utils.py

import torch.nn as nn
import torch.nn.functional as F
import torch

import sys 
import numpy as np 
import matplotlib.pyplot as plt

from torch.utils.data.sampler import SubsetRandomSampler
import torch
import torchvision 
import torch.optim as optim
import torch.nn as nn
from torch.autograd import Variable
from six import add_metaclass
from contextlib import contextmanager
import random
import pickle
import functools

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class BalancedBatchSampler(torch.utils.data.sampler.Sampler):
    def __init__(self, dataset, labels=None):
        self.labels = labels
        self.dataset = dict()
        self.balanced_max = 0
        # Save all the indices for all the classes
        for idx in range(0, len(dataset)):
            label = self._get_label(dataset, idx)
            if label not in self.dataset:
                self.dataset[label] = list()
            self.dataset[label].append(idx)
            self.balanced_max = len(self.dataset[label]) \
                if len(self.dataset[label]) > self.balanced_max else self.balanced_max
        
        # Oversample the classes with fewer elements than the max
        for label in self.dataset:
            while len(self.dataset[label]) < self.balanced_max:
                self.dataset[label].append(random.choice(self.dataset[label]))
        self.keys = list(self.dataset.keys())
        self.currentkey = 0
        self.indices = [-1]*len(self.keys)

    def __iter__(self):
        while self.indices[self.currentkey] < self.balanced_max - 1:
            self.indices[self.currentkey] += 1
            yield self.dataset[self.keys[self.currentkey]][self.indices[self.currentkey]]
            self.currentkey = (self.currentkey + 1) % len(self.keys)
        self.indices = [-1]*len(self.keys)
    
    def _get_label(self, dataset, idx, labels = None):
        if self.labels is not None:
            return self.labels[idx]
        else:
            # Trying guessing
            dataset_type = type(dataset)
            if is_torchvision_installed and dataset_type is torchvision.datasets.MNIST:
                return dataset.train_labels[idx].item()
            elif is_torchvision_installed and dataset_type is torchvision.datasets.ImageFolder:
                return dataset.imgs[idx][1]
            else:
                raise Exception("You should pass the tensor of labels to the constructor as second argument")

    def __len__(self):
        return self.balanced_max*len(self.keys)

    
#cifar10_trainloader = torch.utils.data.DataLoader(trainset, sampler=BalancedBatchSampler(trainset, trainset.dataset.targets), batch_size=120)
#cifar10_testloader_balanced = torch.utils.data.DataLoader(cifar10_testset, sampler=BalancedBatchSampler(cifar10_testset, cifar10_testset.targets), batch_size=120)

##############################################################################
# ReparamModule
##############################################################################

class PatchModules(type):
    def __call__(cls, *args, **kwargs):
        r"""Called when you call ReparamModule(...) """
        net = type.__call__(cls, *args, **kwargs)

        # collect weight (module, name) pairs
        # flatten weights
        w_modules_names = []

        for m in net.modules():
            for n, p in m.named_parameters(recurse=False):
                if p is not None:
                    w_modules_names.append((m, n))
            for n, b in m.named_buffers(recurse=False):
                if b is not None:
                    print((
                        '{} contains buffer {}. The buffer will be treated as '
                        'a constant and assumed not to change during gradient '
                        'steps. If this assumption is violated (e.g., '
                        'BatchNorm*d\'s running_mean/var), the computation will '
                        'be incorrect.').format(m.__class__.__name__, n))

        net._weights_module_names = tuple(w_modules_names)

        # Put to correct device before we do stuff on parameters
        net = net.to(device)

        ws = tuple(m._parameters[n].detach() for m, n in w_modules_names)

        assert len(set(w.dtype for w in ws)) == 1

        # reparam to a single flat parameter
        net._weights_numels = tuple(w.numel() for w in ws)
        net._weights_shapes = tuple(w.shape for w in ws)
        with torch.no_grad():
            flat_w = torch.cat([w.reshape(-1) for w in ws], 0)

        # remove old parameters, assign the names as buffers
        for m, n in net._weights_module_names:
            delattr(m, n)
            m.register_buffer(n, None)

        # register the flat one
        net.register_parameter('flat_w', nn.Parameter(flat_w, requires_grad=True))

        return net


@add_metaclass(PatchModules)
class ReparamModule(nn.Module):
    def _apply(self, *args, **kwargs):
        rv = super(ReparamModule, self)._apply(*args, **kwargs)
        return rv

    def get_param(self, clone=False):
        if clone:
            return self.flat_w.detach().clone().requires_grad_(self.flat_w.requires_grad)
        return self.flat_w
    
    @contextmanager
    def unflatten_weight(self, flat_w):
        ws = (t.view(s) for (t, s) in zip(flat_w.split(self._weights_numels), self._weights_shapes))
        for (m, n), w in zip(self._weights_module_names, ws):
            setattr(m, n, w)
        yield
        for m, n in self._weights_module_names:
            setattr(m, n, None)            

    def forward_with_param(self, inp, new_w):
        with self.unflatten_weight(new_w):
            return nn.Module.__call__(self, inp)

    def __call__(self, inp):
        return self.forward_with_param(inp, self.flat_w)

    # make load_state_dict work on both
    # singleton dicts containing a flattened weight tensor and
    # full dicts containing unflattened weight tensors...
    def load_state_dict(self, state_dict, *args, **kwargs):
        if len(state_dict) == 1 and 'flat_w' in state_dict:
            return super(ReparamModule, self).load_state_dict(state_dict, *args, **kwargs)
        with self.unflatten_weight(self.flat_w):
            flat_w = self.flat_w
            del self.flat_w
            super(ReparamModule, self).load_state_dict(state_dict, *args, **kwargs)
        self.register_parameter('flat_w', flat_w)

    def reset(self, inplace=True):
        if inplace:
            flat_w = self.flat_w
        else:
            flat_w = torch.empty_like(self.flat_w).requires_grad_()
        with torch.no_grad():
            with self.unflatten_weight(flat_w):
                weights_init(self)
        return flat_w
def conv3x3(in_planes, out_planes, stride=1):
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True)

def conv_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        init.xavier_uniform(m.weight, gain=np.sqrt(2))
        init.constant(m.bias, 0)
    elif classname.find('BatchNorm') != -1:
        init.constant(m.weight, 1)
        init.constant(m.bias, 0)

class wide_basic(ReparamModule):
    def __init__(self, in_planes, planes, dropout_rate, stride=1):
        super(wide_basic, self).__init__()
#         self.bn1 = nn.BatchNorm2d(in_planes)
        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, bias=True)
#         self.dropout = nn.Dropout(p=dropout_rate)
#         self.bn2 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=True),
            )

    def forward(self, x):
#         out = self.dropout(self.conv1(F.relu(self.bn1(x))))
#         out = self.conv1(F.leaky_relu(self.bn1(x), 0.1))
#         out = self.conv2(F.leaky_relu(self.bn2(out), 0.1))
        out = self.conv1(F.leaky_relu((x), 0.1))
        out = self.conv2(F.leaky_relu((out), 0.1))
        out += self.shortcut(x)

        return out

class Wide_ResNet(ReparamModule):
    def __init__(self, depth, widen_factor, dropout_rate, num_classes):
        super(Wide_ResNet, self).__init__()
        self.in_planes = 16

        assert ((depth-4)%6 ==0), 'Wide-resnet depth should be 6n+4'
        n = int((depth-4)/6)
        k = widen_factor

#         print('| Wide-Resnet %dx%d' %(depth, k))
        nStages = [16, 16*k, 32*k, 64*k]

        self.conv1 = conv3x3(3,nStages[0])
        self.layer1 = self._wide_layer(wide_basic, nStages[1], n, dropout_rate, stride=1)
        self.layer2 = self._wide_layer(wide_basic, nStages[2], n, dropout_rate, stride=2)
        self.layer3 = self._wide_layer(wide_basic, nStages[3], n, dropout_rate, stride=2)
#         self.bn1 = nn.BatchNorm2d(nStages[3], momentum=0.9)
        self.linear = nn.Linear(nStages[3], num_classes)

    def _wide_layer(self, block, planes, num_blocks, dropout_rate, stride):
        
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        
        for stride in strides:
            layers.append(block(self.in_planes, planes, dropout_rate, stride))
            self.in_planes = planes

        return nn.Sequential(*layers)

    def forward(self, x):
        out = self.conv1(x)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
#         out = F.leaky_relu(self.bn1(out), 0.1)
        out = F.leaky_relu((out), 0.1)
        out = F.avg_pool2d(out, 8)
        out = out.view(out.size(0), -1)
        out = self.linear(out)

        return out
    
############### VGG
class VGG(ReparamModule):

    def __init__(self, num_classes = 10):
        super(VGG, self).__init__()
        
        cfg =  [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M']
        layers = []
        in_channels = 3
        for v in cfg:
            if v == 'M':
                layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
#                 layers += [nn.Conv2d(in_channels, in_channels, kernel_size = 2, stride = 2, padding = 0)]
            else:
                conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)                
                layers += [conv2d, nn.ReLU(inplace=True)]
                in_channels = v
        self.features = nn.Sequential(*layers)
#         self.avgpool = nn.AdaptiveAvgPool2d((7, 7))
        
        self.classifier = nn.Sequential(    
            nn.Linear(512, 64),
            nn.ReLU(True),
#             nn.Dropout(),
            nn.Linear(64, 64),
            nn.ReLU(True),
#             nn.Dropout(),
            nn.Linear(64, num_classes),
        )

    
    def forward(self, x):
        x = self.features(x)
#         x = self.avgpool(x)
#         x = torch.flatten(x, 1)
        x = x.view(x.size(0), -1)
        x = self.classifier(x)
        return x

    
######## ResNet    
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=dilation, groups=groups, bias=False, dilation=dilation)


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


class BasicBlock(ReparamModule):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(BasicBlock, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64:
            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
        if dilation > 1:
            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(inplanes, planes, stride)
#         self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
#         self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
#         out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
#         out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


class Bottleneck(ReparamModule):
    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
    # This variant is also known as ResNet V1.5 and improves accuracy according to
    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.

    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.)) * groups
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width)
#         self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation)
#         self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion)
#         self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
#         out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
#         out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
#         out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

    
class ResNet(ReparamModule):

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False,
                 groups=1, width_per_group=64, replace_stride_with_dilation=None,
                 norm_layer=None):
        super(ResNet, self).__init__()
        if block == 'BasicBlock':
            block = BasicBlock
        else:
            block = Bottleneck
            
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if replace_stride_with_dilation is None:
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError("replace_stride_with_dilation should be None "
                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
        self.groups = groups
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
                               bias=False)
#         self.bn1 = norm_layer(self.inplanes)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
                                       dilate=replace_stride_with_dilation[0])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
                                       dilate=replace_stride_with_dilation[1])
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
                                       dilate=replace_stride_with_dilation[2])
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

#         for m in self.modules():
#             if isinstance(m, nn.Conv2d):
#                 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
#             elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
#                 nn.init.constant_(m.weight, 1)
#                 nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
#         if zero_init_residual:
#             for m in self.modules():
#                 if isinstance(m, Bottleneck):
#                     nn.init.constant_(m.bn3.weight, 0)
#                 elif isinstance(m, BasicBlock):
#                     nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
               # norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def _forward_impl(self, x):
        # See note [TorchScript super()]
        x = self.conv1(x)
#         x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)

        return x

    def forward(self, x):
        return self._forward_impl(x)


def resnet(block, layers):
    model = ResNet(block, layers)

    return model
    
net = resnet(BasicBlock, [3, 4, 6, 3])

def weights_init(m):
    def init_func(m):
        classname = m.__class__.__name__
        if classname.startswith('Conv') or classname == 'Linear':
            if getattr(m, 'bias', None) is not None:
                nn.init.constant_(m.bias, 0.0)
            if getattr(m, 'weight', None) is not None:
                if classname == 'Linear':           
                    nn.init.xavier_normal_(m.weight)
                if classname.startswith('Conv'):
                    nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')

        elif 'Norm' in classname:
            if getattr(m, 'weight', None) is not None:
                m.weight.data.fill_(1)
            if getattr(m, 'bias', None) is not None:
                m.bias.data.zero_()
    m.apply(init_func)
    return(m)