def check_dataset(dataset, dataroot, augment, download):
if dataset == "cifar10":
cifar10 = get_CIFAR10(augment, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = cifar10
if dataset == "svhn":
svhn = get_SVHN(augment, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = svhn
if dataset == "mnist":
mnist = get_MNIST(augment, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = mnist
return input_size, num_classes, train_dataset, test_dataset
def check_dataset(dataset, dataroot, augment, download):
if dataset == 'cifar10':
cifar10 = get_CIFAR10(augment, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = cifar10
if dataset == 'svhn':
svhn = get_SVHN(augment, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = svhn
if dataset == 'mnist':
mnist = get_MNIST(False, dataroot, download)
input_size, num_classes, train_dataset, test_dataset = mnist
return input_size, num_classes, train_dataset, test_dataset
def main(args,kwargs):
output_folder = args.output_folder
model_name = args.model_name
with open(os.path.join(output_folder,'hparams.json')) as json_file:
hparams = json.load(json_file)
image_shape, num_classes, _, test_mnist = get_MNIST(False, hparams['dataroot'], hparams['download'])
test_loader = data.DataLoader(test_mnist, batch_size=32,
shuffle=False, num_workers=6,
drop_last=False)
x, y = test_loader.__iter__().__next__()
x = x.to(device)
model = Glow(image_shape, hparams['hidden_channels'], hparams['K'], hparams['L'], hparams['actnorm_scale'],
hparams['flow_permutation'], hparams['flow_coupling'], hparams['LU_decomposed'], num_classes,
hparams['learn_top'], hparams['y_condition'], False if 'logittransform' not in hparams else hparams['logittransform'],False if 'sn' not in hparams else hparams['sn'])
model.load_state_dict(torch.load(os.path.join(output_folder, model_name)))
model.set_actnorm_init()
model = model.to(device)
model = model.eval()
with torch.no_grad():
# ipdb.set_trace()
images = model(y_onehot=None, temperature=1, batch_size=32, reverse=True).cpu()
better_dup_images = model(y_onehot=None, temperature=1, z=model._last_z, reverse=True, use_last_split=True).cpu()
dup_images = model(y_onehot=None, temperature=1, z=model._last_z, reverse=True).cpu()
worse_dup_images = model(y_onehot=None, temperature=1, z=model._last_z, reverse=True).cpu()
l2_err = torch.pow((images - dup_images).view(images.shape[0], -1), 2).sum(-1).mean()
better_l2_err = torch.pow((images - better_dup_images).view(images.shape[0], -1), 2).sum(-1).mean()
worse_l2_err = torch.pow((images - worse_dup_images).view(images.shape[0], -1), 2).sum(-1).mean()
print(l2_err, better_l2_err, worse_l2_err)
plot_imgs([images, dup_images, better_dup_images, worse_dup_images], '_recons')
#
with torch.no_grad():
# ipdb.set_trace()
z, nll, y_logits = model(x, None)
better_dup_images = model(y_onehot=None, temperature=1, z=z, reverse=True, use_last_split=True).cpu()
plot_imgs([x, better_dup_images], '_data_recons2')
fpath = os.path.join(output_folder, '_recon_evoluation.png')
pad = run_recon_evolution(model, x, fpath)
import matplotlib.pyplot as plt
from datasets import get_CIFAR10, get_SVHN, get_MNIST, postprocess
from model import Glow
import ipdb
import os
device = torch.device("cuda")
output_folder = '/scratch/gobi2/wangkuan/glow/db-gan'
# output_folder = '/scratch/gobi2/wangkuan/glow/mnist-1x1-affine-512-1e-2'
model_name = 'glow_model_1.pth'
with open(os.path.join(output_folder, 'hparams.json')) as json_file:
hparams = json.load(json_file)
image_shape, num_classes, _, test_mnist = get_MNIST(False, hparams['dataroot'],
hparams['download'])
model = Glow(
image_shape, hparams['hidden_channels'], hparams['K'], hparams['L'],
hparams['actnorm_scale'], hparams['flow_permutation'],
hparams['flow_coupling'], hparams['LU_decomposed'], num_classes,
hparams['learn_top'], hparams['y_condition'],
False if 'logittransform' not in hparams else hparams['logittransform'])
model.load_state_dict(torch.load(os.path.join(output_folder, model_name)))
model.set_actnorm_init()
model = model.to(device)
model = model.eval()
network.add_synapses(Synapses(
network.groups['Ae'],
network.groups['Ai'],
w=torch.diag(c_excite * torch.ones_like(torch.Tensor(n_neurons)))),
source='Ae',
target='Ai')
network.add_synapses(Synapses(network.groups['Ai'], network.groups['Ae'], w=-c_inhib * \
(torch.ones_like(torch.Tensor(n_neurons, n_neurons)) - torch.diag(1 \
* torch.ones_like(torch.Tensor(n_neurons))))), source='Ai', target='Ae')
# network.add_monitor(Monitor(obj=network.groups['Ae'], state_vars=['v', 'theta']), name=('Ae', ('v', 'theta')))
# network.add_monitor(Monitor(obj=network.groups['Ai'], state_vars=['v']), name=('Ai', 'v'))
# Get training or test data from disk.
if mode == 'train':
data = get_MNIST(train=True)
elif mode == 'test':
data = get_MNIST(train=False)
X, y = data['X'], data['y']
# Count spikes from each neuron on each example (between update intervals).
outputs = torch.zeros_like(torch.Tensor(update_interval, n_neurons))
# Network simulation times.
image_time = time
rest_time = rest
# Voting schemes and neuron label assignments.
voting_schemes = ['all']
rates = torch.zeros_like(torch.Tensor(n_neurons, 10))