def test_recipe_seed(self, dataset_name, model_name, init_schema,
optimizer_name, tmp_path):
experiment_dir = tmp_path / 'experiment'
experiment = Experiment(directory=experiment_dir)
initial_state = experiment.spawn_new_tree(
dataset_name=dataset_name,
model_name=model_name,
init_schema=init_schema,
optimizer_name=optimizer_name,
seed=999,
)
assert initial_state.get().seed == 999
r = Recipe(train={"n_epochs": 1})
assert r.seed is None
new_state = r(initial_state)
assert r.seed == initial_state.get().seed
logger = logging.getLogger("main")
utils.setup_logging(debug=True)
directory = "../output/08-06-19_seed1"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="models.LeNet",
init_schema="",
optimizer_name="sgd",
seed=1,
)
x = Recipe(train={"n_epochs": 30})(x)
for _ in range(20):
# finetune
pruned = Recipe(
prune_schema="../schemas/pruning_schema_lenet_structuredrandom.py", )(
x)
x = Recipe(
reinit_schema="../schemas/reinit_schema_lt_lenet.py",
train={"n_epochs": 30},
)(pruned) # LT
# x = Recipe(
# train={"n_epochs": 20},
# )(pruned) # finetuning
logger = logging.getLogger("main")
utils.setup_logging(debug=True)
directory = "../output/vgg11_cifar10_seed4"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="cifar-10",
model_name="torchvision.models.vgg11",
init_schema="",
optimizer_name="sgd",
seed=4,
)
x = Recipe(train={"n_epochs": 30})(x)
for _ in range(20):
# finetune
pruned = Recipe(
prune_schema="../schemas/pruning_schema_vgg11_unstructuredl1.py", )(x)
x = Recipe(
reinit_schema="../schemas/reinit_schema_lt_vgg11.py",
train={"n_epochs": 30},
)(pruned) # LT
# x = Recipe(
# train={"n_epochs": 20},
# )(pruned) # finetuning
experiment.run()
import os
import utils
logger = logging.getLogger("main")
utils.setup_logging(debug=True)
directory = "../output/08-06-19_seed3"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="models.LeNet",
init_schema="",
optimizer_name="sgd",
seed=3,
)
x = Recipe(train={"n_epochs": 30})(x)
for _ in range(20):
pruned = Recipe(
prune_schema="../schemas/pruning_schema_lenet_unstructuredl1.py", )(x)
# x = Recipe(
# reinit_schema="../schemas/reinit_schema_lt_lenet.py",
# train={"n_epochs": 30},
# )(pruned) # LT
x = Recipe(train={"n_epochs": 30}, )(pruned) # finetuning
experiment.run()
# This materializes immediately
initial_state = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="models.LeNet",
init_schema="schemas/init_kaiminguniform_lenet.py",
optimizer_name="sgd",
seed=0,
device=torch.device("cpu"),
)
# These are computed lazily until .run() is called. A Recipe represents a
# way to transition from one state (read: model) to the following via
# pruning, finetuning, and reinitializing
state1 = Recipe(
train={"n_epochs": 1},
prune_schema="schemas/pruning_schema_lenet_unstructuredl1.py",
)(initial_state)
state2 = Recipe(reinit_schema="schemas/reinit_schema_lenet.py")(state1)
state3 = Recipe(
train={"n_epochs": 1},
prune_schema="schemas/pruning_schema_lenet_unstructuredl1.py",
reinit_schema="schemas/reinit_schema_lenet.py",
)(state2)
final_state = Recipe(train={"n_epochs": 1})(state3)
# When .run() is called, we execute the full experiment graph
experiment.run()
logger = logging.getLogger("main")
utils.setup_logging(debug=True)
directory = "../output/alexnet_seed4"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="torchvision.models.AlexNet",
init_schema="",
optimizer_name="sgd",
seed=4,
)
x = Recipe(train={"n_epochs": 30})(x)
for _ in range(20):
# finetune
pruned = Recipe(
prune_schema="../schemas/pruning_schema_alexnet_mixedl1.py", )(x)
x = Recipe(
reinit_schema="../schemas/reinit_schema_lt_alexnet.py",
train={"n_epochs": 30},
)(pruned) # LT
# x = Recipe(
# train={"n_epochs": 20},
# )(pruned) # finetuning
experiment.run()
utils.setup_logging(debug=True)
directory = "../output/lenet1ep_seed4" #"../output/08-06-19_seed0"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="models.LeNet",
init_schema="",
optimizer_name="sgd",
seed=4,
)
x = Recipe(
train={"n_epochs": 1} # 30
)(x)
# new
y = Recipe(
train={"n_epochs": 29}
)(x)
for _ in range(20):
# finetune
pruned = Recipe(
prune_schema="../schemas/pruning_schema_lenet_unstructuredl1.py",
)(x)
x = Recipe(
reinit_schema="../schemas/reinit_schema_lt_lenet.py",
train={"n_epochs": 1}, # 30
utils.setup_logging(debug=True)
directory = "../output/08-06-19_seed4"
experiment = Experiment(directory=directory)
# this materializes immediately
x = experiment.spawn_new_tree(
dataset_name="mnist",
model_name="models.LeNet",
init_schema="",
optimizer_name="sgd",
seed=4,
)
x = Recipe(
train={"n_epochs": 30}
)(x)
for _ in range(10):
# finetune
pruned = Recipe(
prune_schema="../schemas/pruning_schema_lenet_unstructuredl1.py",
name='halfhalfR',
)(x)
x = Recipe(
reinit_schema="../schemas/reinit_schema_lt_lenet.py",
train={"n_epochs": 30},
name='halfhalfR',
)(pruned)
for _ in range(10):
