def test_fully_customized(self):
"""Evaluate if ResNet of different sizes initializes and runs"""
custom_sparse_params = dict(
stem=LayerParams(),
filters64=[ # 3 blocks
dict(
conv1x1_1=SparseWeightsLayerParams(
percent_on=0.3,
boost_strength=1.2,
boost_strength_factor=1.0,
local=False,
weight_sparsity=0.3,
),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
],
filters128=[ # 4 blocks
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
],
filters256=[ # 6 blocks
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
],
filters512=[ # 3 blocks
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
],
linear=LayerParams(),
)
net = ResNet(config=dict(
depth=50, num_classes=10, sparse_params=custom_sparse_params))
net(Variable(torch.randn(2, 3, 32, 32)))
self.assertIsInstance(net, ResNet, "Loads ResNet50 fully customized")
def test_custom_per_group(self):
"""Evaluate ResNets customized per group"""
custom_sparse_params = dict(
stem=SparseWeightsLayerParams(),
filters64=dict(
conv1x1_1=SparseWeightsLayerParams(
percent_on=0.3,
boost_strength=1.2,
boost_strength_factor=1.0,
local=False,
weight_sparsity=0.3,
),
conv3x3_2=SparseWeightsLayerParams(
percent_on=0.1,
boost_strength=1.2,
boost_strength_factor=1.0,
local=True,
weight_sparsity=0.1,
),
conv1x1_3=SparseWeightsLayerParams(weight_sparsity=0.1),
shortcut=SparseWeightsLayerParams(percent_on=0.4,
weight_sparsity=0.4),
),
filters128=dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
filters256=dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
filters512=dict(
conv1x1_1=LayerParams(),
conv3x3_2=LayerParams(),
conv1x1_3=LayerParams(),
shortcut=LayerParams(),
),
linear=SparseWeightsLayerParams(weight_sparsity=0.5),
)
net = ResNet(config=dict(
depth=50, num_classes=10, sparse_params=custom_sparse_params))
net(Variable(torch.randn(2, 3, 32, 32)))
self.assertIsInstance(net, ResNet,
"Loads ResNet50 customized per group")