def random_projection_mnist(opt): import data # import seaborn as sns sigmoid = nn.Sigmoid() data = data.MNIST(opt) batches = [data.next_batch_from_class(i) for i in range(10)] from sklearn.manifold import TSNE tsne_model = TSNE(n_components=2, random_state=0) np.set_printoptions(suppress=True) mnist = batches[0] for i in range(1,10): mnist = torch.cat((mnist, batches[i]), 0) output = tsne_model.fit_transform(mnist.numpy()) print(mnist) # for i in range(10): # plt.plot(output[i*100:(i+1)*100, 0], output[i*100:(i+1)*100, 1], '+') # # plt.pause(10000) M = torch.FloatTensor(784, 1024).normal_(0,200) embedding = [torch.round(sigmoid(Variable(torch.mm(batches[i], M)))) for i in range(10)] mnist = embedding[0] for i in range(1,10): mnist = torch.cat((mnist, embedding[i]), 0) print(mnist) output = tsne_model.fit_transform(mnist.data.numpy()) for i in range(10): plt.plot(output[i*100:(i+1)*100, 0], output[i*100:(i+1)*100, 1], '+') plt.pause(10000) print(embedding[0][:10]) print(embedding[4][:10]) hamming = np.zeros((10,10,3)) for i in range(10): for k in range(10): hamming[i,k,0] = hamming_distance(embedding[i], embedding[i]) hamming[i,k,1] = hamming_distance(embedding[k], embedding[k]) hamming[i,k,2] = hamming_distance(embedding[i], embedding[k]) print(i,k,hamming[i,k]) print(hamming[:,:,2]) sns.heatmap(hamming[:,:,2]) plt.pause(100)
def train_as_binary():
num_classes = 1
dataloader = [data_.MNIST(opt, class_=i) for i in range(num_classes)]
classifier = [
mlp.Classifier(opt.input_size, opt.feature_size)
for i in range(num_classes)
]
criterion = [nn.BCELoss() for i in range(num_classes)]
optimiser = [
optim.Adam(classifier[i].parameters(), lr=opt.lr)
for i in range(num_classes)
]
logs = [[] for i in range(num_classes)]
for i in range(num_classes):
for epoch in range(opt.epochs):
for idx in range(len(dataloader[i])):
classifier[i].zero_grad()
dataset, targets = dataloader[i].next()
output = classifier[i](Variable(dataset))
loss = criterion[i](output, Variable(targets))
loss.backward()
optimiser[i].step()
test_dataset, test_targets = dataloader[i].next_test()
test_output = classifier[i](Variable(test_dataset))
loss = criterion[i](test_output, Variable(test_targets))
pred = torch.round(test_output.data)
positive_correct = pred.eq(test_targets) * test_targets.byte()
positive_correct = float(
positive_correct.sum()) / (test_targets.sum())
negative_correct = pred.eq(test_targets) * (1 -
test_targets).byte()
negative_correct = float(
negative_correct.sum()) / (test_targets.sum())
# correct = float(pred.eq(test_targets).cpu().sum()) / test_output.size(0)
logs[i].append(positive_correct)
print("{0} - {1} - {2}".format(epoch, logs[i][-1],
negative_correct))
for k in range(len(logs)):
plt.plot(logs[k])
plt.pause(0.01)
plt.clf()
for k in range(len(logs)):
plt.plot(logs[k])
plt.pause(1000)
def trainGANs(n=100, datadir='data/gan/'):
for i in range(n):
try:
os.mkdir(datadir+str(i))
except FileExistsError:
pass
loader = data.MNIST(batch=128)
model = SimpleGAN(28*28, zdim=64, hd=64, hg=64, lr=2e-4).cuda()
print('Network: ' + str(i) + ', Params: ' + str(utils.count_parameters(model)))
#model = DCGAN(zdim=16, h=4, lr=2e-4).cuda()
trainer = MNISTTrainer(model, loader, datadir+str(i)+'/')
trainer.train(epochs=100)
import numpy as np
import os
import matplotlib.pyplot as plt
import SNN
import data
SAVE_PATH = os.getcwd() + '/weight_mnist'
mnist = data.MNIST(
path=["MNIST/t10k-images.idx3-ubyte", "MNIST/t10k-labels.idx1-ubyte"])
w1 = np.load(SAVE_PATH + '1.npy')
w2 = np.load(SAVE_PATH + '2.npy')
Ts = 1e-3
scale = 2
view_max = 2
l1 = SNN.SNNDiscrete(w1, Ts, scale)
l2 = SNN.SNNDiscrete(w2, Ts, scale)
correct = 0
for i in range(mnist.datasize):
xs, ys = mnist.next_batch(1, shuffle=True)
xs = (1 - xs[0, :]) / Ts
input_mat = np.zeros([784, int(1 / Ts * view_max)])
input_mat[range(784), xs.astype(int)] = 1
l1out = l1.forward(input_mat)
l2out = l2.forward(l1out)
(inputs, outputs)) * 2 - 1) * np.sqrt(2 / (inputs + outputs)))
if bias:
b = Tensor(np.zeros((outputs, )))
return w, b
return w
def accuracy(y, output):
correct = np.sum(np.argmax(y.value, -1) == np.argmax(output.value, -1))
total = y.shape[0]
return correct / total, correct, total
if __name__ == '__main__':
train = data.MNIST('train', one_hot=True)
test = data.MNIST('test', one_hot=True)
train_loader = data.DataLoader(train,
batch_size=32,
shuffle=True,
repeat=False)
test_loader = data.DataLoader(train,
batch_size=32,
shuffle=False,
repeat=False)
w1, b1 = Linear(28 * 28, 256)
w2, b2 = Linear(256, 10)
opt = bp.SGD([w1, b1, w2, b2], lr=1e-3)
## create a data loader for training
ds = Subset(MNISTDataset(x_train, y_train, classes, class_to_idx, transform=tforms.Compose(tforms_train)),
data.get_random_index(len(y_train), len(y_train), args.seed))
self.train = DataLoader(ds, batch_size=args.batch_size, shuffle=True, num_workers=args.n_workers)
## create a data loader for validation
ds = Subset(MNISTDataset(x_val, y_val, classes, class_to_idx, transform=tforms.Compose(tforms_val)),
data.get_random_index(len(y_val), len(y_val), args.seed))
self.val = DataLoader(ds, batch_size=args.batch_size, shuffle=True, num_workers=args.n_workers)
## create a data loader for test
ds = Subset(MNISTDataset(x_test, y_test, classes, class_to_idx, transform=tforms.Compose(tforms_test)),
data.get_random_index(len(y_test), len(y_test), args.seed))
self.test = DataLoader(ds, batch_size=args.batch_size, shuffle=True, num_workers=args.n_workers)
## print data statistics
print(f'#train = {len(self.train.dataset)}, #val = {len(self.val.dataset)}, #test = {len(self.test.dataset)}')
if __name__ == '__main__':
dsld = data.MNIST(types.SimpleNamespace(src='MNIST', batch_size=100, seed=0, n_workers=10))
## MNIST
#train = 60000
#val = 5000
#test = 5000
# help="When given, loads models from LOAD_PATH folder")
# parser.add_argument("--save_path", action="store", type=str, default=None,
# help="When given, saves models to LOAD_PATH folder after all epochs (or every epoch)")
# parser.add_argument("--save_every_epoch", action="store_true", default=False,
# help="When a save path is given, store the model after every epoch instead of only the last")
# parser.add_argument("--img_path", action="store", type=str, default=None,
# help="When given, saves samples to the given directory")
args = parser.parse_args()
output_path = util.output.init_experiment_output_dir("mnist", "gan", args)
dataset = data.MNIST("data/downloads/mnist",
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Lambda(lambda img: img * 2 - 1)
]))
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=12)
G = Generator28(args.l_size, args.h_size, args.use_mish)
D = Discriminator28(args.h_size,
use_bn=args.use_batchnorm_in_D,
use_mish=args.use_mish)
if args.cuda:
G = G.cuda()