def test_multihead_head_selection(self):
# Check if the optimizer is updated correctly
# when heads are created and updated.
model = MultiHeadClassifier(in_features=6)
optimizer = SGD(model.parameters(), lr=1e-3)
criterion = CrossEntropyLoss()
benchmark = get_fast_benchmark(use_task_labels=True, shuffle=False)
strategy = Naive(model, optimizer, criterion,
train_mb_size=100, train_epochs=1,
eval_mb_size=100, device='cpu')
strategy.evaluator.loggers = [TextLogger(sys.stdout)]
# initialize head
strategy.train(benchmark.train_stream[0])
strategy.train(benchmark.train_stream[4])
# create models with fixed head
model_t0 = model.classifiers['0']
model_t4 = model.classifiers['4']
# check head task0
for x, y, t in DataLoader(benchmark.train_stream[0].dataset):
y_mh = model(x, t)
y_t = model_t0(x)
assert ((y_mh - y_t) ** 2).sum() < 1.e-7
break
# check head task4
for x, y, t in DataLoader(benchmark.train_stream[4].dataset):
y_mh = model(x, t)
y_t = model_t4(x)
assert ((y_mh - y_t) ** 2).sum() < 1.e-7
break
def __init__(
self,
num_layers=1,
in_features=784,
hidden_features_per_column=100,
adapter="mlp",
):
"""
:param num_layers: number of layers (default=1)
:param in_features: size of each input sample
:param hidden_features_per_column:
number of hidden units for each column
:param adapter: adapter type. One of {'linear', 'mlp'} (default='mlp')
"""
super().__init__()
assert num_layers >= 1
self.num_layers = num_layers
self.in_features = in_features
self.out_features_per_columns = hidden_features_per_column
self.layers = nn.ModuleList()
self.layers.append(PNNLayer(in_features, hidden_features_per_column))
for _ in range(num_layers - 1):
lay = PNNLayer(
hidden_features_per_column,
hidden_features_per_column,
adapter=adapter,
)
self.layers.append(lay)
self.classifier = MultiHeadClassifier(hidden_features_per_column)
def __init__(self,
num_layers=1,
in_features=784,
hidden_features_per_column=100,
adapter='mlp'):
""" Progressive Neural Network.
The model assumes that each experience is a separate task.
Multiple experiences with the same task label or multiple task labels
within the same experience will result in a runtime error.
:param num_layers: number of layers (default=1)
:param in_features: size of each input sample
:param hidden_features_per_column:
number of hidden units for each column
:param adapter: adapter type. One of {'linear', 'mlp'} (default='mlp')
"""
super().__init__()
assert num_layers >= 1
self.num_layers = num_layers
self.in_features = in_features
self.out_features_per_columns = hidden_features_per_column
self.layers = nn.ModuleList()
self.layers.append(PNNLayer(in_features, hidden_features_per_column))
for _ in range(num_layers - 1):
lay = PNNLayer(hidden_features_per_column,
hidden_features_per_column,
adapter=adapter)
self.layers.append(lay)
self.classifier = MultiHeadClassifier(hidden_features_per_column)
def __init__(self, num_classes=10, input_size=6, hidden_size=50):
super().__init__()
self.classifier = MultiHeadClassifier(self._hidden_size, num_classes)