def test_networks():
dirname = Path('results_networks')
dirname.mkdir(exist_ok=True)
image = np.array(Image.open('lena.png').convert('L'))
image_cuda = torch.tensor(image).float().cuda()[None, None, ...]
config = Config()
config.kn_kernel_size = 2
config.kn_num_channels = 256
config.kn_num_convs = 3
kn = KernelNet().cuda()
print(kn)
assert kn.kernel_cuda.shape == (1, 1, 19, 1)
assert kn(image_cuda).shape == (1, 1, 494, 512)
assert torch.isclose(torch.sum(kn.kernel_cuda), torch.tensor(1).float())
kn_dot = make_dot(image_cuda, kn)
kn_dot.render(dirname.joinpath('kn'))
fig = plt.figure()
kernel = kn.kernel.numpy().squeeze()
plt.plot(kernel)
fig.savefig(dirname.joinpath('kernel.png'))
lrd = LowResDiscriminator().cuda()
print(lrd)
assert lrd(image_cuda).shape == (1, 1, 502, 502)
lrd_dot = make_dot(image_cuda, lrd)
lrd_dot.render(dirname.joinpath('lrd'))
def test_double_input_model():
di = DoubleInputModel()
x1 = torch.randn(1, 8, 16, 16)
x2 = torch.randn(1, 8, 16, 16)
dot = make_dot((x1, x2), di)
dot.render('double_input', format='png')
def test_lstm():
x = torch.randn(1, 128)
lstm_cell = torch.nn.LSTMCell(128, 128)
dot = make_dot(x, lstm_cell)
dot.render('lstm')
def test_double_backprop_model():
x = torch.randn(1, 8)
dp = DoublePropModel()
dot = make_dot(x, dp)
dot.render('double_backprop')
def test_mlp():
x = torch.randn(1, 8)
mlp = MLP()
dot = make_dot(x, mlp)
dot.render('mlp')
test_acc += torch.sum(y_pred == acc_test)
test_loss += loss.data[0]
test_loss /= test_size
test_acc /= test_size
print('\tVal Loss: {:.4f} \t Acc: {:.4f}'.format(test_loss, test_acc))
print('-' * 100)
if __name__ == '__main__':
# main()
# train()
x = Variable(torch.rand(1, 3, 224, 224))
# resNet18 = ResNet(BasicBlock,[2, 2, 2, 2])
# resNet34 = ResNet(BasicBlock,[3,4,6,3])
resNet50 = ResNet(Bottleneck, [3, 4, 6, 3])
# resNet101 = ResNet(Bottleneck, [3, 4, 23, 3])
# with SummaryWriter(log_dir="./logs/troch",comment='resNet101') as w:
# w.add_graph(resNet101, (x, ))
y = resNet50(x)
# model = tv.models.resnet50()
# y = model(x)
print(y)
dot = make_dot(y, name="./images/resNet50",params=dict(list(resNet50.named_parameters()) + [('x', x)]))
# dot = make_dot(y, name="./images/resNet50",params=dict(list(model.named_parameters()) + [('x', x)]))
dot.format = "pdf"
dot.render()
#!/usr/bin/env python
import torch
from pytorchviz import make_dot
# from torchviz import make_dot
x = torch.randn(1, 2, 3)
class Minimum(torch.nn.Module):
def forward(self, x):
return x + 2
x.requires_grad = True
y = Minimum()(x)
print(y.grad_fn.next_functions)
dot = make_dot(x, Minimum())
# x.requires_grad = True
# model = Minimum()
# params = dict(model.named_parameters())
# params['x'] = x
# dot = make_dot(model(x), params)
dot.render('minimum')
#!/usr/bin/env python
import torch
from pytorch_unet import UNet
from pytorchviz import make_dot
class Wrapper(torch.nn.Module):
def __init__(self, net):
super().__init__()
self.net = net
def forward(self, x):
y = self.net(x)
if isinstance(y, dict):
y = tuple(y.values())
return y
x = torch.rand(1, 1, 16, 16, 16).cuda().float()
unet = UNet(1, 4, 3, 8, output_levels=[3, 0, 1]).cuda()
unet = Wrapper(unet)
print(unet)
y = unet(x)
dot = make_dot(x, unet)
dot.format = 'pdf'
dot.render('unet')