code
stringlengths
17
6.64M
class NoiseLayer(nn.Module): def __init__(self, in_planes, out_planes, level): super(NoiseLayer, self).__init__() self.noise = nn.Parameter(torch.Tensor(0), requires_grad=False).to(device) self.level = level self.layers = nn.Sequential(nn.ReLU(True), nn.BatchNorm2d(in_planes), nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1)) def forward(self, x): if (self.noise.numel() == 0): self.noise.resize_(x.data[0].shape).uniform_() self.noise = (((2 * self.noise) - 1) * self.level) y = torch.add(x, self.noise) z = self.layers(y) return z
class NoiseBasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1, shortcut=None, level=0.2): super(NoiseBasicBlock, self).__init__() self.layers = nn.Sequential(NoiseLayer(in_planes, planes, level), nn.MaxPool2d(stride, stride), nn.BatchNorm2d(planes), nn.ReLU(True), NoiseLayer(planes, planes, level), nn.BatchNorm2d(planes)) self.shortcut = shortcut def forward(self, x): residual = x y = self.layers(x) if self.shortcut: residual = self.shortcut(x) y += residual y = F.relu(y) return y
class NoiseBottleneck(nn.Module): expansion = 4 def __init__(self, in_planes, planes, stride=1, shortcut=None, level=0.2): super(NoiseBottleneck, self).__init__() self.layers = nn.Sequential(nn.Conv2d(in_planes, planes, kernel_size=1, bias=False), nn.BatchNorm2d(planes), nn.ReLU(True), NoiseLayer(planes, planes, level), nn.MaxPool2d(stride, stride), nn.BatchNorm2d(planes), nn.ReLU(True), nn.Conv2d(planes, (planes * 4), kernel_size=1, bias=False), nn.BatchNorm2d((planes * 4))) self.shortcut = shortcut def forward(self, x): residual = x y = self.layers(x) if self.shortcut: residual = self.shortcut(x) y += residual y = F.relu(y) return y
class NoiseResNet(nn.Module): def __init__(self, block, nblocks, nchannels, nfilters, nclasses, pool, level): super(NoiseResNet, self).__init__() self.in_planes = nfilters self.pre_layers = nn.Sequential(nn.Conv2d(nchannels, nfilters, kernel_size=7, stride=2, padding=3, bias=False), nn.BatchNorm2d(nfilters), nn.ReLU(True), nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) self.layer1 = self._make_layer(block, (1 * nfilters), nblocks[0], level=level) self.layer2 = self._make_layer(block, (2 * nfilters), nblocks[1], stride=2, level=level) self.layer3 = self._make_layer(block, (4 * nfilters), nblocks[2], stride=2, level=level) self.layer4 = self._make_layer(block, (8 * nfilters), nblocks[3], stride=2, level=level) self.avgpool = nn.AvgPool2d(pool, stride=1) self.linear = nn.Linear(((8 * nfilters) * block.expansion), nclasses) def _make_layer(self, block, planes, nblocks, stride=1, level=0.2): shortcut = None if ((stride != 1) or (self.in_planes != (planes * block.expansion))): shortcut = nn.Sequential(nn.Conv2d(self.in_planes, (planes * block.expansion), kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d((planes * block.expansion))) layers = [] layers.append(block(self.in_planes, planes, stride, shortcut, level=level)) self.in_planes = (planes * block.expansion) for i in range(1, nblocks): layers.append(block(self.in_planes, planes, level=level)) return nn.Sequential(*layers) def forward(self, x): x1 = self.pre_layers(x) x2 = self.layer1(x1) x3 = self.layer2(x2) x4 = self.layer3(x3) x5 = self.layer4(x4) x6 = self.avgpool(x5) x7 = x6.view(x6.size(0), (- 1)) x8 = self.linear(x7) return x8
def noiseresnet18(nchannels, nfilters, nclasses, pool=7, level=0.1): return NoiseResNet(NoiseBasicBlock, [2, 2, 2, 2], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses, pool=pool, level=level)
def noiseresnet34(nchannels, nfilters, nclasses, pool=7, level=0.1): return NoiseResNet(NoiseBasicBlock, [3, 4, 6, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses, pool=pool, level=level)
def noiseresnet50(nchannels, nfilters, nclasses, pool=7, level=0.1): return NoiseResNet(NoiseBottleneck, [3, 4, 6, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses, pool=pool, level=level)
def noiseresnet101(nchannels, nfilters, nclasses, pool=7, level=0.1): return NoiseResNet(NoiseBottleneck, [3, 4, 23, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses, pool=pool, level=level)
def noiseresnet152(nchannels, nfilters, nclasses, pool=7, level=0.1): return NoiseResNet(NoiseBottleneck, [3, 8, 36, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses, pool=pool, level=level)
class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(3, 6, 5) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(6, 16, 5) self.fc1 = nn.Linear(((16 * 5) * 5), 120) self.fc2 = nn.Linear(120, 84) self.fc3 = nn.Linear(84, 10) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) x = self.pool(F.relu(self.conv2(x))) x = x.view((- 1), ((16 * 5) * 5)) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) return x
def conv3x3(in_planes, out_planes, stride=1): '3x3 convolution with padding' return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = 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): residual = self.downsample(x) out += residual out = self.relu(out) return out
class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, (planes * 4), kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d((planes * 4)) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = 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): residual = self.downsample(x) out += residual out = self.relu(out) return out
class ResNet(nn.Module): def __init__(self, block, layers, nchannels, nfilters, nclasses=1000): self.inplanes = nfilters super(ResNet, self).__init__() self.conv1 = nn.Conv2d(nchannels, nfilters, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(nfilters) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, nfilters, layers[0]) self.layer2 = self._make_layer(block, (2 * nfilters), layers[1], stride=2) self.layer3 = self._make_layer(block, (4 * nfilters), layers[2], stride=2) self.layer4 = self._make_layer(block, (8 * nfilters), layers[3], stride=2) self.avgpool = nn.AvgPool2d(7, stride=1) self.fc = nn.Linear(((8 * nfilters) * block.expansion), nclasses) for m in self.modules(): if isinstance(m, nn.Conv2d): n = ((m.kernel_size[0] * m.kernel_size[1]) * m.out_channels) m.weight.data.normal_(0, math.sqrt((2.0 / n))) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if ((stride != 1) or (self.inplanes != (planes * block.expansion))): downsample = nn.Sequential(nn.Conv2d(self.inplanes, (planes * block.expansion), kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d((planes * block.expansion))) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = (planes * block.expansion) for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x0): x = self.conv1(x0) x = self.bn1(x) x = self.relu(x) x1 = self.maxpool(x) x2 = self.layer1(x1) x3 = self.layer2(x2) x4 = self.layer3(x3) x5 = self.layer4(x4) x = self.avgpool(x5) x = x.view(x.size(0), (- 1)) x6 = self.fc(x) return [x0, x1, x2, x3, x4, x5]
def resnet18(nchannels, nfilters, nclasses): return ResNet(BasicBlock, [2, 2, 2, 2], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses)
def resnet34(nchannels, nfilters, nclasses): return ResNet(BasicBlock, [3, 4, 6, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses)
def resnet50(nchannels, nfilters, nclasses): return ResNet(Bottleneck, [3, 4, 6, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses)
def resnet101(nchannels, nfilters, nclasses): return ResNet(Bottleneck, [3, 4, 23, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses)
def resnet152(nchannels, nfilters, nclasses): return ResNet(Bottleneck, [3, 8, 36, 3], nchannels=nchannels, nfilters=nfilters, nclasses=nclasses)
class Image(): def __init__(self, path, ext='png'): if (os.path.isdir(path) == False): os.makedirs(path) self.path = path self.names = [] self.ext = ext self.iteration = 1 self.num = 0 def register(self, modules): self.num = (self.num + len(modules)) for tmp in modules: self.names.append(tmp) def update(self, modules): for i in range(self.num): name = os.path.join(self.path, ('%s_%03d.png' % (self.names[i], self.iteration))) nrow = math.ceil(math.sqrt(modules[i].size(0))) vutils.save_image(modules[i], name, nrow=nrow, padding=0, normalize=True, scale_each=True) self.iteration = (self.iteration + 1)
class Logger(): def __init__(self, path, filename): self.num = 0 if (os.path.isdir(path) == False): os.makedirs(path) self.filename = os.path.join(path, filename) self.fid = open(self.filename, 'w') self.fid.close() def register(self, modules): self.num = (self.num + len(modules)) tmpstr = '' for tmp in modules: tmpstr = ((tmpstr + tmp) + '\t') tmpstr = (tmpstr + '\n') self.fid = open(self.filename, 'a') self.fid.write(tmpstr) self.fid.close() def update(self, modules): tmpstr = '' for tmp in modules: tmpstr = ((tmpstr + ('%.4f' % modules[tmp])) + '\t') tmpstr = (tmpstr + '\n') self.fid = open(self.filename, 'a') self.fid.write(tmpstr) self.fid.close()
class Monitor(): def __init__(self, smoothing=True, smoothness=0.7): self.keys = [] self.losses = {} self.smoothing = smoothing self.smoothness = smoothness self.num = 0 def register(self, modules): for m in modules: self.keys.append(m) self.losses[m] = 0 def reset(self): self.num = 0 for (key, value) in self.losses.items(): value = 0 def update(self, modules, batch_size): if (self.smoothing == False): for (key, value) in modules.items(): self.losses[key] = (((self.losses[key] * self.num) + (value * batch_size)) / (self.num + batch_size)) if (self.smoothing == True): for (key, value) in modules.items(): temp = (((self.losses[key] * self.num) + (value * batch_size)) / (self.num + batch_size)) self.losses[key] = ((self.losses[key] * self.smoothness) + (value * (1 - self.smoothness))) self.num += batch_size def getvalues(self, key=None): if (key != None): return self.losses[key] if (key == None): return OrderedDict([(key, self.losses[key]) for key in self.keys])
class Visualizer(): def __init__(self, port, title): self.keys = [] self.values = {} self.viz = visdom.Visdom(port=port) self.iteration = 0 self.title = title def register(self, modules): for key in modules: self.keys.append(key) self.values[key] = {} self.values[key]['dtype'] = modules[key]['dtype'] self.values[key]['vtype'] = modules[key]['vtype'] if (modules[key]['vtype'] == 'plot'): self.values[key]['value'] = [] self.values[key]['win'] = self.viz.line(X=np.array([0]), Y=np.array([0]), opts=dict(title=self.title, xlabel='Epoch', ylabel=key)) elif (modules[key]['vtype'] == 'image'): self.values[key]['value'] = None elif (modules[key]['vtype'] == 'images'): self.values[key]['value'] = None else: sys.exit('Data type not supported, please update the visualizer plugin and rerun !!') def update(self, modules): for key in modules: if (self.values[key]['dtype'] == 'scalar'): self.values[key]['value'].append(modules[key]) elif (self.values[key]['dtype'] == 'image'): self.values[key]['value'] = modules[key] elif (self.values[key]['dtype'] == 'images'): self.values[key]['value'] = modules[key] else: sys.exit('Data type not supported, please update the visualizer plugin and rerun !!') for key in self.keys: if (self.values[key]['vtype'] == 'plot'): self.viz.updateTrace(X=np.array([self.iteration]), Y=np.array([self.values[key]['value'][(- 1)]]), win=self.values[key]['win']) elif (self.values[key]['vtype'] == 'image'): temp = self.values[key]['value'].numpy() for i in range(temp.shape[0]): temp[i] = (temp[i] - temp[i].min()) temp[i] = (temp[i] / temp[i].max()) if (self.iteration == 0): self.values[key]['win'] = self.viz.image(temp, opts=dict(title=key, caption=self.iteration)) else: self.viz.image(temp, opts=dict(title=key, caption=self.iteration), win=self.values[key]['win']) elif (self.values[key]['vtype'] == 'images'): temp = self.values[key]['value'].numpy() for i in range(temp.shape[0]): for j in range(temp.shape[1]): temp[i][j] = (temp[i][j] - temp[i][j].min()) temp[i][j] = (temp[i][j] / temp[i][j].max()) if (self.iteration == 0): self.values[key]['win'] = self.viz.images(temp, opts=dict(title=key, caption=self.iteration)) else: self.viz.images(temp, opts=dict(title=key, caption=self.iteration), win=self.values[key]['win']) else: sys.exit('Visualization type not supported, please update the visualizer plugin and rerun !!') self.iteration = (self.iteration + 1)
class Trainer(): def __init__(self, args, model, criterion): self.args = args self.model = model self.criterion = criterion self.port = args.port self.dir_save = args.save self.cuda = args.cuda self.nepochs = args.nepochs self.nclasses = args.nclasses self.nchannels = args.nchannels self.batch_size = args.batch_size self.resolution_high = args.resolution_high self.resolution_wide = args.resolution_wide self.lr = args.learning_rate self.momentum = args.momentum self.adam_beta1 = args.adam_beta1 self.adam_beta2 = args.adam_beta2 self.weight_decay = args.weight_decay self.optim_method = args.optim_method self.dataset_train_name = args.dataset_train parameters = filter((lambda p: p.requires_grad), model.parameters()) if (self.optim_method == 'Adam'): self.optimizer = optim.Adam(parameters, lr=self.lr, betas=(self.adam_beta1, self.adam_beta2), weight_decay=self.weight_decay) elif (self.optim_method == 'RMSprop'): self.optimizer = optim.RMSprop(parameters, lr=self.lr, momentum=self.momentum, weight_decay=self.weight_decay) elif (self.optim_method == 'SGD'): self.optimizer = optim.SGD(parameters, lr=self.lr, momentum=self.momentum, weight_decay=self.weight_decay, nesterov=True) else: raise Exception('Unknown Optimization Method') self.label = torch.zeros(self.batch_size).long() self.input = torch.zeros(self.batch_size, self.nchannels, self.resolution_high, self.resolution_wide) if args.cuda: self.label = self.label.cuda() self.input = self.input.cuda() self.input = Variable(self.input) self.label = Variable(self.label) self.log_loss_train = plugins.Logger(args.logs, 'TrainLogger.txt') self.params_loss_train = ['Loss', 'Accuracy'] self.log_loss_train.register(self.params_loss_train) self.log_loss_test = plugins.Logger(args.logs, 'TestLogger.txt') self.params_loss_test = ['Loss', 'Accuracy'] self.log_loss_test.register(self.params_loss_test) self.monitor_train = plugins.Monitor() self.params_monitor_train = ['Loss', 'Accuracy'] self.monitor_train.register(self.params_monitor_train) self.monitor_test = plugins.Monitor() self.params_monitor_test = ['Loss', 'Accuracy'] self.monitor_test.register(self.params_monitor_test) self.visualizer_train = plugins.Visualizer(self.port, 'Train') self.params_visualizer_train = {'Loss': {'dtype': 'scalar', 'vtype': 'plot'}, 'Accuracy': {'dtype': 'scalar', 'vtype': 'plot'}} self.visualizer_train.register(self.params_visualizer_train) self.visualizer_test = plugins.Visualizer(self.port, 'Test') self.params_visualizer_test = {'Loss': {'dtype': 'scalar', 'vtype': 'plot'}, 'Accuracy': {'dtype': 'scalar', 'vtype': 'plot'}} self.visualizer_test.register(self.params_visualizer_test) self.print_train = '[%d/%d][%d/%d] ' for item in self.params_loss_train: self.print_train = ((self.print_train + item) + ' %.4f ') self.print_test = '[%d/%d][%d/%d] ' for item in self.params_loss_test: self.print_test = ((self.print_test + item) + ' %.4f ') self.evalmodules = [] self.giterations = 0 self.losses_test = {} self.losses_train = {} def learning_rate(self, epoch): if (self.dataset_train_name == 'CIFAR10'): return (self.lr * (((0.1 ** int((epoch >= 60))) * (0.1 ** int((epoch >= 90)))) * (0.1 ** int((epoch >= 120))))) elif (self.dataset_train_name == 'CIFAR100'): return (self.lr * (((0.1 ** int((epoch >= 80))) * (0.1 ** int((epoch >= 120)))) * (0.1 ** int((epoch >= 160))))) elif (self.dataset_train_name == 'MNIST'): return (self.lr * (((0.1 ** int((epoch >= 80))) * (0.1 ** int((epoch >= 120)))) * (0.1 ** int((epoch >= 160))))) elif (self.dataset_train_name == 'FRGC'): return (self.lr * (((0.1 ** int((epoch >= 80))) * (0.1 ** int((epoch >= 120)))) * (0.1 ** int((epoch >= 160))))) elif (self.dataset_train_name == 'ImageNet'): decay = math.floor(((epoch - 1) / 30)) return (self.lr * math.pow(0.1, decay)) def get_optimizer(self, epoch, optimizer): lr = self.learning_rate(epoch) for param_group in optimizer.param_groups: param_group['lr'] = lr return optimizer def model_eval(self): self.model.eval() for m in self.model.modules(): for i in range(len(self.evalmodules)): if isinstance(m, self.evalmodules[i]): m.train() def model_train(self): self.model.train() def train(self, epoch, dataloader): self.monitor_train.reset() data_iter = iter(dataloader) self.input.volatile = False self.label.volatile = False self.optimizer = self.get_optimizer((epoch + 1), self.optimizer) self.model_train() i = 0 while (i < len(dataloader)): (input, label) = data_iter.next() i += 1 batch_size = input.size(0) if (batch_size == self.batch_size): self.input.data.resize_(input.size()).copy_(input) self.label.data.resize_(label.size()).copy_(label) output = self.model(self.input) loss = self.criterion(output, self.label) self.optimizer.zero_grad() loss.backward() self.optimizer.step() pred = output.data.max(1)[1] acc = (float((pred.eq(self.label.data).cpu().sum() * 100.0)) / float(batch_size)) self.losses_train['Accuracy'] = float(acc) self.losses_train['Loss'] = float(loss.data[0]) self.monitor_train.update(self.losses_train, batch_size) print((self.print_train % tuple(([epoch, self.nepochs, i, len(dataloader)] + [self.losses_train[key] for key in self.params_monitor_train])))) loss = self.monitor_train.getvalues() self.log_loss_train.update(loss) self.visualizer_train.update(loss) return self.monitor_train.getvalues('Accuracy') def test(self, epoch, dataloader): self.monitor_test.reset() data_iter = iter(dataloader) self.input.volatile = True self.label.volatile = True self.model_eval() i = 0 while (i < len(dataloader)): (input, label) = data_iter.next() i += 1 batch_size = input.size(0) if (batch_size == self.batch_size): self.input.data.resize_(input.size()).copy_(input) self.label.data.resize_(label.size()).copy_(label) self.model.zero_grad() output = self.model(self.input) loss = self.criterion(output, self.label) pred = output.data.max(1)[1] acc = (float((pred.eq(self.label.data).cpu().sum() * 100.0)) / float(batch_size)) self.losses_test['Accuracy'] = float(acc) self.losses_test['Loss'] = float(loss.data[0]) self.monitor_test.update(self.losses_test, batch_size) print((self.print_test % tuple(([epoch, self.nepochs, i, len(dataloader)] + [self.losses_test[key] for key in self.params_monitor_test])))) loss = self.monitor_test.getvalues() self.log_loss_test.update(loss) self.visualizer_test.update(loss) return self.monitor_test.getvalues('Accuracy')
def readtextfile(filename): with open(filename) as f: content = f.readlines() f.close() return content
def writetextfile(data, filename): with open(filename, 'w') as f: f.writelines(data) f.close()
def delete_file(filename): if (os.path.isfile(filename) == True): os.remove(filename)
def eformat(f, prec, exp_digits): s = ('%.*e' % (prec, f)) (mantissa, exp) = s.split('e') return ('%se%+0*d' % (mantissa, (exp_digits + 1), int(exp)))
def saveargs(args): path = args.logs if (os.path.isdir(path) == False): os.makedirs(path) with open(os.path.join(path, 'args.txt'), 'w') as f: for arg in vars(args): f.write((((arg + ' ') + str(getattr(args, arg))) + '\n'))
class Dataloader(): def __init__(self, args, input_size): self.args = args self.dataset_test_name = args.dataset_test self.dataset_train_name = args.dataset_train self.input_size = input_size if (self.dataset_train_name == 'LSUN'): self.dataset_train = getattr(datasets, self.dataset_train_name)(db_path=args.dataroot, classes=['bedroom_train'], transform=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif ((self.dataset_train_name == 'CIFAR10') or (self.dataset_train_name == 'CIFAR100')): self.dataset_train = getattr(datasets, self.dataset_train_name)(root=self.args.dataroot, train=True, download=True, transform=transforms.Compose([transforms.RandomCrop(self.input_size, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))])) elif ((self.dataset_train_name == 'CocoCaption') or (self.dataset_train_name == 'CocoDetection')): self.dataset_train = getattr(datasets, self.dataset_train_name)(root=self.args.dataroot, train=True, download=True, transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif ((self.dataset_train_name == 'STL10') or (self.dataset_train_name == 'SVHN')): self.dataset_train = getattr(datasets, self.dataset_train_name)(root=self.args.dataroot, split='train', download=True, transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_train_name == 'MNIST'): self.dataset_train = getattr(datasets, self.dataset_train_name)(root=self.args.dataroot, train=True, download=True, transform=transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])) elif (self.dataset_train_name == 'ImageNet'): normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) self.dataset_train = datasets.ImageFolder(root=os.path.join(self.args.dataroot, self.args.input_filename_train), transform=transforms.Compose([transforms.RandomSizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize])) elif (self.dataset_train_name == 'FRGC'): self.dataset_train = datasets.ImageFolder(root=(self.args.dataroot + self.args.input_filename_train), transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_train_name == 'Folder'): self.dataset_train = datasets.ImageFolder(root=(self.args.dataroot + self.args.input_filename_train), transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_train_name == 'FileList'): self.dataset_train = datasets.FileList(self.input_filename_train, self.label_filename_train, self.split_train, self.split_test, train=True, transform_train=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), transform_test=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), loader_input=self.loader_input, loader_label=self.loader_label) elif (self.dataset_train_name == 'FolderList'): self.dataset_train = datasets.FileList(self.input_filename_train, self.label_filename_train, self.split_train, self.split_test, train=True, transform_train=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), transform_test=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), loader_input=self.loader_input, loader_label=self.loader_label) else: raise Exception('Unknown Dataset') if (self.dataset_test_name == 'LSUN'): self.dataset_test = getattr(datasets, self.dataset_test_name)(db_path=args.dataroot, classes=['bedroom_val'], transform=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif ((self.dataset_test_name == 'CIFAR10') or (self.dataset_test_name == 'CIFAR100')): self.dataset_test = getattr(datasets, self.dataset_test_name)(root=self.args.dataroot, train=False, download=True, transform=transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))])) elif ((self.dataset_test_name == 'CocoCaption') or (self.dataset_test_name == 'CocoDetection')): self.dataset_test = getattr(datasets, self.dataset_test_name)(root=self.args.dataroot, train=False, download=True, transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif ((self.dataset_test_name == 'STL10') or (self.dataset_test_name == 'SVHN')): self.dataset_test = getattr(datasets, self.dataset_test_name)(root=self.args.dataroot, split='test', download=True, transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_test_name == 'MNIST'): self.dataset_test = getattr(datasets, self.dataset_test_name)(root=self.args.dataroot, train=False, download=True, transform=transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])) elif (self.dataset_test_name == 'ImageNet'): normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) self.dataset_test = datasets.ImageFolder(root=os.path.join(self.args.dataroot, self.args.input_filename_test), transform=transforms.Compose([transforms.Scale(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize])) elif (self.dataset_test_name == 'FRGC'): self.dataset_test = datasets.ImageFolder(root=(self.args.dataroot + self.args.input_filename_test), transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_test_name == 'Folder'): self.dataset_test = datasets.ImageFolder(root=(self.args.dataroot + self.args.input_filename_test), transform=transforms.Compose([transforms.Scale(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) elif (self.dataset_test_name == 'FileList'): self.dataset_test = datasets.FileList(self.input_filename_test, self.label_filename_test, self.split_train, self.split_test, train=True, transform_train=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), loader_input=self.loader_input, loader_label=self.loader_label) elif (self.dataset_test_name == 'FolderList'): self.dataset_test = datasets.FileList(self.input_filename_test, self.label_filename_test, self.split_train, self.split_test, train=True, transform_train=transforms.Compose([transforms.Scale(self.input_size), transforms.CenterCrop(self.input_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]), loader_input=self.loader_input, loader_label=self.loader_label) else: raise Exception('Unknown Dataset') def create(self, flag=None): if (flag == 'Train'): dataloader_train = torch.utils.data.DataLoader(self.dataset_train, batch_size=self.args.batch_size, shuffle=True, num_workers=int(self.args.nthreads), pin_memory=True) return dataloader_train if (flag == 'Test'): dataloader_test = torch.utils.data.DataLoader(self.dataset_test, batch_size=self.args.batch_size, shuffle=False, num_workers=int(self.args.nthreads), pin_memory=True) return dataloader_test if (flag == None): dataloader_train = torch.utils.data.DataLoader(self.dataset_train, batch_size=self.args.batch_size, shuffle=True, num_workers=int(self.args.nthreads), pin_memory=True) dataloader_test = torch.utils.data.DataLoader(self.dataset_test, batch_size=self.args.batch_size, shuffle=False, num_workers=int(self.args.nthreads), pin_memory=True) return (dataloader_train, dataloader_test)
class FileList(data.Dataset): def __init__(self, ifile, lfile=None, split_train=1.0, split_test=0.0, train=True, transform_train=None, transform_test=None, loader_input=loaders.loader_image, loader_label=loaders.loader_torch): self.ifile = ifile self.lfile = lfile self.train = train self.split_test = split_test self.split_train = split_train self.transform_test = transform_test self.transform_train = transform_train self.loader_input = loader_input self.loader_label = loader_label if (loader_input == 'image'): self.loader_input = loaders.loader_image if (loader_input == 'torch'): self.loader_input = loaders.loader_torch if (loader_input == 'numpy'): self.loader_input = loaders.loader_numpy if (loader_label == 'image'): self.loader_label = loaders.loader_image if (loader_label == 'torch'): self.loader_label = loaders.loader_torch if (loader_label == 'numpy'): self.loader_label = loaders.loader_numpy if (ifile != None): imagelist = utils.readtextfile(ifile) imagelist = [x.rstrip('\n') for x in imagelist] else: imagelist = [] if (lfile != None): labellist = utils.readtextfile(lfile) labellist = [x.rstrip('\n') for x in labellist] else: labellist = [] if (len(imagelist) == len(labellist)): shuffle(imagelist, labellist) if ((len(imagelist) > 0) and (len(labellist) == 0)): shuffle(imagelist) if ((len(labellist) > 0) and (len(imagelist) == 0)): shuffle(labellist) if ((self.split_train < 1.0) & (self.split_train > 0.0)): if (len(imagelist) > 0): num = math.floor((self.split * len(imagelist))) self.images_train = imagelist[0:num] self.images_test = images[(num + 1):len(imagelist)] if (len(labellist) > 0): num = math.floor((self.split * len(labellist))) self.labels_train = labellist[0:num] self.labels_test = labellist[(num + 1):len(labellist)] elif (self.split_train == 1.0): if (len(imagelist) > 0): self.images_train = imagelist if (len(labellist) > 0): self.labels_train = labellist elif (self.split_test == 1.0): if (len(imagelist) > 0): self.images_test = imagelist if (len(labellist) > 0): self.labels_test = labellist def __len__(self): if (self.train == True): return len(self.images_train) if (self.train == False): return len(self.images_test) def __getitem__(self, index): input = {} if (self.train == True): if (len(self.images_train) > 0): path = self.images_train[index] input['inp'] = self.loader_input(path) if (len(self.labels_train) > 0): path = self.labels_train[index] input['tgt'] = self.loader_label(path) if (self.transform_train is not None): input = self.transform_train(input) image = input['inp'] label = input['tgt'] if (self.train == False): if (len(self.images_test) > 0): path = self.images_test[index] input['inp'] = self.loader_input(path) if (len(self.labels_test) > 0): path = self.labels_test[index] input['tgt'] = self.loader_label(path) if (self.transform_test is not None): input = self.transform_test(input) image = input['inp'] label = input['tgt'] return (image, label)
def is_image_file(filename): return any((filename.endswith(extension) for extension in IMG_EXTENSIONS))
def make_dataset(classlist, labellist=None): images = [] labels = [] classes = utils.readtextfile(ifile) classes = [x.rstrip('\n') for x in classes] classes.sort() for i in len(classes): for fname in os.listdir(classes[i]): if is_image_file(fname): label = {} label['class'] = os.path.split(classes[i]) images.append(fname) labels.append(label) if (labellist != None): labels = utils.readtextfile(ifile) labels = [x.rstrip('\n') for x in labels] labels.sort() for i in len(labels): for fname in os.listdir(labels[i]): if is_image_file(fname): labels.append(os.path.split(classes[i])) return (images, labels)
class FolderList(data.Dataset): def __init__(self, ifile, lfile=None, split_train=1.0, split_test=0.0, train=True, transform_train=None, transform_test=None, loader_input=loaders.loader_image, loader_label=loaders.loader_torch): (imagelist, labellist) = make_dataset(ifile, lfile) if (len(imagelist) == 0): raise RuntimeError('No images found') if (len(labellist) == 0): raise RuntimeError('No labels found') self.loader_input = loader_input self.loader_label = loader_label if (loader_input == 'image'): self.loader_input = loaders.loader_image if (loader_input == 'torch'): self.loader_input = loaders.loader_torch if (loader_input == 'numpy'): self.loader_input = loaders.loader_numpy if (loader_label == 'image'): self.loader_label = loaders.loader_image if (loader_label == 'torch'): self.loader_label = loaders.loader_torch if (loader_label == 'numpy'): self.loader_label = loaders.loader_numpy self.imagelist = imagelist self.labellist = labellist self.transform_test = transform_test self.transform_train = transform_train if (len(imagelist) == len(labellist)): shuffle(imagelist, labellist) if ((len(imagelist) > 0) and (len(labellist) == 0)): shuffle(imagelist) if ((len(labellist) > 0) and (len(imagelist) == 0)): shuffle(labellist) if ((args.split_train < 1.0) & (args.split_train > 0.0)): if (len(imagelist) > 0): num = math.floor((args.split * len(imagelist))) self.images_train = imagelist[0:num] self.images_test = images[(num + 1):len(imagelist)] if (len(labellist) > 0): num = math.floor((args.split * len(labellist))) self.labels_train = labellist[0:num] self.labels_test = labellist[(num + 1):len(labellist)] elif (args.split_train == 1.0): if (len(imagelist) > 0): self.images_train = imagelist if (len(labellist) > 0): self.labels_train = labellist elif (args.split_test == 1.0): if (len(imagelist) > 0): self.images_test = imagelist if (len(labellist) > 0): self.labels_test = labellist def __len__(self): if (self.train == True): return len(self.images_train) if (self.train == False): return len(self.images_test) def __getitem__(self, index): if (self.train == True): if (len(self.images_train) > 0): path = self.images_train[index] input['inp'] = self.loader_input(path) if (len(self.labels_train) > 0): path = self.labels_train[index] input['tgt'] = self.loader_label(path) if (self.transform_train is not None): input = self.transform_train(input) image = input['inp'] label = input['tgt'] if (self.train == False): if (len(self.images_test) > 0): path = self.images_test[index] input['inp'] = self.loader_input(path) if (len(self.labels_test) > 0): path = self.labels_test[index] input['tgt'] = self.loader_label(path) if (self.transform_test is not None): input = self.transform_test(input) image = input['inp'] label = input['tgt'] return (image, label)
def loader_image(path): return Image.open(path).convert('RGB')
def loader_torch(path): return torch.load(path)
def loader_numpy(path): return np.load(path)
class Model(): def __init__(self, args): self.cuda = torch.cuda.is_available() self.lr = args.learning_rate self.dataset_train_name = args.dataset_train self.nfilters = args.nfilters self.batch_size = args.batch_size self.level = args.level self.net_type = args.net_type self.nmasks = args.nmasks self.unique_masks = args.unique_masks self.filter_size = args.filter_size self.first_filter_size = args.first_filter_size self.scale_noise = args.scale_noise self.noise_type = args.noise_type self.act = args.act self.use_act = args.use_act self.dropout = args.dropout self.train_masks = args.train_masks self.debug = args.debug self.pool_type = args.pool_type self.mix_maps = args.mix_maps if self.dataset_train_name.startswith('CIFAR'): self.input_size = 32 self.nclasses = 10 if (self.filter_size < 7): self.avgpool = 4 elif (self.filter_size == 7): self.avgpool = 1 elif self.dataset_train_name.startswith('MNIST'): self.nclasses = 10 self.input_size = 28 if (self.filter_size < 7): self.avgpool = 14 elif (self.filter_size == 7): self.avgpool = 7 self.model = getattr(models, self.net_type)(nfilters=self.nfilters, avgpool=self.avgpool, nclasses=self.nclasses, nmasks=self.nmasks, unique_masks=self.unique_masks, level=self.level, filter_size=self.filter_size, first_filter_size=self.first_filter_size, act=self.act, scale_noise=self.scale_noise, noise_type=self.noise_type, use_act=self.use_act, dropout=self.dropout, train_masks=self.train_masks, pool_type=self.pool_type, debug=self.debug, input_size=self.input_size, mix_maps=self.mix_maps) self.loss_fn = nn.CrossEntropyLoss() if self.cuda: self.model = self.model.cuda() self.loss_fn = self.loss_fn.cuda() parameters = filter((lambda p: p.requires_grad), self.model.parameters()) if (args.optim_method == 'Adam'): self.optimizer = optim.Adam(parameters, lr=self.lr, betas=(args.adam_beta1, args.adam_beta2), weight_decay=args.weight_decay) elif (args.optim_method == 'RMSprop'): self.optimizer = optim.RMSprop(parameters, lr=self.lr, momentum=args.momentum, weight_decay=args.weight_decay) elif (args.optim_method == 'SGD'): self.optimizer = optim.SGD(parameters, lr=self.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True) "\n # use this to set different learning rates for training noise masks and regular parameters:\n self.optimizer = optim.SGD([{'params': [param for name, param in self.model.named_parameters() if 'noise' not in name]},\n {'params': [param for name, param in self.model.named_parameters() if 'noise' in name], 'lr': self.lr * 10},\n ], lr=self.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True) #" else: raise Exception('Unknown Optimization Method') def learning_rate(self, epoch): if (self.dataset_train_name == 'CIFAR10'): new_lr = (self.lr * (((((0.2 ** int((epoch >= 150))) * (0.2 ** int((epoch >= 250)))) * (0.2 ** int((epoch >= 300)))) * (0.2 ** int((epoch >= 350)))) * (0.2 ** int((epoch >= 400))))) elif (self.dataset_train_name == 'CIFAR100'): new_lr = (self.lr * (((0.1 ** int((epoch >= 80))) * (0.1 ** int((epoch >= 120)))) * (0.1 ** int((epoch >= 160))))) elif (self.dataset_train_name == 'MNIST'): new_lr = (self.lr * (((0.2 ** int((epoch >= 30))) * (0.2 ** int((epoch >= 60)))) * (0.2 ** int((epoch >= 90))))) elif (self.dataset_train_name == 'FRGC'): new_lr = (self.lr * (((0.1 ** int((epoch >= 80))) * (0.1 ** int((epoch >= 120)))) * (0.1 ** int((epoch >= 160))))) elif (self.dataset_train_name == 'ImageNet'): decay = math.floor(((epoch - 1) / 30)) new_lr = (self.lr * math.pow(0.1, decay)) return new_lr def train(self, epoch, dataloader): self.model.train() lr = self.learning_rate((epoch + 1)) for param_group in self.optimizer.param_groups: param_group['lr'] = lr losses = [] accuracies = [] for (i, (input, label)) in enumerate(dataloader): if self.cuda: label = label.cuda() input = input.cuda() output = self.model(input) loss = self.loss_fn(output, label) if self.debug: print('\nBatch:', i) self.optimizer.zero_grad() loss.backward() self.optimizer.step() pred = output.data.max(1)[1] acc = ((pred.eq(label.data).cpu().sum() * 100.0) / self.batch_size) losses.append(loss.item()) accuracies.append(acc) return (np.mean(losses), np.mean(accuracies)) def test(self, dataloader): self.model.eval() losses = [] accuracies = [] with torch.no_grad(): for (i, (input, label)) in enumerate(dataloader): if self.cuda: label = label.cuda() input = input.cuda() output = self.model(input) loss = self.loss_fn(output, label) pred = output.data.max(1)[1] acc = ((pred.eq(label.data).cpu().sum() * 100.0) / self.batch_size) losses.append(loss.item()) accuracies.append(acc) return (np.mean(losses), np.mean(accuracies))
class PerturbLayerFirst(nn.Module): def __init__(self, in_channels=None, out_channels=None, nmasks=None, level=None, filter_size=None, debug=False, use_act=False, stride=1, act=None, unique_masks=False, mix_maps=None, train_masks=False, noise_type='uniform', input_size=None): super(PerturbLayerFirst, self).__init__() self.nmasks = nmasks self.unique_masks = unique_masks self.train_masks = train_masks self.level = level self.filter_size = filter_size self.use_act = use_act self.act = act_fn('sigmoid') self.debug = debug self.noise_type = noise_type self.in_channels = in_channels self.input_size = input_size self.mix_maps = mix_maps if (filter_size == 1): padding = 0 bias = True elif ((filter_size == 3) or (filter_size == 5)): padding = 1 bias = False elif (filter_size == 7): stride = 2 padding = 3 bias = False if (self.filter_size > 0): self.noise = None self.layers = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size=filter_size, padding=padding, stride=stride, bias=bias), nn.BatchNorm2d(out_channels), self.act) else: noise_channels = (in_channels if self.unique_masks else 1) shape = (1, noise_channels, self.nmasks, input_size, input_size) self.noise = nn.Parameter(torch.Tensor(*shape), requires_grad=self.train_masks) if (noise_type == 'uniform'): self.noise.data.uniform_((- 1), 1) elif (self.noise_type == 'normal'): self.noise.data.normal_() else: print('\n\nNoise type {} is not supported / understood\n\n'.format(self.noise_type)) if (nmasks != 1): if ((out_channels % in_channels) != 0): print('\n\n\nnfilters must be divisible by 3 if using multiple noise masks per input channel\n\n\n') groups = in_channels else: groups = 1 self.layers = nn.Sequential(nn.BatchNorm2d((in_channels * self.nmasks)), self.act, nn.Conv2d((in_channels * self.nmasks), out_channels, kernel_size=1, stride=1, groups=groups), nn.BatchNorm2d(out_channels), self.act) if self.mix_maps: self.mix_layers = nn.Sequential(nn.Conv2d(out_channels, out_channels, kernel_size=1, stride=1, groups=1), nn.BatchNorm2d(out_channels), self.act) def forward(self, x): if (self.filter_size > 0): return self.layers(x) else: y = torch.add(x.unsqueeze(2), (self.noise * self.level)) if self.debug: print_values(x, self.noise, y, self.unique_masks) y = y.view((- 1), (self.in_channels * self.nmasks), self.input_size, self.input_size) y = self.layers(y) if self.mix_maps: y = self.mix_layers(y) return y
class PerturbLayer(nn.Module): def __init__(self, in_channels=None, out_channels=None, nmasks=None, level=None, filter_size=None, debug=False, use_act=False, stride=1, act=None, unique_masks=False, mix_maps=None, train_masks=False, noise_type='uniform', input_size=None): super(PerturbLayer, self).__init__() self.nmasks = nmasks self.unique_masks = unique_masks self.train_masks = train_masks self.level = level self.filter_size = filter_size self.use_act = use_act self.act = act_fn(act) self.debug = debug self.noise_type = noise_type self.in_channels = in_channels self.input_size = input_size self.mix_maps = mix_maps if (filter_size == 1): padding = 0 bias = True elif ((filter_size == 3) or (filter_size == 5)): padding = 1 bias = False elif (filter_size == 7): stride = 2 padding = 3 bias = False if (self.filter_size > 0): self.noise = None self.layers = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size=filter_size, padding=padding, stride=stride, bias=bias), nn.BatchNorm2d(out_channels), self.act) else: noise_channels = (in_channels if self.unique_masks else 1) shape = (1, noise_channels, self.nmasks, input_size, input_size) self.noise = nn.Parameter(torch.Tensor(*shape), requires_grad=self.train_masks) if (noise_type == 'uniform'): self.noise.data.uniform_((- 1), 1) elif (self.noise_type == 'normal'): self.noise.data.normal_() else: print('\n\nNoise type {} is not supported / understood\n\n'.format(self.noise_type)) if (nmasks != 1): if ((out_channels % in_channels) != 0): print('\n\n\nnfilters must be divisible by 3 if using multiple noise masks per input channel\n\n\n') groups = in_channels else: groups = 1 self.layers = nn.Sequential(nn.Conv2d((in_channels * self.nmasks), out_channels, kernel_size=1, stride=1, groups=groups), nn.BatchNorm2d(out_channels), self.act) if self.mix_maps: self.mix_layers = nn.Sequential(nn.Conv2d(out_channels, out_channels, kernel_size=1, stride=1, groups=1), nn.BatchNorm2d(out_channels), self.act) def forward(self, x): if (self.filter_size > 0): return self.layers(x) else: y = torch.add(x.unsqueeze(2), (self.noise * self.level)) if self.debug: print_values(x, self.noise, y, self.unique_masks) if self.use_act: y = self.act(y) y = y.view((- 1), (self.in_channels * self.nmasks), self.input_size, self.input_size) y = self.layers(y) if self.mix_maps: y = self.mix_layers(y) return y
class PerturbBasicBlock(nn.Module): expansion = 1 def __init__(self, in_channels=None, out_channels=None, stride=1, shortcut=None, nmasks=None, train_masks=False, level=None, use_act=False, filter_size=None, act=None, unique_masks=False, noise_type=None, input_size=None, pool_type=None, mix_maps=None): super(PerturbBasicBlock, self).__init__() self.shortcut = shortcut if (pool_type == 'max'): pool = nn.MaxPool2d elif (pool_type == 'avg'): pool = nn.AvgPool2d else: print('\n\nPool Type {} is not supported/understood\n\n'.format(pool_type)) return self.layers = nn.Sequential(PerturbLayer(in_channels=in_channels, out_channels=out_channels, nmasks=nmasks, input_size=input_size, level=level, filter_size=filter_size, use_act=use_act, train_masks=train_masks, act=act, unique_masks=unique_masks, noise_type=noise_type, mix_maps=mix_maps), pool(stride, stride), PerturbLayer(in_channels=out_channels, out_channels=out_channels, nmasks=nmasks, input_size=(input_size // stride), level=level, filter_size=filter_size, use_act=use_act, train_masks=train_masks, act=act, unique_masks=unique_masks, noise_type=noise_type, mix_maps=mix_maps)) def forward(self, x): residual = x y = self.layers(x) if self.shortcut: residual = self.shortcut(x) y += residual y = F.relu(y) return y
class PerturbResNet(nn.Module): def __init__(self, block, nblocks=None, avgpool=None, nfilters=None, nclasses=None, nmasks=None, input_size=32, level=None, filter_size=None, first_filter_size=None, use_act=False, train_masks=False, mix_maps=None, act=None, scale_noise=1, unique_masks=False, debug=False, noise_type=None, pool_type=None): super(PerturbResNet, self).__init__() self.nfilters = nfilters self.unique_masks = unique_masks self.noise_type = noise_type self.train_masks = train_masks self.pool_type = pool_type self.mix_maps = mix_maps self.act = act_fn(act) layers = [PerturbLayerFirst(in_channels=3, out_channels=(3 * nfilters), nmasks=(nfilters * 5), level=((level * scale_noise) * 20), debug=debug, filter_size=first_filter_size, use_act=use_act, train_masks=train_masks, input_size=input_size, act=act, unique_masks=self.unique_masks, noise_type=self.noise_type, mix_maps=mix_maps)] if (first_filter_size == 7): layers.append(nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) self.pre_layers = nn.Sequential(*layers, nn.Conv2d(((self.nfilters * 3) * 1), self.nfilters, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(self.nfilters), self.act) self.layer1 = self._make_layer(block, (1 * nfilters), nblocks[0], stride=1, level=level, nmasks=nmasks, use_act=True, filter_size=filter_size, act=act, input_size=input_size) self.layer2 = self._make_layer(block, (2 * nfilters), nblocks[1], stride=2, level=level, nmasks=nmasks, use_act=True, filter_size=filter_size, act=act, input_size=input_size) self.layer3 = self._make_layer(block, (4 * nfilters), nblocks[2], stride=2, level=level, nmasks=nmasks, use_act=True, filter_size=filter_size, act=act, input_size=(input_size // 2)) self.layer4 = self._make_layer(block, (8 * nfilters), nblocks[3], stride=2, level=level, nmasks=nmasks, use_act=True, filter_size=filter_size, act=act, input_size=(input_size // 4)) self.avgpool = nn.AvgPool2d(avgpool, stride=1) self.linear = nn.Linear(((8 * nfilters) * block.expansion), nclasses) def _make_layer(self, block, out_channels, nblocks, stride=1, level=0.2, nmasks=None, use_act=False, filter_size=None, act=None, input_size=None): shortcut = None if ((stride != 1) or (self.nfilters != (out_channels * block.expansion))): shortcut = nn.Sequential(nn.Conv2d(self.nfilters, (out_channels * block.expansion), kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d((out_channels * block.expansion))) layers = [] layers.append(block(self.nfilters, out_channels, stride, shortcut, level=level, nmasks=nmasks, use_act=use_act, filter_size=filter_size, act=act, unique_masks=self.unique_masks, noise_type=self.noise_type, train_masks=self.train_masks, input_size=input_size, pool_type=self.pool_type, mix_maps=self.mix_maps)) self.nfilters = (out_channels * block.expansion) for i in range(1, nblocks): layers.append(block(self.nfilters, out_channels, level=level, nmasks=nmasks, use_act=use_act, train_masks=self.train_masks, filter_size=filter_size, act=act, unique_masks=self.unique_masks, noise_type=self.noise_type, input_size=(input_size // stride), pool_type=self.pool_type, mix_maps=self.mix_maps)) return nn.Sequential(*layers) def forward(self, x): x = self.pre_layers(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) x = x.view(x.size(0), (- 1)) x = self.linear(x) return x
class LeNet(nn.Module): def __init__(self, nfilters=None, nclasses=None, nmasks=None, level=None, filter_size=None, linear=128, input_size=28, debug=False, scale_noise=1, act='relu', use_act=False, first_filter_size=None, pool_type=None, dropout=None, unique_masks=False, train_masks=False, noise_type='uniform', mix_maps=None): super(LeNet, self).__init__() if (filter_size == 5): n = 5 else: n = 4 if (input_size == 32): first_channels = 3 elif (input_size == 28): first_channels = 1 if (pool_type == 'max'): pool = nn.MaxPool2d elif (pool_type == 'avg'): pool = nn.AvgPool2d else: print('\n\nPool Type {} is not supported/understood\n\n'.format(pool_type)) return self.linear1 = nn.Linear(((nfilters * n) * n), linear) self.linear2 = nn.Linear(linear, nclasses) self.dropout = nn.Dropout(p=dropout) self.act = act_fn(act) self.batch_norm = nn.BatchNorm1d(linear) self.first_layers = nn.Sequential(PerturbLayer(in_channels=first_channels, out_channels=nfilters, nmasks=nmasks, level=(level * scale_noise), filter_size=first_filter_size, use_act=use_act, act=act, unique_masks=unique_masks, train_masks=train_masks, noise_type=noise_type, input_size=input_size, mix_maps=mix_maps), pool(kernel_size=3, stride=2, padding=1), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, debug=debug, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 2), mix_maps=mix_maps), pool(kernel_size=3, stride=2, padding=1), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 4), mix_maps=mix_maps), pool(kernel_size=3, stride=2, padding=1)) self.last_layers = nn.Sequential(self.dropout, self.linear1, self.batch_norm, self.act, self.dropout, self.linear2) def forward(self, x): x = self.first_layers(x) x = x.view(x.size(0), (- 1)) x = self.last_layers(x) return x
class CifarNet(nn.Module): def __init__(self, nfilters=None, nclasses=None, nmasks=None, level=None, filter_size=None, input_size=32, linear=256, scale_noise=1, act='relu', use_act=False, first_filter_size=None, pool_type=None, dropout=None, unique_masks=False, debug=False, train_masks=False, noise_type='uniform', mix_maps=None): super(CifarNet, self).__init__() if (filter_size == 5): n = 5 else: n = 4 if (input_size == 32): first_channels = 3 elif (input_size == 28): first_channels = 1 if (pool_type == 'max'): pool = nn.MaxPool2d elif (pool_type == 'avg'): pool = nn.AvgPool2d else: print('\n\nPool Type {} is not supported/understood\n\n'.format(pool_type)) return self.linear1 = nn.Linear(((nfilters * n) * n), linear) self.linear2 = nn.Linear(linear, nclasses) self.dropout = nn.Dropout(p=dropout) self.act = act_fn(act) self.batch_norm = nn.BatchNorm1d(linear) self.first_layers = nn.Sequential(PerturbLayer(in_channels=first_channels, out_channels=nfilters, nmasks=nmasks, level=(level * scale_noise), unique_masks=unique_masks, filter_size=first_filter_size, use_act=use_act, input_size=input_size, act=act, train_masks=train_masks, noise_type=noise_type, mix_maps=mix_maps), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, debug=debug, use_act=True, act=act, mix_maps=mix_maps, unique_masks=unique_masks, train_masks=train_masks, noise_type=noise_type, input_size=input_size), pool(kernel_size=3, stride=2, padding=1), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, mix_maps=mix_maps, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 2)), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, mix_maps=mix_maps, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 2)), pool(kernel_size=3, stride=2, padding=1), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, mix_maps=mix_maps, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 4)), PerturbLayer(in_channels=nfilters, out_channels=nfilters, nmasks=nmasks, level=level, filter_size=filter_size, use_act=True, act=act, unique_masks=unique_masks, mix_maps=mix_maps, train_masks=train_masks, noise_type=noise_type, input_size=(input_size // 4)), pool(kernel_size=3, stride=2, padding=1)) self.last_layers = nn.Sequential(self.dropout, self.linear1, self.batch_norm, self.act, self.dropout, self.linear2) def forward(self, x): x = self.first_layers(x) x = x.view(x.size(0), (- 1)) x = self.last_layers(x) return x
class NoiseLayer(nn.Module): def __init__(self, in_planes, out_planes, level): super(NoiseLayer, self).__init__() self.noise = nn.Parameter(torch.Tensor(0), requires_grad=False).to(device) self.level = level self.layers = nn.Sequential(nn.ReLU(True), nn.BatchNorm2d(in_planes), nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1)) def forward(self, x): if (self.noise.numel() == 0): self.noise.resize_(x.data[0].shape).uniform_() self.noise = (((2 * self.noise) - 1) * self.level) y = torch.add(x, self.noise) return self.layers(y)
class NoiseBasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1, shortcut=None, level=0.2): super(NoiseBasicBlock, self).__init__() self.layers = nn.Sequential(NoiseLayer(in_planes, planes, level), nn.MaxPool2d(stride, stride), nn.BatchNorm2d(planes), nn.ReLU(True), NoiseLayer(planes, planes, level), nn.BatchNorm2d(planes)) self.shortcut = shortcut def forward(self, x): residual = x y = self.layers(x) if self.shortcut: residual = self.shortcut(x) y += residual y = F.relu(y) return y
class NoiseResNet(nn.Module): def __init__(self, block, nblocks, nfilters, nclasses, pool, level, first_filter_size=3): super(NoiseResNet, self).__init__() self.in_planes = nfilters if (first_filter_size == 7): pool = 1 self.pre_layers = nn.Sequential(nn.Conv2d(3, nfilters, kernel_size=first_filter_size, stride=2, padding=3, bias=False), nn.BatchNorm2d(nfilters), nn.ReLU(True), nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) elif (first_filter_size == 3): pool = 4 self.pre_layers = nn.Sequential(nn.Conv2d(3, nfilters, kernel_size=first_filter_size, stride=1, padding=1, bias=False), nn.BatchNorm2d(nfilters), nn.ReLU(True)) elif (first_filter_size == 0): print('\n\nThe original noiseresnet18 model does not support noise masks in the first layer, use perturb_resnet18 model, or set first_filter_size to 3 or 7\n\n') return self.pre_layers[0].weight.requires_grad = False self.layer1 = self._make_layer(block, (1 * nfilters), nblocks[0], stride=1, level=level) self.layer2 = self._make_layer(block, (2 * nfilters), nblocks[1], stride=2, level=level) self.layer3 = self._make_layer(block, (4 * nfilters), nblocks[2], stride=2, level=level) self.layer4 = self._make_layer(block, (8 * nfilters), nblocks[3], stride=2, level=level) self.avgpool = nn.AvgPool2d(pool, stride=1) self.linear = nn.Linear(((8 * nfilters) * block.expansion), nclasses) def _make_layer(self, block, planes, nblocks, stride=1, level=0.2, filter_size=1): shortcut = None if ((stride != 1) or (self.in_planes != (planes * block.expansion))): shortcut = nn.Sequential(nn.Conv2d(self.in_planes, (planes * block.expansion), kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d((planes * block.expansion))) layers = [] layers.append(block(self.in_planes, planes, stride, shortcut, level=level)) self.in_planes = (planes * block.expansion) for i in range(1, nblocks): layers.append(block(self.in_planes, planes, level=level)) return nn.Sequential(*layers) def forward(self, x): x1 = self.pre_layers(x) x2 = self.layer1(x1) x3 = self.layer2(x2) x4 = self.layer3(x3) x5 = self.layer4(x4) x6 = self.avgpool(x5) x7 = x6.view(x6.size(0), (- 1)) x8 = self.linear(x7) return x8
class BasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.shortcut = nn.Sequential() if ((stride != 1) or (in_planes != (self.expansion * planes))): self.shortcut = nn.Sequential(nn.Conv2d(in_planes, (self.expansion * planes), kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d((self.expansion * planes))) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) out += self.shortcut(x) out = F.relu(out) return out
class ResNet(nn.Module): def __init__(self, block, num_blocks, nfilters=64, avgpool=4, nclasses=10): super(ResNet, self).__init__() self.in_planes = nfilters self.avgpool = avgpool self.conv1 = nn.Conv2d(3, nfilters, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(nfilters) self.layer1 = self._make_layer(block, nfilters, num_blocks[0], stride=1) self.layer2 = self._make_layer(block, (nfilters * 2), num_blocks[1], stride=2) self.layer3 = self._make_layer(block, (nfilters * 4), num_blocks[2], stride=2) self.layer4 = self._make_layer(block, (nfilters * 8), num_blocks[3], stride=2) self.linear = nn.Linear(((nfilters * 8) * block.expansion), nclasses) def _make_layer(self, block, planes, num_blocks, stride): strides = ([stride] + ([1] * (num_blocks - 1))) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, stride)) self.in_planes = (planes * block.expansion) return nn.Sequential(*layers) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = self.layer4(out) out = F.avg_pool2d(out, self.avgpool) out = out.view(out.size(0), (- 1)) out = self.linear(out) return out
def resnet18(nfilters, avgpool=4, nclasses=10, nmasks=32, level=0.1, filter_size=0, first_filter_size=0, pool_type=None, input_size=None, scale_noise=1, act='relu', use_act=True, dropout=0.5, unique_masks=False, noise_type='uniform', train_masks=False, debug=False, mix_maps=None): return ResNet(BasicBlock, [2, 2, 2, 2], nfilters=nfilters, avgpool=avgpool, nclasses=nclasses)
def noiseresnet18(nfilters, avgpool=4, nclasses=10, nmasks=32, level=0.1, filter_size=0, first_filter_size=7, pool_type=None, input_size=None, scale_noise=1, act='relu', use_act=True, dropout=0.5, unique_masks=False, debug=False, noise_type='uniform', train_masks=False, mix_maps=None): return NoiseResNet(NoiseBasicBlock, [2, 2, 2, 2], nfilters=nfilters, pool=avgpool, nclasses=nclasses, level=level, first_filter_size=first_filter_size)
def perturb_resnet18(nfilters, avgpool=4, nclasses=10, nmasks=32, level=0.1, filter_size=0, first_filter_size=0, pool_type=None, input_size=None, scale_noise=1, act='relu', use_act=True, dropout=0.5, unique_masks=False, debug=False, noise_type='uniform', train_masks=False, mix_maps=None): return PerturbResNet(PerturbBasicBlock, [2, 2, 2, 2], nfilters=nfilters, avgpool=avgpool, nclasses=nclasses, pool_type=pool_type, scale_noise=scale_noise, nmasks=nmasks, level=level, filter_size=filter_size, train_masks=train_masks, first_filter_size=first_filter_size, act=act, use_act=use_act, unique_masks=unique_masks, debug=debug, noise_type=noise_type, input_size=input_size, mix_maps=mix_maps)
def lenet(nfilters, avgpool=None, nclasses=10, nmasks=32, level=0.1, filter_size=3, first_filter_size=0, pool_type=None, input_size=None, scale_noise=1, act='relu', use_act=True, dropout=0.5, unique_masks=False, debug=False, noise_type='uniform', train_masks=False, mix_maps=None): return LeNet(nfilters=nfilters, nclasses=nclasses, nmasks=nmasks, level=level, filter_size=filter_size, pool_type=pool_type, scale_noise=scale_noise, act=act, first_filter_size=first_filter_size, input_size=input_size, mix_maps=mix_maps, use_act=use_act, dropout=dropout, unique_masks=unique_masks, debug=debug, noise_type=noise_type, train_masks=train_masks)
def cifarnet(nfilters, avgpool=None, nclasses=10, nmasks=32, level=0.1, filter_size=3, first_filter_size=0, pool_type=None, input_size=None, scale_noise=1, act='relu', use_act=True, dropout=0.5, unique_masks=False, debug=False, noise_type='uniform', train_masks=False, mix_maps=None): return CifarNet(nfilters=nfilters, nclasses=nclasses, nmasks=nmasks, level=level, filter_size=filter_size, pool_type=pool_type, scale_noise=scale_noise, act=act, use_act=use_act, first_filter_size=first_filter_size, input_size=input_size, dropout=dropout, unique_masks=unique_masks, debug=debug, noise_type=noise_type, train_masks=train_masks, mix_maps=mix_maps)
class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.linear = nn.Linear(((9 * 6) * 6), 10) self.noise = nn.Parameter(torch.Tensor(1, 1, 28, 28), requires_grad=True) self.noise.data.uniform_((- 1), 1) self.layers = nn.Sequential(nn.Conv2d(1, 9, kernel_size=5, stride=2, bias=False), nn.MaxPool2d(2, 2), nn.ReLU()) def forward(self, x): x = torch.add(x, self.noise) x = self.layers(x) x = x.view(x.size(0), (- 1)) x = self.linear(x) print('{:.5f}'.format(self.noise.data[(0, 0, 0, 0)].cpu().numpy())) return x
class Normalize(nn.Module): def __init__(self, mean, std): super(Normalize, self).__init__() self.register_buffer('mean', torch.Tensor(mean)) self.register_buffer('std', torch.Tensor(std)) def forward(self, x): mean = self.mean.reshape(1, 3, 1, 1) std = self.std.reshape(1, 3, 1, 1) return ((x - mean) / std)
def add_data_normalization(model, mean, std): norm_layer = Normalize(mean=mean, std=std) model_ = torch.nn.Sequential(norm_layer, model) return model_
def apply_attack_on_dataset(model, dataloader, attack, epsilons, device, verbose=True): robust_accuracy = [] c_a = [] for (images, labels) in dataloader: (images, labels) = (images.to(device), labels.to(device)) outputs = model(images) (_, pre) = torch.max(outputs.data, 1) correct_predictions = (pre == labels) c_a.append((correct_predictions.sum() / len(correct_predictions)).cpu().numpy()) clean_accuracy = np.mean(c_a) print('Clean accuracy: ', clean_accuracy) for epsilon in epsilons: attack.eps = epsilon r_a = [] if verbose: print('Epsilon: ', epsilon) t = trange(len(dataloader)) for (images, labels) in dataloader: (images, labels) = (images.to(device), labels.to(device)) adv_images = attack(images, labels) outputs = model(adv_images) (_, pre) = torch.max(outputs.data, 1) correct_predictions = (pre == labels) r_a.append((correct_predictions.sum() / len(correct_predictions)).cpu().numpy()) if verbose: t.update(1) robust_acc = np.mean(r_a) if verbose: print('Robust accuracy: ', robust_acc) robust_accuracy.append(robust_acc) return (clean_accuracy, robust_accuracy)
def apply_attack_on_batch(model, images, labels, attack, device): (images, labels) = (images.to(device), labels.to(device)) outputs = model(images) (_, pre) = torch.max(outputs.data, 1) correct_predictions = (pre == labels) correct_predictions = correct_predictions.cpu().numpy() clean_accuracy = (correct_predictions.sum() / len(correct_predictions)) adv_images = attack(images, labels) outputs = model(adv_images) (_, pre) = torch.max(outputs.data, 1) correct_predictions_adv = (pre == labels) correct_predictions_adv = correct_predictions_adv.cpu().numpy() robust_accuracy = (correct_predictions_adv.sum() / len(correct_predictions_adv)) adversarial_success = [] for (pred_c, pred_r) in zip(correct_predictions, correct_predictions_adv): if (pred_c and (not pred_r)): adversarial_success.append(True) else: adversarial_success.append(False) print('Clean Accuracy on Batch: {}%'.format(clean_accuracy)) print('Robust Accuracy on Batch: {}%'.format(robust_accuracy)) return (adv_images.cpu(), adversarial_success, clean_accuracy, robust_accuracy)
def plot_accuracy(x, accuracy, methods, title, xlabel='x', ylabel='accuracy'): for i in range(len(methods)): plt.plot(x, accuracy[i], label=methods[i]) plt.legend() plt.xlabel(xlabel) plt.ylabel(ylabel) plt.title(title) plt.show()
def imshow(img, title): npimg = img.numpy() fig = plt.figure(figsize=(15, 15)) plt.imshow(np.transpose(npimg, (1, 2, 0))) plt.title(title) plt.show()
class Conv2dGrad(autograd.Function): @staticmethod def forward(context, input, weight, bias, stride, padding, dilation, groups): (context.stride, context.padding, context.dilation, context.groups) = (stride, padding, dilation, groups) context.save_for_backward(input, weight, bias) output = torch.nn.functional.conv2d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) return output @staticmethod def backward(context, grad_output): (input, weight, bias) = context.saved_tensors grad_input = grad_weight = grad_bias = None if context.needs_input_grad[0]: grad_input = torch.nn.grad.conv2d_input(input_size=input.shape, weight=weight, grad_output=grad_output, stride=context.stride, padding=context.padding, dilation=context.dilation, groups=context.groups) if context.needs_input_grad[1]: grad_weight = torch.nn.grad.conv2d_weight(input=input, weight_size=weight.shape, grad_output=grad_output, stride=context.stride, padding=context.padding, dilation=context.dilation, groups=context.groups) if ((bias is not None) and context.needs_input_grad[2]): grad_bias = grad_output.sum(0).sum(2).sum(1) return (grad_input, grad_weight, grad_bias, None, None, None, None)
class LinearGrad(autograd.Function): @staticmethod def forward(context, input, weight, bias=None): context.save_for_backward(input, weight, bias) output = torch.nn.functional.linear(input, weight, bias) return output @staticmethod def backward(context, grad_output): (input, weight, bias) = context.saved_tensors grad_input = grad_weight = grad_bias = None if context.needs_input_grad[0]: grad_input = grad_output.mm(weight) if context.needs_input_grad[1]: grad_weight = grad_output.t().mm(input) if ((bias is not None) and context.needs_input_grad[2]): grad_bias = grad_output.sum(0).squeeze(0) return (grad_input, grad_weight, grad_bias)
class Conv2dGrad(autograd.Function): '\n Autograd Function that Does a backward pass using the weight_backward matrix of the layer\n ' @staticmethod def forward(context, input, weight, weight_backward, bias, bias_backward, stride, padding, dilation, groups): (context.stride, context.padding, context.dilation, context.groups) = (stride, padding, dilation, groups) context.save_for_backward(input, weight, weight_backward, bias, bias_backward) output = torch.nn.functional.conv2d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) return output @staticmethod def backward(context, grad_output): (input, weight, weight_backward, bias, bias_backward) = context.saved_tensors grad_input = grad_weight = grad_weight_backward = grad_bias = grad_bias_backward = None if context.needs_input_grad[0]: grad_input = torch.nn.grad.conv2d_input(input_size=input.shape, weight=weight_backward, grad_output=grad_output, stride=context.stride, padding=context.padding, dilation=context.dilation, groups=context.groups) if context.needs_input_grad[1]: grad_weight = torch.nn.grad.conv2d_weight(input=input, weight_size=weight_backward.shape, grad_output=grad_output, stride=context.stride, padding=context.padding, dilation=context.dilation, groups=context.groups) if ((bias is not None) and context.needs_input_grad[3]): grad_bias = grad_output.sum(0).sum(2).sum(1) return (grad_input, grad_weight, grad_weight_backward, grad_bias, grad_bias_backward, None, None, None, None)
class LinearGrad(autograd.Function): '\n Autograd Function that Does a backward pass using the weight_backward matrix of the layer\n ' @staticmethod def forward(context, input, weight, weight_backward, bias=None, bias_backward=None): context.save_for_backward(input, weight, weight_backward, bias, bias_backward) output = input.mm(weight.t()) if (bias is not None): output += bias.unsqueeze(0).expand_as(output) return output @staticmethod def backward(context, grad_output): (input, weight, weight_backward, bias, bias_backward) = context.saved_tensors grad_input = grad_weight = grad_weight_backward = grad_bias = grad_bias_backward = None if context.needs_input_grad[0]: grad_input = grad_output.mm(weight_backward) if context.needs_input_grad[1]: grad_weight = grad_output.t().mm(input) if ((bias is not None) and context.needs_input_grad[3]): grad_bias = grad_output.sum(0).squeeze(0) return (grad_input, grad_weight, grad_weight_backward, grad_bias, grad_bias_backward)
def select_loss_function(loss_function_config): if (loss_function_config['name'] == 'cross_entropy'): return torch.nn.CrossEntropyLoss()
def create_lr_scheduler(lr_scheduler_config, optimizer): gamma = lr_scheduler_config['gamma'] if (lr_scheduler_config['type'] == 'multistep_lr'): lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=lr_scheduler_config['milestones'], gamma=gamma, verbose=True) else: raise ValueError('Optimizer type {} not supported'.format(lr_scheduler_config['type'])) return lr_scheduler
def create_optimizer(optimizer_config, model): lr = optimizer_config['lr'] weight_decay = optimizer_config['weight_decay'] momentum = optimizer_config['momentum'] if (optimizer_config['type'] == 'RMSProp'): optimizer = torch.optim.RMSprop(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay) elif (optimizer_config['type'] == 'Adam'): betas = optimizer_config['betas'] optimizer = torch.optim.Adam(model.parameters(), lr=lr, betas=betas, weight_decay=weight_decay) elif (optimizer_config['type'] == 'SGD'): optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay) else: raise ValueError('Optimizer type {} not supported'.format(optimizer_config['type'])) return optimizer
class Benchmark(): def __init__(self, config_file): self.config_file_path = config_file self.config_file = read_yaml(config_file) validate_config(self.config_file, 'benchmark', defaults=True) torch.manual_seed(self.config_file['experiment']['seed']) random.seed(self.config_file['experiment']['seed']) np.random.seed(self.config_file['experiment']['seed']) self.deterministic = self.config_file['experiment']['deterministic'] if self.deterministic: cudnn.benchmark = False torch.use_deterministic_algorithms(True) else: cudnn.benchmark = True self.gpus = self.config_file['infrastructure']['gpus'] self.model_config = self.config_file['model'] if ('training' not in self.config_file): self.benchmark_mode = 'evaluation' else: self.benchmark_mode = 'training' self.hyperparameters = self.config_file['training']['hyperparameters'] self.metrics_config = self.config_file['training']['metrics'] self.optimizer_config = self.config_file['training']['optimizer'] self.lr_scheduler_config = self.config_file['training']['lr_scheduler'] self.mode_names = sorted((name for name in models.__dict__ if (name.islower() and (not name.startswith('__')) and isinstance(models.__dict__[name], ModuleType)))) self.mode = self.model_config['mode']['type'] self.layer_config = {} self.layer_config = {'type': self.mode} if ('options' in self.model_config['mode']): self.mode_options = self.model_config['mode']['options'] self.layer_config['options'] = self.mode_options if (self.mode not in self.mode_names): raise ValueError('Mode not {} supported'.format(self.mode)) options = models.__dict__[self.mode].__dict__ self.model_names = sorted((name for name in options if (name.islower() and (not name.startswith('__')) and callable(options[name])))) self.loss_function_config = self.config_file['model']['loss_function'] self.data_config = self.config_file['data'] self.num_workers = self.config_file['data']['num_workers'] self.multi_gpu = False if (isinstance(self.gpus, int) and (self.gpus <= (- 1))): self.device = 'cpu' else: if (not torch.cuda.is_available()): raise ValueError('You selected {} GPUs but there are no GPUs available'.format(self.gpus)) self.device = 'cuda' if isinstance(self.gpus, int): self.device += (':' + str(self.gpus)) elif isinstance(self.gpus, list): self.device += (':' + str(self.gpus[0])) self.multi_gpu = True self.output_dir = os.path.join(self.config_file['experiment']['output_dir'], self.config_file['experiment']['name']) mkdir(self.output_dir) shutil.copy2(self.config_file_path, os.path.join(self.output_dir, 'config.yaml')) def run(self): self.epochs = self.hyperparameters['epochs'] self.batch_size = self.hyperparameters['batch_size'] self.target_size = self.data_config['target_size'] self.dataset_creator = DatasetSelector(self.data_config['dataset']).get_dataset() if (self.data_config['dataset_path'] is not None): self.dataset = self.dataset_creator(self.target_size, dataset_path=self.data_config['dataset_path']) else: self.dataset = self.dataset_creator(self.target_size) self.train_dataloader = self.dataset.create_train_dataloader(self.batch_size, deterministic=self.deterministic, num_workers=self.num_workers) self.val_dataloader = self.dataset.create_val_dataloader(self.batch_size, deterministic=self.deterministic, num_workers=self.num_workers) self.num_classes = self.dataset.num_classes if ((self.model_config['architecture'] is not None) and (self.model_config['architecture'] in self.model_names)): arch = self.model_config['architecture'] if self.model_config['pretrained']: print("=> Using pre-trained model '{}'".format(self.model_config['architecture'])) else: print("=> Creating model from scratch '{}'".format(self.model_config['architecture'])) self.model = models.__dict__[self.mode].__dict__[arch](pretrained=self.model_config['pretrained'], num_classes=self.num_classes, layer_config=self.layer_config) elif (self.model_config['checkpoint'] is not None): print('Loading model checkpoint from ', self.model_config['checkpoint']) self.model = torch.load(self.model_config['checkpoint'], map_location=self.device) self.model.to(self.device) if isinstance(self.gpus, list): self.model = nn.DataParallel(self.model, self.gpus) self.loss_function = select_loss_function(self.loss_function_config) self.optimizer = create_optimizer(self.optimizer_config, self.model) self.lr_scheduler = create_lr_scheduler(self.lr_scheduler_config, self.optimizer) print('\nBenchmarking model on {}'.format(str(self.dataset))) print(self.metrics_config) trainer = Trainer(model=self.model, mode=self.mode, loss_function=self.loss_function, optimizer=self.optimizer, lr_scheduler=self.lr_scheduler, train_dataloader=self.train_dataloader, val_dataloader=self.val_dataloader, device=self.device, epochs=self.epochs, output_dir=self.output_dir, metrics_config=self.metrics_config, multi_gpu=self.multi_gpu) trainer.run() if self.config_file['evaluation']: self.model = torch.load(os.path.join(self.output_dir, 'model_best_acc.pth')) self.test_dataloader = self.dataset.create_test_dataloader(self.batch_size, deterministic=self.deterministic, num_workers=self.num_workers) self.evaluator = Evaluator(self.model, self.mode, self.loss_function, self.test_dataloader, self.device, self.output_dir) self.evaluate(self.evaluator) def run_eval(self): if (self.model_config['checkpoint'] is not None): print('Loading model checkpoint from ', self.model_config['checkpoint']) self.model = torch.load(self.model_config['checkpoint'], map_location=self.device) else: raise ValueError('A model checkpoint must be specified') self.target_size = self.data_config['target_size'] self.batch_size = self.data_config['batch_size'] self.dataset_creator = DatasetSelector(self.data_config['dataset']).get_dataset() if (self.data_config['dataset_path'] is not None): self.dataset = self.dataset_creator(self.target_size, dataset_path=self.data_config['dataset_path']) else: self.dataset = self.dataset_creator(self.target_size) self.test_dataloader = self.dataset.create_test_dataloader(self.batch_size, deterministic=self.deterministic, num_workers=self.num_workers) self.loss_function = select_loss_function(self.loss_function_config) self.evaluator = Evaluator(self.model, None, self.loss_function, self.test_dataloader, self.device, self.output_dir) self.evaluate(self.evaluator) def evaluate(self, evaluator): (self.test_acc, self.test_loss) = evaluator.run() self.results_df = pd.DataFrame({'model_name': [self.config_file['experiment']['name']], 'dataset': [self.data_config['dataset']], 'accuracy': [float(self.test_acc)], 'error': [(100.0 - float(self.test_acc))], 'loss': [self.test_loss]}) print('Test Results') print(self.results_df) csv_results = os.path.join(self.output_dir, 'results.csv') json_results = os.path.join(self.output_dir, 'results.json') print('Test Results saved in: \n{}\n{}'.format(csv_results, json_results)) self.results_df.to_csv(csv_results) self.results_df.to_json(json_results, indent=2, orient='records')
def __main__(): parser = argparse.ArgumentParser(description='BioTorch') parser.add_argument('--config_file', help='Path to the configuration file') try: args = parser.parse_args() benchmark = Benchmark(args.config_file) if (benchmark.benchmark_mode == 'training'): benchmark.run() else: benchmark.run_eval() except Exception as e: message = 'an unexpected error occurred: {}: {}'.format(type(e).__name__, ((e.message if hasattr(e, 'message') else '') or str(e))) raise ValueError(message)
class CIFAR100(Dataset): def __str__(self): return 'CIFAR-100 Dataset' def __init__(self, target_size, dataset_path='./datasets/cifar100', train_transforms=None, test_transforms=None): self.mean = (0.5071, 0.4867, 0.4408) self.std = (0.2675, 0.2565, 0.2761) self.num_classes = 100 super(CIFAR100, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Preparing {} and storing data in {}'.format(str(self), dataset_path)) self.train_dataset = datasets.CIFAR100(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.train_transforms)) self.val_dataset = datasets.CIFAR100(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms)) self.test_dataset = datasets.CIFAR100(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms))
class CIFAR10(Dataset): def __str__(self): return 'CIFAR-10 Dataset' def __init__(self, target_size, dataset_path='./datasets/cifar10', train_transforms=None, test_transforms=None): self.mean = (0.4914, 0.4821, 0.4465) self.std = (0.247, 0.2435, 0.2616) self.num_classes = 10 super(CIFAR10, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Preparing {} and storing data in {}'.format(str(self), dataset_path)) self.train_dataset = datasets.CIFAR10(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.train_transforms)) self.val_dataset = datasets.CIFAR10(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms)) self.test_dataset = datasets.CIFAR10(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms))
class CIFAR10Benchmark(Dataset): def __str__(self): return 'CIFAR-10 Benchmark Dataset' def __init__(self, target_size, dataset_path='./datasets/cifar10', train_transforms=None, test_transforms=None): self.mean = (0.4914, 0.4821, 0.4465) self.std = (0.247, 0.2435, 0.2616) self.num_classes = 10 super(CIFAR10Benchmark, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Preparing {} and storing data in {}'.format(str(self), dataset_path)) random.seed(0) self.train_transforms = [transforms.RandomHorizontalFlip(), transforms.RandomCrop(32, padding=4), transforms.ToTensor(), transforms.Normalize(self.mean, self.std)] self.train_dataset = datasets.CIFAR10(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.train_transforms)) val_indices = random.sample(range(0, len(self.train_dataset.data)), 5000) self.train_dataset.data = np.delete(self.train_dataset.data, val_indices, axis=0) self.train_dataset.targets = np.delete(self.train_dataset.targets, val_indices, axis=0) self.val_dataset = datasets.CIFAR10(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.test_transforms)) self.val_dataset.data = self.val_dataset.data[val_indices] self.val_dataset.targets = list(np.array(self.val_dataset.targets)[val_indices]) self.test_dataset = datasets.CIFAR10(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms))
class Dataset(object): def __init__(self, target_size, dataset_path, mean=None, std=None, train_transforms=None, test_transforms=None): self.dataset_path = dataset_path self.target_size = target_size self.mean = mean self.std = std self.train_transforms = train_transforms self.test_transforms = test_transforms default_transforms = [transforms.Resize((self.target_size, self.target_size)), transforms.ToTensor()] if ((self.mean is not None) and (self.std is not None)): default_transforms.append(transforms.Normalize(self.mean, self.std)) if (self.train_transforms is None): self.train_transforms = default_transforms if (self.test_transforms is None): self.test_transforms = default_transforms @staticmethod def seed_worker(worker_id): worker_seed = (torch.initial_seed() % (2 ** 32)) numpy.random.seed(worker_seed) random.seed(worker_seed) def _create_dataloader(self, mode, batch_size, deterministic=False, shuffle=True, drop_last=True, num_workers=0): gen = None worker_init_fn = None if (deterministic and (num_workers > 0)): worker_init_fn = self.seed_worker gen = torch.Generator() gen.manual_seed(0) if (mode == 'train'): return DataLoader(self.train_dataset, batch_size=batch_size, shuffle=shuffle, drop_last=drop_last, num_workers=num_workers, worker_init_fn=worker_init_fn, generator=gen) elif (mode == 'val'): return DataLoader(self.val_dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=num_workers, worker_init_fn=worker_init_fn, generator=gen) elif (mode == 'test'): return DataLoader(self.test_dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=num_workers, worker_init_fn=worker_init_fn, generator=gen) def create_train_dataloader(self, batch_size, deterministic=False, num_workers=0): return self._create_dataloader('train', batch_size, deterministic=deterministic, num_workers=num_workers) def create_val_dataloader(self, batch_size, deterministic=False, num_workers=0): return self._create_dataloader('val', batch_size, deterministic=deterministic, num_workers=num_workers) def create_test_dataloader(self, batch_size, deterministic=False, num_workers=0): return self._create_dataloader('test', batch_size, deterministic=deterministic, num_workers=num_workers)
class FashionMNIST(Dataset): def __str__(self): return 'Fashion MNIST Dataset' def __init__(self, target_size, dataset_path='./datasets/fashion-mnist', train_transforms=None, test_transforms=None): self.mean = (0.2859,) self.std = (0.353,) self.num_classes = 10 super(FashionMNIST, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Preparing {} and storing data in {}'.format(str(self), dataset_path)) self.train_dataset = datasets.FashionMNIST(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.train_transforms)) self.val_dataset = datasets.FashionMNIST(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms)) self.test_dataset = datasets.FashionMNIST(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms))
class ImageNet(Dataset): def __str__(self): return 'Imagenet Dataset' def __init__(self, target_size, dataset_path='./datasets/imagenet', train_transforms=None, test_transforms=None): self.mean = (0.485, 0.456, 0.406) self.std = (0.229, 0.224, 0.225) self.num_classes = 1000 normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if (train_transforms is None): train_transforms = [transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize] if (test_transforms is None): test_transforms = [transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize] super(ImageNet, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Reading {} data from {}'.format(str(self), dataset_path)) self.train_dataset = datasets.ImageFolder(os.path.join(self.dataset_path, 'train'), transform=transforms.Compose(self.train_transforms)) self.val_dataset = datasets.ImageFolder(os.path.join(self.dataset_path, 'val'), transform=transforms.Compose(self.test_transforms)) self.test_dataset = datasets.ImageFolder(os.path.join(self.dataset_path, 'val'), transform=transforms.Compose(self.test_transforms))
class MNIST(Dataset): def __str__(self): return 'MNIST Dataset' def __init__(self, target_size, dataset_path='./datasets/mnist', train_transforms=None, test_transforms=None): self.mean = (0.1307,) self.std = (0.3081,) self.num_classes = 10 super(MNIST, self).__init__(target_size=target_size, dataset_path=dataset_path, mean=self.mean, std=self.std, train_transforms=train_transforms, test_transforms=test_transforms) print('Preparing {} and storing data in {}'.format(str(self), dataset_path)) self.train_dataset = datasets.MNIST(self.dataset_path, train=True, download=True, transform=transforms.Compose(self.train_transforms)) self.val_dataset = datasets.MNIST(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms)) self.test_dataset = datasets.MNIST(self.dataset_path, train=False, download=True, transform=transforms.Compose(self.test_transforms))
class DatasetSelector(): def __init__(self, dataset_name): if (dataset_name not in DATASETS_AVAILABLE): raise ValueError('Dataset name specified: {} not in the list of available datasets {}'.format(dataset_name, DATASETS_AVAILABLE)) self.dataset_name = dataset_name def get_dataset(self): if (self.dataset_name == 'cifar10'): return CIFAR10 elif (self.dataset_name == 'cifar10_benchmark'): return CIFAR10Benchmark elif (self.dataset_name == 'cifar100'): return CIFAR100 elif (self.dataset_name == 'mnist'): return MNIST elif (self.dataset_name == 'fashion_mnist'): return FashionMNIST elif (self.dataset_name == 'imagenet'): return ImageNet
class Evaluator(): def __init__(self, model, mode, loss_function, dataloader, device, output_dir, multi_gpu=False): self.model = model self.mode = mode self.output_dir = output_dir self.logs_dir = os.path.join(output_dir, 'logs') self.loss_function = loss_function self.dataloader = dataloader self.device = device self.multi_gpu = multi_gpu def run(self): (acc, loss) = test(model=self.model, loss_function=self.loss_function, test_dataloader=self.dataloader, device=self.device) return (acc.cpu().numpy(), loss)
class Conv2d(nn.Conv2d): def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode) self.layer_config = layer_config if (self.layer_config is None): self.layer_config = {'type': 'backpropagation'} if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'gradient_clip': False, 'init': 'xavier'} self.options = self.layer_config['options'] self.type = self.layer_config['type'] self.init = self.options['init'] self.init_parameters() def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) if (self.bias is not None): nn.init.constant_(self.bias, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound)
class Linear(nn.Linear): def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: super(Linear, self).__init__(in_features, out_features, bias) self.layer_config = layer_config if (self.layer_config is None): self.layer_config = {'type': 'backpropagation'} if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'gradient_clip': False, 'init': 'xavier'} self.options = self.layer_config['options'] self.type = self.layer_config['type'] self.init = self.options['init'] self.init_parameters() def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) if (self.bias is not None): nn.init.constant_(self.bias, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound)
class Conv2d(fa_constructor.Conv2d): '\n Implements the method from How Important Is Weight Symmetry in Backpropagation?\n with the modification of taking the absolute value of the Backward Matrix\n\n Batchwise Random Magnitude Sign-concordant Feedbacks (brSF):\n weight_backward = |M| ◦ sign(weight), where M is redrawn after each update of W (i.e., each mini-batch).\n\n (https://arxiv.org/pdf/1510.05067.pdf)\n ' def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): if (layer_config is None): layer_config = {} layer_config['type'] = 'brsf' super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode, layer_config)
class Linear(fa_constructor.Linear): '\n Implements the method from How Important Is Weight Symmetry in Backpropagation?\n with the modification of taking the absolute value of the Backward Matrix\n\n Batchwise Random Magnitude Sign-concordant Feedbacks (brSF):\n weight_backward = |M| ◦ sign(weight), where M is redrawn after each update of W (i.e., each mini-batch).\n\n (https://arxiv.org/pdf/1510.05067.pdf)\n\n ' def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: if (layer_config is None): layer_config = {} layer_config['type'] = 'brsf' super(Linear, self).__init__(in_features, out_features, bias, layer_config)
class Conv2d(nn.Conv2d): def __init__(self, in_channels: int, out_channels: int, output_dim: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode) self.layer_config = layer_config if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'init': 'xavier', 'gradient_clip': False} self.options = self.layer_config['options'] self.init = self.options['init'] self.loss_gradient = None self.weight_backward = nn.Parameter(torch.Tensor(size=(output_dim, self.in_channels, self.kernel_size[0], self.kernel_size[1])), requires_grad=False) self.bias_backward = None if (self.bias is not None): self.bias_backward = nn.Parameter(torch.Tensor(size=(output_dim, self.in_channels)), requires_grad=False) self.init_parameters() if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.norm_initial_weights = torch.linalg.norm(self.weight) self.register_backward_hook(self.dfa_backward_hook) self.weight_ratio = 0 def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) nn.init.xavier_uniform_(self.weight_backward) self.scaling_factor = math.sqrt((2.0 / float((fan_in + fan_out)))) if (self.bias is not None): nn.init.constant_(self.bias, 0) nn.init.constant_(self.bias_backward, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) nn.init.kaiming_uniform_(self.weight_backward, a=math.sqrt(5)) self.scaling_factor = (1 / math.sqrt((3 * fan_in))) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound) nn.init.uniform_(self.bias_backward, (- bound), bound) def compute_weight_ratio(self): with torch.no_grad(): self.weight_diff = (torch.linalg.norm(self.weight_backward) / torch.linalg.norm(self.weight)) return self.weight_diff def forward(self, x): with torch.no_grad(): if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.weight = torch.nn.Parameter(((self.weight * self.norm_initial_weights) / torch.linalg.norm(self.weight))) return Conv2dGrad.apply(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups) @staticmethod def dfa_backward_hook(module, grad_input, grad_output): if (grad_input[0] is None): return grad_input else: out_grad = module.loss_gradient.unsqueeze(2).repeat(1, 1, grad_output[0].size()[2]) out_grad = out_grad.unsqueeze(3).repeat(1, 1, 1, grad_output[0].size()[3]) grad_dfa = torch.nn.grad.conv2d_input(input_size=grad_input[0].shape, weight=module.weight_backward, grad_output=out_grad, stride=module.stride, padding=module.padding, dilation=module.dilation, groups=module.groups) if (len(grad_input) == 2): return (grad_dfa, grad_input[1]) else: return (grad_dfa, grad_input[1], grad_input[2])
class Linear(nn.Linear): def __init__(self, in_features: int, out_features: int, output_dim: int, bias: bool=True, layer_config: dict=None) -> None: super(Linear, self).__init__(in_features, out_features, bias) self.layer_config = layer_config if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'gradient_clip': False, 'init': 'xavier'} self.options = self.layer_config['options'] self.init = self.options['init'] self.loss_gradient = None self.weight_backward = nn.Parameter(torch.Tensor(size=(output_dim, self.in_features)), requires_grad=False) self.bias_backward = None if (self.bias is not None): self.bias_backward = nn.Parameter(torch.Tensor(size=(output_dim, self.in_features)), requires_grad=False) self.init_parameters() if (('constrain_weights' in self.options) and self.options['constrain_weights']): with torch.no_grad(): self.norm_initial_weights = torch.linalg.norm(self.weight) self.register_backward_hook(self.dfa_backward_hook) self.weight_ratio = 0 def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) nn.init.xavier_uniform_(self.weight_backward) self.scaling_factor = math.sqrt((2.0 / float((fan_in + fan_out)))) if (self.bias is not None): nn.init.constant_(self.bias, 0) nn.init.constant_(self.bias_backward, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) nn.init.kaiming_uniform_(self.weight_backward, a=math.sqrt(5)) self.scaling_factor = (1 / math.sqrt((3 * fan_in))) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound) nn.init.uniform_(self.bias_backward, (- bound), bound) def forward(self, x): with torch.no_grad(): if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.weight = torch.nn.Parameter(((self.weight * self.norm_initial_weights) / torch.linalg.norm(self.weight))) return LinearGrad.apply(x, self.weight, self.bias) def compute_weight_ratio(self): with torch.no_grad(): self.weight_diff = (torch.linalg.norm(self.weight_backward) / torch.linalg.norm(self.weight)) return self.weight_diff @staticmethod def dfa_backward_hook(module, grad_input, grad_output): if (grad_input[0] is None): return grad_input else: grad_dfa = module.loss_gradient.mm(module.weight_backward) if (len(grad_input) == 2): return (grad_dfa, grad_input[1]) else: return (grad_dfa, grad_input[1], grad_input[2])
class Conv2d(fa_constructor.Conv2d): def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): if (layer_config is None): layer_config = {} layer_config['type'] = 'fa' super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode, layer_config)
class Linear(fa_constructor.Linear): def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: if (layer_config is None): layer_config = {} layer_config['type'] = 'fa' super(Linear, self).__init__(in_features, out_features, bias, layer_config)
class Conv2d(nn.Conv2d): def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode) self.layer_config = layer_config if (self.layer_config is None): self.layer_config = {'type': 'fa'} if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'gradient_clip': False, 'init': 'xavier'} self.options = self.layer_config['options'] self.type = self.layer_config['type'] self.init = self.options['init'] self.weight_backward = nn.Parameter(torch.Tensor(self.weight.size()), requires_grad=False) if (self.bias is not None): self.bias_backward = nn.Parameter(torch.Tensor(self.bias.size()), requires_grad=False) else: self.register_parameter('bias', None) self.bias_backward = None self.init_parameters() if (self.type == 'frsf'): self.weight_backward = nn.Parameter(torch.abs(self.weight_backward), requires_grad=False) if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.norm_initial_weights = torch.linalg.norm(self.weight) if ((self.type == 'usf') or (self.type == 'brsf')): with torch.no_grad(): self.weight_backward = torch.nn.Parameter((self.scaling_factor * torch.sign(self.weight)), requires_grad=False) self.alignment = 0 self.weight_ratio = 0 if (('gradient_clip' in self.options) and self.options['gradient_clip']): self.register_backward_hook(self.gradient_clip) def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) nn.init.xavier_uniform_(self.weight_backward) self.scaling_factor = math.sqrt((2.0 / float((fan_in + fan_out)))) if (self.bias is not None): nn.init.constant_(self.bias, 0) nn.init.constant_(self.bias_backward, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) nn.init.kaiming_uniform_(self.weight_backward, a=math.sqrt(5)) self.scaling_factor = (1 / math.sqrt((3 * fan_in))) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound) nn.init.uniform_(self.bias_backward, (- bound), bound) def forward(self, x: Tensor) -> Tensor: weight_backward = None with torch.no_grad(): if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.weight = torch.nn.Parameter(((self.weight * self.norm_initial_weights) / torch.linalg.norm(self.weight))) if (self.type == 'usf'): weight_backward = torch.nn.Parameter(torch.sign(self.weight), requires_grad=False) weight_backward = torch.nn.Parameter((self.scaling_factor * weight_backward), requires_grad=False) self.weight_backward = weight_backward elif (self.type == 'brsf'): wb = torch.Tensor(self.weight.size()).to(self.weight.device) if (self.init == 'xavier'): torch.nn.init.xavier_uniform_(wb) else: init.kaiming_uniform_(wb, a=math.sqrt(5)) weight_backward = torch.nn.Parameter((torch.abs(wb) * torch.sign(self.weight)), requires_grad=False) self.weight_backward = weight_backward elif (self.type == 'frsf'): weight_backward = torch.nn.Parameter((self.weight_backward * torch.sign(self.weight)), requires_grad=False) if (weight_backward is None): weight_backward = self.weight_backward return Conv2dGrad.apply(x, self.weight, weight_backward, self.bias, self.bias_backward, self.stride, self.padding, self.dilation, self.groups) def compute_alignment(self): self.alignment = compute_matrix_angle(self.weight_backward, self.weight) return self.alignment def compute_weight_ratio(self): with torch.no_grad(): self.weight_diff = (torch.linalg.norm(self.weight_backward) / torch.linalg.norm(self.weight)) return self.weight_diff @staticmethod def gradient_clip(module, grad_input, grad_output): grad_input = list(grad_input) for i in range(len(grad_input)): if (grad_input[i] is not None): grad_input[i] = torch.clamp(grad_input[i], (- 1), 1) return tuple(grad_input)
class Linear(nn.Linear): def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: super(Linear, self).__init__(in_features, out_features, bias) self.layer_config = layer_config if (self.layer_config is None): self.layer_config = {'type': 'fa'} if ('options' not in self.layer_config): self.layer_config['options'] = {'constrain_weights': False, 'gradient_clip': False, 'init': 'xavier'} self.options = self.layer_config['options'] self.type = self.layer_config['type'] self.init = self.options['init'] self.weight_backward = nn.Parameter(torch.Tensor(self.weight.size()), requires_grad=False) if (self.bias is not None): self.bias_backward = nn.Parameter(torch.Tensor(self.bias.size()), requires_grad=False) else: self.register_parameter('bias', None) self.bias_backward = None self.init_parameters() if (self.type == 'frsf'): self.weight_backward = nn.Parameter(torch.abs(self.weight_backward), requires_grad=False) if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.norm_initial_weights = torch.linalg.norm(self.weight) if ((self.type == 'usf') or (self.type == 'brsf')): with torch.no_grad(): self.weight_backward = torch.nn.Parameter((self.scaling_factor * torch.sign(self.weight)), requires_grad=False) self.alignment = 0 self.weight_ratio = 0 if (('gradient_clip' in self.options) and self.options['gradient_clip']): self.register_backward_hook(self.gradient_clip) def init_parameters(self) -> None: (fan_in, fan_out) = nn.init._calculate_fan_in_and_fan_out(self.weight) if (self.init == 'xavier'): nn.init.xavier_uniform_(self.weight) nn.init.xavier_uniform_(self.weight_backward) self.scaling_factor = math.sqrt((2.0 / float((fan_in + fan_out)))) if (self.bias is not None): nn.init.constant_(self.bias, 0) nn.init.constant_(self.bias_backward, 0) else: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) nn.init.kaiming_uniform_(self.weight_backward, a=math.sqrt(5)) self.scaling_factor = (1 / math.sqrt((3 * fan_in))) if (self.bias is not None): bound = ((1 / math.sqrt(fan_in)) if (fan_in > 0) else 0) nn.init.uniform_(self.bias, (- bound), bound) nn.init.uniform_(self.bias_backward, (- bound), bound) def forward(self, x: Tensor) -> Tensor: weight_backward = None with torch.no_grad(): if (('constrain_weights' in self.options) and self.options['constrain_weights']): self.weight = torch.nn.Parameter(((self.weight * self.norm_initial_weights) / torch.linalg.norm(self.weight))) if (self.type == 'usf'): weight_backward = torch.nn.Parameter(torch.sign(self.weight), requires_grad=False) weight_backward = torch.nn.Parameter((self.scaling_factor * weight_backward), requires_grad=False) self.weight_backward = weight_backward elif (self.type == 'brsf'): wb = torch.Tensor(self.weight.size()).to(self.weight.device) if (self.init == 'xavier'): torch.nn.init.xavier_uniform_(wb) else: init.kaiming_uniform_(wb, a=math.sqrt(5)) weight_backward = torch.nn.Parameter((torch.abs(wb) * torch.sign(self.weight)), requires_grad=False) self.weight_backward = weight_backward elif (self.type == 'frsf'): weight_backward = torch.nn.Parameter((torch.abs(self.weight_backward) * torch.sign(self.weight)), requires_grad=False) if (weight_backward is None): weight_backward = self.weight_backward return LinearGrad.apply(x, self.weight, weight_backward, self.bias, self.bias_backward) def compute_alignment(self): self.alignment = compute_matrix_angle(self.weight_backward, self.weight) return self.alignment def compute_weight_ratio(self): with torch.no_grad(): self.weight_diff = (torch.linalg.norm(self.weight_backward) / torch.linalg.norm(self.weight)) return self.weight_diff @staticmethod def gradient_clip(module, grad_input, grad_output): grad_input = list(grad_input) for i in range(len(grad_input)): if (grad_input[i] is not None): grad_input[i] = torch.clamp(grad_input[i], (- 1), 1) return tuple(grad_input)
class Conv2d(fa_constructor.Conv2d): '\n Implements the method from How Important Is Weight Symmetry in Backpropagation?\n with the modification of taking the absolute value of the Backward Matrix\n\n Fixed Random Magnitude Sign-concordant Feedbacks (frSF):\n weight_backward = |M| ◦ sign(weight), where M is initialized once and fixed throughout each experiment\n\n (https://arxiv.org/pdf/1510.05067.pdf)\n ' def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): if (layer_config is None): layer_config = {} layer_config['type'] = 'frsf' super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode, layer_config)
class Linear(fa_constructor.Linear): '\n Implements the method from How Important Is Weight Symmetry in Backpropagation?\n with the modification of taking the absolute value of the Backward Matrix\n\n Fixed Random Magnitude Sign-concordant Feedbacks (frSF):\n weight_backward = |M| ◦ sign(weight), where M is initialized once and fixed throughout each experiment\n\n ' def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: if (layer_config is None): layer_config = {} layer_config['type'] = 'frsf' super(Linear, self).__init__(in_features, out_features, bias, layer_config)
def compute_matrix_angle(A, B): with torch.no_grad(): flat_A = torch.reshape(A, ((- 1),)) normalized_flat_A = (flat_A / torch.norm(flat_A)) flat_B = torch.reshape(B, ((- 1),)) normalized_flat_B = (flat_B / torch.norm(flat_B)) angle = ((180.0 / math.pi) * torch.arccos(torch.clip(torch.dot(normalized_flat_A, normalized_flat_B), (- 1.0), 1.0))) return angle
class Conv2d(fa_constructor.Conv2d): '\n Implements the method from How Important Is Weight Symmetry in Backpropagation?\n\n Uniform Sign-concordant Feedbacks (uSF):\n Backward Weights = sign(W)\n\n (https://arxiv.org/pdf/1510.05067.pdf)\n ' def __init__(self, in_channels: int, out_channels: int, kernel_size: _size_2_t, stride: _size_2_t=1, padding: Union[(str, _size_2_t)]=0, dilation: _size_2_t=1, groups: int=1, bias: bool=True, padding_mode: str='zeros', layer_config: dict=None): if (layer_config is None): layer_config = {} layer_config['type'] = 'usf' super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias, padding_mode, layer_config)
class Linear(fa_constructor.Linear): '\n Method from [How Important Is Weight Symmetry in Backpropagation?](https://arxiv.org/pdf/1510.05067.pdf)\n\n Uniform Sign-concordant Feedbacks (uSF):\n weight_backward = sign(weight)\n\n ' def __init__(self, in_features: int, out_features: int, bias: bool=True, layer_config: dict=None) -> None: if (layer_config is None): layer_config = {} layer_config['type'] = 'usf' super(Linear, self).__init__(in_features, out_features, bias, layer_config)
def convert_layer(layer, mode, copy_weights, layer_config=None, output_dim=None): (layer_bias, bias_weight) = (False, None) if (('weight' in layer.__dict__['_parameters']) and copy_weights): weight = layer.weight if (('bias' in layer.__dict__['_parameters']) and (layer.bias is not None)): bias_weight = layer.bias layer_bias = True new_layer = None if (layer_config is None): layer_config = {} layer_config['type'] = mode if isinstance(layer, nn.Conv2d): if (mode in ['fa', 'usf', 'brsf', 'frsf']): new_layer = fa_constructor.Conv2d(layer.in_channels, layer.out_channels, layer.kernel_size, layer.stride, layer.padding, layer.dilation, layer.groups, layer_bias, layer.padding_mode, layer_config) elif (mode == 'dfa'): new_layer = dfa_layers.Conv2d(layer.in_channels, layer.out_channels, output_dim, layer.kernel_size, layer.stride, layer.padding, layer.dilation, layer.groups, layer_bias, layer.padding_mode, layer_config) elif (mode == 'backpropagation'): new_layer = bp_layers.Conv2d(layer.in_channels, layer.out_channels, layer.kernel_size, layer.stride, layer.padding, layer.dilation, layer.groups, layer_bias, layer.padding_mode, layer_config) elif isinstance(layer, nn.Linear): if (mode in ['fa', 'usf', 'brsf', 'frsf']): new_layer = fa_constructor.Linear(layer.in_features, layer.out_features, layer_bias, layer_config) elif (mode == 'dfa'): new_layer = dfa_layers.Linear(layer.in_features, layer.out_features, output_dim, layer_bias, layer_config) elif (mode == 'backpropagation'): new_layer = bp_layers.Linear(layer.in_features, layer.out_features, layer_bias, layer_config) if ((new_layer is not None) and copy_weights): new_layer.weight = weight new_layer.bias = bias_weight return new_layer
def alexnet(pretrained: bool=False, progress: bool=True, num_classes: int=1000, layer_config=None) -> AlexNet: 'AlexNet model architecture from the\n `"One weird trick..." <https://arxiv.org/abs/1404.5997>`_ paper.\n The required minimum input size of the model is 63x63.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n progress (bool): If True, displays a progress bar of the download to stderr\n num_classes (int): Output dimension of the last linear layer\n layer_config (dict): Custom biologically plausible method layer configuration\n ' print('Converting AlexNet to {} mode'.format(MODE_STRING)) return create_torchvision_biomodel(models.alexnet, MODE, layer_config, pretrained, progress, num_classes)
def densenet121(pretrained: bool=False, progress: bool=True, num_classes: int=1000, layer_config=None) -> DenseNet: 'Densenet-121 model from\n `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_.\n The required minimum input size of the model is 29x29.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n progress (bool): If True, displays a progress bar of the download to stderr\n num_classes (int): Output dimension of the last linear layer\n layer_config (dict): Custom biologically plausible method layer configuration\n ' print('Converting Densenet-121 to {} mode'.format(MODE_STRING)) return create_torchvision_biomodel(models.densenet121, MODE, layer_config, pretrained, progress, num_classes)
def densenet161(pretrained: bool=False, progress: bool=True, num_classes: int=1000, layer_config=None) -> DenseNet: 'Densenet-161 model from\n `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_.\n The required minimum input size of the model is 29x29.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n progress (bool): If True, displays a progress bar of the download to stderr\n num_classes (int): Output dimension of the last linear layer\n layer_config (dict): Custom biologically plausible method layer configuration\n ' print('Converting Densenet-161 to {} mode'.format(MODE_STRING)) return create_torchvision_biomodel(models.densenet161, MODE, layer_config, pretrained, progress, num_classes)
def densenet169(pretrained: bool=False, progress: bool=True, num_classes: int=1000, layer_config=None) -> DenseNet: 'Densenet-169 model from\n `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_.\n The required minimum input size of the model is 29x29.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n progress (bool): If True, displays a progress bar of the download to stderr\n num_classes (int): Output dimension of the last linear layer\n layer_config (dict): Custom biologically plausible method layer configuration\n ' print('Converting Densenet-169 to {} mode'.format(MODE_STRING)) return create_torchvision_biomodel(models.densenet169, MODE, layer_config, pretrained, progress, num_classes)
def densenet201(pretrained: bool=False, progress: bool=True, num_classes: int=1000, layer_config=None) -> DenseNet: 'Densenet-201 model from\n `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_.\n The required minimum input size of the model is 29x29.\n Args:\n pretrained (bool): If True, returns a model pre-trained on ImageNet\n progress (bool): If True, displays a progress bar of the download to stderr\n num_classes (int): Output dimension of the last linear layer\n layer_config (dict): Custom biologically plausible method layer configuration\n ' print('Converting Densenet-201 to {} mode'.format(MODE_STRING)) return create_torchvision_biomodel(models.densenet201, MODE, layer_config, pretrained, progress, num_classes)