+            mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
+
+        # to calculate area under PR curve, look for points
+        # where X axis (recall) changes value
+        i = np.where(mrec[1:] != mrec[:-1])[0]
+
+        # and sum (\Delta recall) * prec
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
+    return ap
+
+def tensor2im(image_tensor, imtype=np.uint8, cent=1., factor=255./2.):
+# def tensor2im(image_tensor, imtype=np.uint8, cent=1., factor=1.):
+    image_numpy = image_tensor[0].cpu().float().numpy()
+    image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + cent) * factor
+    return image_numpy.astype(imtype)
+
+def im2tensor(image, imtype=np.uint8, cent=1., factor=255./2.):
+# def im2tensor(image, imtype=np.uint8, cent=1., factor=1.):
+    return torch.Tensor((image / factor - cent)
+                        [:, :, :, np.newaxis].transpose((3, 2, 0, 1)))

examples/GANwithSelf-taughtLearning/GAN/lpips/base_model.py

@@ -0,0 +1,58 @@
+import os
+import torch
+from torch.autograd import Variable
+from pdb import set_trace as st
+from IPython import embed
+
+class BaseModel():
+    def __init__(self):
+        pass;
+
+    def name(self):
+        return 'BaseModel'
+
+    def initialize(self, use_gpu=True, gpu_ids=[0]):
+        self.use_gpu = use_gpu
+        self.gpu_ids = gpu_ids
+
+    def forward(self):
+        pass
+
+    def get_image_paths(self):
+        pass
+
+    def optimize_parameters(self):
+        pass
+
+    def get_current_visuals(self):
+        return self.input
+
+    def get_current_errors(self):
+        return {}
+
+    def save(self, label):
+        pass
+
+    # helper saving function that can be used by subclasses
+    def save_network(self, network, path, network_label, epoch_label):
+        save_filename = '%s_net_%s.pth' % (epoch_label, network_label)
+        save_path = os.path.join(path, save_filename)
+        torch.save(network.state_dict(), save_path)
+
+    # helper loading function that can be used by subclasses
+    def load_network(self, network, network_label, epoch_label):
+        save_filename = '%s_net_%s.pth' % (epoch_label, network_label)
+        save_path = os.path.join(self.save_dir, save_filename)
+        print('Loading network from %s'%save_path)
+        network.load_state_dict(torch.load(save_path))
+
+    def update_learning_rate():
+        pass
+
+    def get_image_paths(self):
+        return self.image_paths
+
+    def save_done(self, flag=False):
+        np.save(os.path.join(self.save_dir, 'done_flag'),flag)
+        np.savetxt(os.path.join(self.save_dir, 'done_flag'),[flag,],fmt='%i')
+