def test_swap_image_channel(self):
x = np.random.rand(1, 10, 10)
out = swap_image_channel(x)
self.assertTrue(out.shape == (10, 10, 1))
x = np.random.rand(1, 10, 10, 3)
out = swap_image_channel(x)
self.assertTrue(out.shape == (1, 3, 10, 10))
def test_DataContainer_CIFAR10(self):
dataname = 'CIFAR10'
p = {
'range': (0., 1.),
'num_classes': 10,
'type': 'image',
'dim': (3, 32, 32),
'mean': [0.49139947, 0.48215836, 0.44653094],
'std': [0.20230092, 0.1994128, 0.20096162],
'train_shape': (50000, 32, 32, 3),
'test_shape': (10000, 32, 32, 3),
}
dc = self.init_datacontainer(dataname)
# basic props
mu = dc.train_mean
np.testing.assert_array_almost_equal(mu, p['mean'])
std = dc.train_std
np.testing.assert_array_almost_equal(std, p['std'])
# train set
train_np = dc.x_train
self.assertEqual(train_np.shape, p['train_shape'])
y_np = dc.y_train
self.assertEqual(y_np.shape, (p['train_shape'][0], ))
train_loader = dc.get_dataloader(batch_size=8,
is_train=True,
shuffle=False)
self.assertEqual(len(train_loader.dataset), p['train_shape'][0])
x_pt, y_pt = next(iter(train_loader))
self.assertEqual(x_pt.size(), tuple([8] + list(p['dim'])))
self.assertEqual(y_pt.size(), (8, ))
x1 = train_np[:8]
x2 = swap_image_channel(x_pt.cpu().detach().numpy())
np.testing.assert_equal(x1, x2)
y1 = y_np[:8]
y2 = y_pt[:8].cpu().detach().numpy()
np.testing.assert_equal(y1, y2)
# test set
test_np = dc.x_test
self.assertEqual(test_np.shape, p['test_shape'])
y_np = dc.y_test
self.assertEqual(y_np.shape, (p['test_shape'][0], ))
test_loader = dc.get_dataloader(batch_size=8,
is_train=False,
shuffle=False)
self.assertEqual(len(test_loader.dataset), p['test_shape'][0])
x_pt, y_pt = next(iter(test_loader))
self.assertEqual(x_pt.size(), tuple([8] + list(p['dim'])))
self.assertEqual(y_pt.size(), (8, ))
x1 = test_np[:8]
x2 = swap_image_channel(x_pt.cpu().detach().numpy())
np.testing.assert_equal(x1, x2)
y1 = y_np[:8]
y2 = y_pt[:8].cpu().detach().numpy()
np.testing.assert_equal(y1, y2)
def test_DataContainer_SVHN(self):
dataname = 'SVHN'
p = {
'range': (0., 1.),
'num_classes': 10,
'type': 'image',
'dim': (3, 32, 32),
'mean': [0.43768215, 0.4437698, 0.47280422],
'std': [0.12008651, 0.12313706, 0.10520393],
'train_shape': (73257, 32, 32, 3),
'test_shape': (26032, 32, 32, 3),
}
dc = self.init_datacontainer(dataname)
# train set
train_np = dc.x_train
self.assertEqual(train_np.shape, p['train_shape'])
y_np = dc.y_train
self.assertEqual(y_np.shape, (p['train_shape'][0], ))
train_loader = dc.get_dataloader(batch_size=8,
is_train=True,
shuffle=False)
self.assertEqual(len(train_loader.dataset), p['train_shape'][0])
x_pt, y_pt = next(iter(train_loader))
self.assertEqual(x_pt.size(), tuple([8] + list(p['dim'])))
self.assertEqual(y_pt.size(), (8, ))
x1 = train_np[:8]
x2 = swap_image_channel(x_pt.cpu().detach().numpy())
np.testing.assert_equal(x1, x2)
y1 = y_np[:8]
y2 = y_pt[:8].cpu().detach().numpy()
np.testing.assert_equal(y1, y2)
def test_load(self):
self.mc.save(os.path.join('test', 'test_mnist'), overwrite=True)
full_path = os.path.join('save', 'test', 'test_mnist.pt')
self.mc2.load(full_path)
x = torch.tensor(swap_image_channel(self.x)).to(self.mc.device)
s1 = self.mc.model(x).cpu().detach().numpy()
s2 = self.mc2.model(x).cpu().detach().numpy()
np.testing.assert_almost_equal(s1, s2)
def test_predict_one(self):
x = torch.tensor(swap_image_channel(self.x[0])).to(self.mc.device)
p1 = self.mc.predict_one(x)
p2 = self.mc.predict_one(self.x[0])
self.assertTrue((p1 == p2).all())
def test_DataContainer_MNIST(self):
dataname = 'MNIST'
p = {
'range': (0., 1.),
'num_classes': 10,
'type': 'image',
'dim': (1, 28, 28),
'mean': [0.13066046],
'std': [0.30150425],
'train_shape': (60000, 28, 28, 1),
'test_shape': (10000, 28, 28, 1)
}
dc = self.init_datacontainer(dataname)
# basic props
r = dc.data_range
self.assertEqual(r, p['range'])
name = dc.name
self.assertEqual(name, dataname)
m = dc.num_classes
self.assertEqual(m, p['num_classes'])
dtype = dc.data_type
self.assertEqual(dtype, p['type'])
dim = dc.dim_data
self.assertEqual(dim, p['dim'])
mu = dc.train_mean
np.testing.assert_array_almost_equal(mu, p['mean'])
std = dc.train_std
np.testing.assert_array_almost_equal(std, p['std'])
# train set
train_np = dc.x_train
self.assertEqual(train_np.shape, p['train_shape'])
y_np = dc.y_train
self.assertEqual(y_np.shape, (p['train_shape'][0], ))
train_loader = dc.get_dataloader(batch_size=8,
is_train=True,
shuffle=False)
self.assertEqual(len(train_loader.dataset), p['train_shape'][0])
x_pt, y_pt = next(iter(train_loader))
self.assertEqual(x_pt.size(), tuple([8] + list(p['dim'])))
self.assertEqual(y_pt.size(), (8, ))
x1 = train_np[:8]
x2 = swap_image_channel(x_pt.cpu().detach().numpy())
np.testing.assert_equal(x1, x2)
y1 = y_np[:8]
y2 = y_pt[:8].cpu().detach().numpy()
np.testing.assert_equal(y1, y2)
# test set
test_np = dc.x_test
self.assertEqual(test_np.shape, p['test_shape'])
y_np = dc.y_test
self.assertEqual(y_np.shape, (p['test_shape'][0], ))
test_loader = dc.get_dataloader(batch_size=8,
is_train=False,
shuffle=False)
self.assertEqual(len(test_loader.dataset), p['test_shape'][0])
x_pt, y_pt = next(iter(test_loader))
self.assertEqual(x_pt.size(), tuple([8] + list(p['dim'])))
self.assertEqual(y_pt.size(), (8, ))
x1 = test_np[:8]
x2 = swap_image_channel(x_pt.cpu().detach().numpy())
np.testing.assert_equal(x1, x2)
y1 = y_np[:8]
y2 = y_pt[:8].cpu().detach().numpy()
np.testing.assert_equal(y1, y2)