def run(self):
try:
path = replace_standard_paths('{path_tmp}/tmp.run_config')
self.interactive.to_json(path)
Main.new_task(path).run()
except Exception as e:
LoggerManager().get_logger().error(str(e), exc_info=e)
tkm.showwarning(message=str(e))
def test_deterministic_method(self):
"""
make sure that using the same seed results in the same outcome when:
- using deterministic cudnn settings
- using a single worker for data loading
note that this does not test:
- the lightning trainer
- multi gpu setups, ddp distribution
- advanced augmentation strategies, e.g. mixup
"""
for i, config in enumerate([
super1_fairnas,
super3,
search_darts_config,
retrain_darts_cifar_config,
]):
for seed in range(2):
args_changes = {
"{cls_task}.seed": seed,
"{cls_task}.is_deterministic": True,
"{cls_task}.is_test_run": True,
"{cls_task}.save_del_old": True,
"{cls_device}.num_devices": 1,
"{cls_device}.use_cudnn_benchmark": False,
"cls_trainer": "SimpleTrainer",
"{cls_trainer}.max_epochs": 1,
"cls_data": "Cifar10Data",
"{cls_data}.fake": False,
"{cls_data}.batch_size_train": 2,
"{cls_data}.batch_size_test": -1,
"{cls_data}.dir": "{path_data}/cifar_data/",
"{cls_data}.download": True,
"{cls_data}.num_workers": 0,
"{cls_schedulers#0}.warmup_epochs": 0,
}
exp1 = Main.new_task(config,
args_changes=args_changes.copy()).run()
name = exp1.get_method().__class__.__name__
data = exp1.get_method().data_set.sample_random_data(
batch_size=4).cuda()
outputs1 = [
o.clone().detach() for o in exp1.get_method()(data)
]
del exp1
exp2 = Main.new_task(config, args_changes=args_changes).run()
outputs2 = [
o.clone().detach() for o in exp2.get_method()(data)
]
del exp2
for o1, o2 in zip(outputs1, outputs2):
self._assert_same_tensors(
'i=%d seed=%d method=%s' % (i, seed, name), o1, o2)
def test_search_param_updates(self):
"""
set gradients of network/architecture weights to zero and ensure that they are not updated, while the others are
expecting two arc weights:
28x8, with gradients, the actual weights
28x8, not requiring gradients, only masks
"""
for mask_idx in range(2):
exp1 = Main.new_task(
search_darts_config,
args_changes={
"test_mask_idx": mask_idx,
"cls_method": "TestMaskGradientsSearchMethod",
"{cls_task}.is_test_run": True,
"{cls_trainer}.max_epochs": 1,
"cls_data": "Cifar10Data",
"{cls_data}.fake": True,
"{cls_data}.download": False,
"{cls_data}.batch_size_train": 2,
"{cls_data}.batch_size_test": -1,
"{StackedCellsNetworkBody}.cell_order": "n, r, n, r, n",
"cls_optimizers": "DebugOptimizer, DebugOptimizer",
# "cls_optimizers": "SGDOptimizer, SGDOptimizer",
"{cls_optimizers#0}.weight_decay": 0.0,
"{cls_optimizers#1}.weight_decay": 0.0,
},
raise_unparsed=False)
# get initial weights, make copies
optimizers, _ = exp1.get_method().configure_optimizers()
assert len(optimizers) == 1, "expecting only one multi optimizer"
weights = [
opt.param_groups[0]['params'][0]
for opt in optimizers[0].optimizers
]
weight_copies = [w.clone().detach().cpu() for w in weights]
s = ['Network', 'Architecture']
for i, w in enumerate(weights):
print('%s sample-weight shape:' % s[i], w.shape)
# run, thus change weights
exp1.run()
for i, (w, w_copy) in enumerate(zip(weights, weight_copies)):
diff = (w.cpu() - w_copy).abs().sum().item()
if i == mask_idx:
if diff > 0.00001:
print(w[0:3, 0:3])
print(w_copy[0:3, 0:3])
assert False, '%s gradients were masked but weights changed anyway; diff: %s' % (
s[i], diff)
else:
if diff < 0.00001:
print(w[0:3, 0:3])
print(w_copy[0:3, 0:3])
assert False, '%s gradients were not masked but weights did not change; diff: %s' % (
s[i], diff)
del exp1
def assert_stats_match(name,
task_cfg,
cfg: dict,
num_params=None,
num_macs=None):
cfg_dir = replace_standard_paths('{path_tmp}/tests/cfgs/')
cfg_path = Builder.save_config(cfg, cfg_dir, name)
exp = Main.new_task(
task_cfg,
args_changes={
'{cls_data}.fake': True,
'{cls_data}.batch_size_train': 2,
'{cls_data}.batch_size_test': -1,
'{cls_task}.is_test_run': True,
'{cls_task}.save_dir': '{path_tmp}/tests/workdir/',
"{cls_network}.config_path": cfg_path,
"{cls_trainer}.ema_decay": -1,
'cls_network_heads':
'ClassificationHead', # necessary for the DARTS search space to disable the aux heads
},
raise_unparsed=False)
net = exp.get_method().get_network()
macs = exp.get_method().profile_macs()
net.eval()
# print(net)
cp = count_parameters(net)
if num_params is not None:
assert cp == num_params, 'Got unexpected num params for %s: %d, expected %d, diff: %d'\
% (name, cp, num_params, abs(cp - num_params))
if num_macs is not None:
assert macs == num_macs, 'Got unexpected num macs for %s: %d, expected %d, diff: %d'\
% (name, macs, num_macs, abs(macs - num_macs))
def generate_from_name(name: str, save=True, verbose=True):
genotype, compact = compact_from_name(name, verbose=verbose)
run_configs = '{path_conf_tasks}/d1_dartsv1.run_config, {path_conf_net_search}darts.run_config'
# create weight sharing cell model
changes = {
'cls_data':
'Cifar10Data',
'{cls_data}.fake':
True,
'{cls_task}.save_del_old':
False,
'{cls_network_body}.cell_order':
'n, r',
'{cls_network_body}.features_first_cell':
36 * 4,
'{cls_network_stem}.features':
36 * 3,
'cls_network_cells_primitives':
"%s, %s" % (compact.get('primitives'), compact.get('primitives')),
}
task = Main.new_task(run_configs, args_changes=changes)
net = task.get_method().get_network()
args = task.args
wss = StrategyManager().get_strategies()
assert len(wss) == 1
ws = wss[list(wss.keys())[0]]
# fix arc, all block inputs use different weights
# go through all weights in the search cell
for n, w in ws.named_parameters_single():
# figure out cell type ("normal", "reduce"), block index, and if it's the first, second, ... op of that block
c_type, block_idx, num_inputs, num_idx = n.split('/')[-4:]
block_idx = int(block_idx.split('-')[-1])
num_idx = int(num_idx.split('-')[-1])
# set all paths weights to zero
w.data.zero_()
# go through the cell description of the genotype, if input and op number match, set the weight to be higher
for op_idx, from_idx in compact.get(c_type)[block_idx]:
if num_idx == from_idx:
w[op_idx] = 1
ws.forward()
# saving config now will only use the highest weighted connections, since we have a search network
cfg = net.config(finalize=True, num_block_ops=2)
if save:
path = Builder.save_config(cfg, get_net_config_dir(genotype.source),
name)
print('Saved config: %s' % path)
return net, cfg, args
def visualize_config(config: dict, save_path: str):
save_path = replace_standard_paths(save_path)
cfg_path = Builder.save_config(config, replace_standard_paths('{path_tmp}/viz/'), 'viz')
exp = Main.new_task(run_config, args_changes={
'{cls_data}.fake': True,
'{cls_data}.batch_size_train': 4,
'{cls_task}.is_test_run': True,
'{cls_task}.save_dir': '{path_tmp}/viz/task/',
'{cls_task}.save_del_old': True,
"{cls_network}.config_path": cfg_path,
})
net = exp.get_method().get_network()
vt = VizTree(net)
vt.print()
vt.plot(save_path + 'net', add_subgraphs=True)
print('Saved cell viz to %s' % save_path)
def visualize_config(config: dict, save_path: str):
save_path = replace_standard_paths(save_path)
cfg_path = Builder.save_config(config, replace_standard_paths('{path_tmp}/viz/'), 'viz')
exp = Main.new_task(run_config, args_changes={
'{cls_data}.fake': True,
'{cls_data}.batch_size_train': 2,
'{cls_task}.is_test_run': True,
'{cls_task}.save_dir': '{path_tmp}/viz/task/',
'{cls_task}.save_del_old': True,
"{cls_task}.note": "viz",
"{cls_network}.config_path": cfg_path,
})
net = exp.get_method().get_network()
for s in ['n', 'r']:
for cell in net.get_cells():
if cell.name.startswith(s):
visualize_cell(cell, save_path, s)
break
print('Saved cell viz to %s' % save_path)
def generate(self,
save_dir: str,
name: str,
verbose=True) -> (nn.Module, dict, Namespace):
run_configs = '{path_conf_tasks}/s1_random.run_config, {path_conf_net_search}/%s.run_config' % self.search_net
task = Main.new_task(run_configs,
args_changes={
'{cls_data}.fake': True,
'{cls_task}.save_del_old': False,
'{cls_task}.save_dir': '{path_tmp}/generate/',
'{cls_trainer}.ema_decay': -1,
})
net = task.get_method().get_network()
args = task.args
# fix arc
net.forward_strategy(fixed_arc=self.gene)
cfg = self.save(net, save_dir, name, verbose)
return net, cfg, args
# "{cls_initializers#0}.strict": True,
"cls_criterion": "CrossEntropyCriterion",
"cls_optimizers": "SGDOptimizer",
"{cls_optimizers#0}.lr": 0.00,
"{cls_optimizers#0}.momentum": 0.9,
"{cls_optimizers#0}.accumulate_batches": 2,
"cls_schedulers": "CosineScheduler",
"{cls_schedulers#0}.warmup_epochs": 2,
"{cls_schedulers#0}.warmup_lr": 0.01,
"cls_regularizers": "DropOutRegularizer",
# "cls_regularizers": "DropOutRegularizer, DropPathRegularizer",
"{cls_regularizers#0}.prob": 0.0,
# "{cls_regularizers#1}.max_prob": 0.1,
# "{cls_method}.amp_enabled": False,
# "{cls_optimizers#0}.weight_decay": 4e-3,
# "{cls_optimizers#0}.weight_decay_filter": True,
}
if __name__ == "__main__":
# ignore the command line, use "args" instead
task = Main.new_task([], args_changes=args)
# print(task.get_method().get_network())
task.load()
# task.load('{path_tmp}/s3_2/')
task.run()
config_files = "{path_conf_tasks}/d1_dartsv1.run_config, {path_conf_net_search}/bench201.run_config"
# config_files = "{path_conf_tasks}/d1_asap.run_config, {path_conf_net_search}/bench201.run_config"
# config_files = "{path_conf_tasks}/d1_gdas.run_config, {path_conf_net_search}/bench201.run_config"
# config_files = "{path_conf_tasks}/d1_mdenas.run_config, {path_conf_net_search}/bench201.run_config"
changes = {
"{cls_task}.is_test_run": True,
"{cls_task}.save_dir": "{path_tmp}/run_bench_d1/",
"{cls_task}.save_del_old": True,
"cls_benchmark": "MiniNASTabularBenchmark",
"{cls_benchmark}.path": "{path_data}/nats_bench_1.1_mini.pt",
"{cls_benchmark}.default_data_set": "cifar100",
"{cls_benchmark}.default_result_type": "test",
"{cls_trainer}.max_epochs": 4,
# "cls_data": "Cifar10Data",
"{cls_data}.fake": True,
"{cls_data}.dir": "{path_data}/ImageNet_ILSVRC2012/",
"{cls_data}.batch_size_train": 2,
"{cls_data}.batch_size_test": -1,
}
if __name__ == "__main__":
task = Main.new_task(config_files, args_changes=changes)
task.load()
task.run()
def test_model_save_load(self):
"""
make sure that saving+loading for methods/models works correctly, so that we get the same outputs after loading
"""
for i, (config, trainer, ds, change_net, fix_topology) in enumerate([
(dna1_config, 'SimpleTrainer', 'Imagenet1000Data', False, True),
(super1_fairnas, 'SimpleTrainer', 'Imagenet1000Data', False, True),
(search_darts_config, 'SimpleTrainer', 'Cifar10Data', True, False),
(retrain_darts_cifar_config, 'SimpleTrainer', 'Cifar10Data', True,
False),
# (search_darts_config, 'LightningTrainer'),
# (retrain_darts_cifar_config, 'LightningTrainer'),
]):
save_dir = "{path_tmp}/tests/%d/"
arg_changes = {
"cls_data": ds,
"{cls_data}.fake": True,
"{cls_data}.batch_size_train": 4,
"{cls_data}.batch_size_test": 4,
"cls_trainer": trainer,
"{cls_trainer}.max_epochs": 2,
"{cls_task}.seed": 0,
"{cls_task}.is_test_run": True,
"{cls_task}.save_dir": save_dir % 1,
"{cls_task}.save_del_old": True,
"{cls_schedulers#0}.warmup_epochs": 0,
}
if change_net:
arg_changes.update({
"{cls_network_body}.features_first_cell": 8,
"{cls_network_body}.cell_order": "n, r, n, r, n",
"{cls_network_stem}.features": 4,
})
print(config)
exp1 = Main.new_task(config, args_changes=arg_changes).run()
data = exp1.get_method().get_data_set().sample_random_data(
batch_size=4).cuda()
net = exp1.get_method().get_network()
if fix_topology and isinstance(net, SearchUninasNetwork):
net.set_forward_strategy(False)
net.get_strategy_manager().forward_const(0)
with torch.no_grad():
outputs1 = exp1.get_method()(data)
arg_changes["{cls_task}.save_dir"] = save_dir % 2
arg_changes["{cls_task}.seed"] += 1
exp2 = Main.new_task(config, args_changes=arg_changes).run().load(
save_dir % 1)
net = exp2.get_method().get_network()
if fix_topology and isinstance(net, SearchUninasNetwork):
net.set_forward_strategy(False)
net.get_strategy_manager().forward_const(0)
with torch.no_grad():
outputs2 = exp2.get_method()(data)
for o1, o2 in zip(outputs1, outputs2):
self._assert_same_tensors(
'i=%d method=%s' %
(i, exp1.get_method().__class__.__name__), o1, o2)