def test_torch_fpgm_pruner(self):
"""
With filters(kernels) weights defined as above (w), it is obvious that w[4] and w[5] is the Geometric Median
which minimize the total geometric distance by defination of Geometric Median in this paper:
Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration,
https://arxiv.org/pdf/1811.00250.pdf
So if sparsity is 0.2, the expected masks should mask out w[4] and w[5], this can be verified through:
`all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([45., 45., 45., 45., 0., 0., 45., 45., 45., 45.]))`
If sparsity is 0.6, the expected masks should mask out w[2] - w[7], this can be verified through:
`all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([45., 45., 0., 0., 0., 0., 0., 0., 45., 45.]))`
"""
model = TorchModel()
config_list = [{'sparsity': 0.2, 'op_types': ['Conv2d']}, {'sparsity': 0.6, 'op_types': ['Conv2d']}]
pruner = torch_compressor.FPGMPruner(model, config_list)
model.conv2.weight.data = torch.tensor(w).float()
layer = torch_compressor.compressor.LayerInfo('conv2', model.conv2)
masks = pruner.calc_mask(layer, config_list[0])
assert all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([45., 45., 45., 45., 0., 0., 45., 45., 45., 45.]))
pruner.update_epoch(1)
model.conv2.weight.data = torch.tensor(w).float()
masks = pruner.calc_mask(layer, config_list[1])
assert all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([45., 45., 0., 0., 0., 0., 0., 0., 45., 45.]))
def test_torch_fpgm_pruner(self):
"""
With filters(kernels) weights defined as above (w), it is obvious that w[4] and w[5] is the Geometric Median
which minimize the total geometric distance by defination of Geometric Median in this paper:
Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration,
https://arxiv.org/pdf/1811.00250.pdf
So if sparsity is 0.2, the expected masks should mask out w[4] and w[5], this can be verified through:
`all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([125., 125., 125., 125., 0., 0., 125., 125., 125., 125.]))`
If sparsity is 0.6, the expected masks should mask out w[2] - w[7], this can be verified through:
`all(torch.sum(masks, (1, 2, 3)).numpy() == np.array([125., 125., 0., 0., 0., 0., 0., 0., 125., 125.]))`
"""
w = np.array([np.ones((5, 5, 5)) * (i + 1)
for i in range(10)]).astype(np.float32)
model = TorchModel()
config_list = [{
'sparsity': 0.6,
'op_types': ['Conv2d']
}, {
'sparsity': 0.2,
'op_types': ['Conv2d']
}]
pruner = torch_compressor.FPGMPruner(
model, config_list, torch.optim.SGD(model.parameters(), lr=0.01))
model.conv2.module.weight.data = torch.tensor(w).float()
masks = pruner.calc_mask(model.conv2)
assert all(
torch.sum(masks['weight_mask'], (1, 2, 3)).numpy() == np.array(
[125., 125., 125., 125., 0., 0., 125., 125., 125., 125.]))
model.conv2.module.weight.data = torch.tensor(w).float()
model.conv2.if_calculated = False
model.conv2.config = config_list[0]
masks = pruner.calc_mask(model.conv2)
assert all(
torch.sum(masks['weight_mask'], (1, 2, 3)).numpy() == np.array(
[125., 125., 0., 0., 0., 0., 0., 0., 125., 125.]))
def test_torch_fpgm_pruner(self):
model = TorchMnist()
configure_list = [{'sparsity': 0.5, 'op_types': ['Conv2d']}]
torch_compressor.FPGMPruner(model, configure_list).compress()
