def test_get_early_stop_history_list_from_files(self):
""" Should load fake EarlyStopHistoryList from pth files. """
plan_fc = [2]
net0 = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
net1 = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
history_list = EarlyStopHistoryList()
history_list.setup(2, 0)
history_list.histories[0].state_dicts[0] = deepcopy(net0.state_dict())
history_list.histories[1].state_dicts[0] = deepcopy(net1.state_dict())
history_list.histories[0].indices[0] = 3
history_list.histories[1].indices[0] = 42
specs = get_specs_lenet_toy()
specs.save_early_stop = True
specs.net_count = 2
specs.prune_count = 0
with TemporaryDirectory() as tmp_dir_name:
# save checkpoints
result_saver.save_early_stop_history_list(tmp_dir_name, 'prefix',
history_list)
# load and validate histories from file
experiment_path_prefix = f"{tmp_dir_name}/prefix"
loaded_history_list = result_loader.get_early_stop_history_list_from_files(
experiment_path_prefix, specs)
self.assertEqual(loaded_history_list, history_list)
net0.load_state_dict(history_list.histories[0].state_dicts[0])
net1.load_state_dict(history_list.histories[1].state_dicts[0])
def test_save_nets(self):
""" Should save two small Lenet instances into pth files. """
plan_fc = [5]
net_list = [
Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc),
Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
]
with TemporaryDirectory() as tmp_dir_name:
result_saver.save_nets(tmp_dir_name, 'prefix', net_list)
# load and reconstruct nets from their files
result_file_path0 = os.path.join(tmp_dir_name, 'prefix-net0.pth')
result_file_path1 = os.path.join(tmp_dir_name, 'prefix-net1.pth')
for result_file_path in [result_file_path0, result_file_path1]:
with open(result_file_path, 'rb') as result_file:
checkpoint = t_load(result_file)
net = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
net.load_state_dict(checkpoint)
def generate_model_from_state_dict(state_dict, specs):
""" Generate a model specified by 'specs' and load the given 'state_dict'. """
net = Net(specs.net, specs.dataset, specs.plan_conv, specs.plan_fc)
net.load_state_dict(state_dict)
net.prune_net(
0., 0., reset=False) # apply pruned masks, but do not modify the masks
return net
def generate_randomly_reinitialized_net(specs, state_dict):
""" Build a net from 'state_dict' and randomly reinitialize its weights.
The net has the same masks like the net specified by 'state_dict'. """
assert isinstance(
specs, ExperimentSpecs
), f"'specs' needs to be ExperimentSpecs, but is {type(specs)}."
net = Net(specs.net, specs.dataset, specs.plan_conv, specs.plan_fc)
net.load_state_dict(state_dict)
net.apply(gaussian_glorot)
net.store_initial_weights()
net.prune_net(0.0, 0.0)
return net
def test_save_early_stop_history_list(self):
""" Should save two fake EarlyStopHistories into two pth files. """
plan_fc = [2]
net0 = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
net1 = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
history_list = EarlyStopHistoryList()
history_list.setup(2, 0)
history_list.histories[0].state_dicts[0] = deepcopy(net0.state_dict())
history_list.histories[1].state_dicts[0] = deepcopy(net1.state_dict())
history_list.histories[0].indices[0] = 3
history_list.histories[1].indices[0] = 42
with TemporaryDirectory() as tmp_dir_name:
result_saver.save_early_stop_history_list(
tmp_dir_name, 'prefix', history_list) # save checkpoints
# load and validate histories from file
result_file_path0 = os.path.join(tmp_dir_name,
'prefix-early-stop0.pth')
result_file_path1 = os.path.join(tmp_dir_name,
'prefix-early-stop1.pth')
for net_num, result_file_path in enumerate(
[result_file_path0, result_file_path1]):
with open(result_file_path, 'rb') as result_file:
reconstructed_hist = t_load(result_file)
net = Net(NetNames.LENET,
DatasetNames.MNIST,
plan_conv=[],
plan_fc=plan_fc)
np.testing.assert_array_equal(
reconstructed_hist.indices,
history_list.histories[net_num].indices)
net.load_state_dict(reconstructed_hist.state_dicts[0])