def create_hook():
# With the following SaveConfig, we will save tensors for every 100 steps
save_config = SaveConfig(save_interval=100)
# Create a hook that logs weights, biases and gradients while training the model.
hook = Hook(save_config=save_config, save_all=True)
return hook
def test_hook_save_every_step(tmpdir):
save_config = SaveConfig(save_interval=1)
out_dir = os.path.join(tmpdir, str(uuid.uuid4()))
hook = Hook(out_dir=out_dir, save_config=save_config)
run_xgboost_model(hook=hook)
trial = create_trial(out_dir)
assert trial.steps() == list(range(10))
def test_lstm_and_generator(out_dir):
# init hook
hook = KerasHook(
out_dir,
include_collections=[
CollectionKeys.WEIGHTS,
CollectionKeys.LOSSES,
CollectionKeys.GRADIENTS,
],
save_config=SaveConfig(save_steps=[0, 1, 2, 3]),
)
# init model
num_steps = 100
hidden_size = 100
vocabulary = 1000
model = Sequential()
model.add(Embedding(vocabulary, hidden_size, input_length=num_steps))
model.add(LSTM(hidden_size, return_sequences=True))
model.add(LSTM(hidden_size, return_sequences=True))
model.add(Dropout(0.2))
model.add(TimeDistributed(Dense(vocabulary)))
model.add(Activation("softmax"))
model.compile(
loss="categorical_crossentropy",
optimizer=hook.wrap_optimizer(Adam()),
metrics=["categorical_accuracy"],
)
train(3, 32, model, num_steps, hook)
tr = create_trial(out_dir)
assert len(tr.tensor_names(collection=CollectionKeys.LOSSES)) > 0
assert len(tr.tensor_names(collection=CollectionKeys.WEIGHTS)) > 0
def test_loss_collection_with_no_other_collections():
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir = "/tmp/" + run_id
hook = t_hook(out_dir=out_dir,
save_config=save_config,
include_collections=[])
assert has_training_ended(out_dir) == False
run_mnist_gluon_model(hook=hook,
num_steps_train=10,
num_steps_eval=10,
register_to_loss_block=True)
print("Created the trial with out_dir {0}".format(out_dir))
tr = create_trial(out_dir)
assert tr
assert len(tr.steps()) == 4
print(tr.tensor_names())
tname = tr.tensor_names(regex=".*loss")[0]
loss_tensor = tr.tensor(tname)
loss_val = loss_tensor.value(step_num=1)
assert len(loss_val) > 0
shutil.rmtree(out_dir)
def helper_test_modes(hook=None, out_dir="/tmp/test_output/test_hook_modes/"):
prefix = str(uuid.uuid4())
device = torch.device("cpu")
save_steps = [i for i in range(5)]
model = Net(to_save=save_steps).to(device)
json = hook is not None
if hook is None:
out_dir = str(Path(out_dir, prefix))
hook = Hook(
out_dir=out_dir,
save_config=SaveConfig({modes.TRAIN: SaveConfigMode(save_steps=save_steps)}),
include_collections=[
CollectionKeys.WEIGHTS,
CollectionKeys.BIASES,
CollectionKeys.GRADIENTS,
CollectionKeys.DEFAULT,
CollectionKeys.LOSSES,
],
)
hook.register_module(model)
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
hook.set_mode(mode=modes.TRAIN)
train(model, device, optimizer, num_steps=10, save_steps=save_steps)
trial = create_trial(path=out_dir, name="test output")
assert len(trial.modes()) == 1
assert len(trial.steps()) == 5
assert len(trial.steps(mode=modes.TRAIN)) == 5
assert len(trial.steps(mode=modes.EVAL)) == 0
if hook is None:
shutil.rmtree(out_dir)
def test_hook_all_zero(hook=None, out_dir=None):
hook_created = False
if hook is None:
hook_created = True
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir = "/tmp/newlogsRunTest/" + run_id
print("Registering the hook with out_dir {0}".format(out_dir))
shutil.rmtree(out_dir, ignore_errors=True)
hook = t_hook(
out_dir=out_dir,
save_config=save_config,
include_collections=[
"ReluActivation", "weights", "biases", "gradients"
],
)
hook.get_collection("ReluActivation").include(["relu*", "input_*"])
run_mnist_gluon_model(hook=hook,
num_steps_train=10,
num_steps_eval=10,
make_input_zero=True)
print("Created the trial with out_dir {0}".format(out_dir))
tr = create_trial(out_dir)
assert tr
assert len(tr.steps()) == 4
tnames = tr.tensor_names(regex="conv._input")
tname = tr.tensor_names(regex="conv._input")[0]
conv_tensor_value = tr.tensor(tname).value(step_num=0)
is_zero = np.all(conv_tensor_value == 0)
assert is_zero == True
if hook_created:
shutil.rmtree(out_dir)
def test_hook_save_config_collections(tmpdir):
out_dir = os.path.join(tmpdir, str(uuid.uuid4()))
hook = Hook(out_dir=out_dir,
include_collections=["metrics", "feature_importance"])
hook.get_collection("metrics").save_config = SaveConfig(save_interval=2)
hook.get_collection("feature_importance").save_config = SaveConfig(
save_interval=3)
run_xgboost_model(hook=hook)
trial = create_trial(out_dir)
metric_steps = trial.tensor("train-rmse").steps()
assert all(step % 2 == 0 for step in metric_steps[:-1])
fimps = [
t for t in trial.tensor_names() if t.startswith("feature_importance/")
]
fimp_steps = trial.tensor(fimps[0]).steps()
assert all(step % 3 == 0 for step in fimp_steps[:-1])
def create_hook(output_dir, module=None, hook_type="saveall", save_steps=None):
# Create a hook that logs weights, biases, gradients and inputs/ouputs of model
if hook_type == "saveall":
hook = Hook(out_dir=output_dir,
save_config=SaveConfig(save_steps=save_steps),
save_all=True)
elif hook_type == "module-input-output":
# The names of input and output tensors of a module are in following format
# Inputs : <module_name>_input_<input_index>, and
# Output : <module_name>_output
# In order to log the inputs and output of a module, we will create a collection as follows:
assert module is not None
# Create a hook that logs weights, biases, gradients and inputs/outputs of model
hook = Hook(
out_dir=output_dir,
save_config=SaveConfig(save_steps=save_steps),
include_collections=[
CollectionKeys.WEIGHTS,
CollectionKeys.GRADIENTS,
CollectionKeys.BIASES,
"l_mod",
],
)
hook.get_collection("l_mod").add_module_tensors(module,
inputs=True,
outputs=True)
elif hook_type == "weights-bias-gradients":
save_config = SaveConfig(save_steps=save_steps)
# Create a hook that logs ONLY weights, biases, and gradients
hook = Hook(
out_dir=output_dir,
save_config=save_config,
include_collections=[
CollectionKeys.WEIGHTS,
CollectionKeys.BIASES,
CollectionKeys.GRADIENTS,
CollectionKeys.DEFAULT,
CollectionKeys.LOSSES,
],
)
return hook
def test_hook():
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir = "/tmp/newlogsRunTest/" + run_id
hook = t_hook(out_dir=out_dir, save_config=save_config)
assert has_training_ended(out_dir) == False
run_mnist_gluon_model(hook=hook,
num_steps_train=10,
num_steps_eval=10,
register_to_loss_block=True)
shutil.rmtree(out_dir)
def create_smdebug_hook(out_dir, train_data=None, validation_data=None, frequency=1, collections=None,):
save_config = SaveConfig(save_interval=frequency)
hook = Hook(
out_dir=out_dir,
train_data=train_data,
validation_data=validation_data,
save_config=save_config,
include_collections=collections,
)
return hook
def create_hook_from_json_config(hook_cls,
json_config_path,
default_values=None):
"""Returns a SessionHook object corresponding to either TF, PT, or MXNet.
If json_config_path is None, an environment variable must be set.
Here we compare HookParameters with CollectionConfiguration and set all the defaults.
"""
params_dict = get_json_config_as_dict(json_config_path=json_config_path)
hook_params = collect_hook_config_params(params_dict)
out_dir = hook_params.get("out_dir")
dry_run = hook_params.get("dry_run", False)
reduction_config = hook_params.get(CONFIG_RDN_CFG_KEY, None)
save_config = SaveConfig.from_dict(hook_params.get("save_config_modes"),
default_values)
include_regex = hook_params.get(CONFIG_INCLUDE_REGEX_KEY)
include_collections = get_include_collections(params_dict)
save_all = hook_params.get(CONFIG_SAVE_ALL_KEY, False)
include_workers = hook_params.get(CONFIG_INCLUDE_WORKERS_KEY, "one")
# If Sagemaker, emit TB only if JSON file exists
if is_sagemaker_job():
tensorboard_dir = get_tensorboard_dir_from_json_config()
export_tensorboard = bool(tensorboard_dir is not None)
# Otherwise, place TB artifacts in out_dir
else:
tensorboard_dir = hook_params[TENSORBOARD_DIR_KEY]
export_tensorboard = hook_params[EXPORT_TENSORBOARD_KEY]
hook = hook_cls(
out_dir=out_dir,
export_tensorboard=export_tensorboard,
tensorboard_dir=tensorboard_dir,
dry_run=dry_run,
reduction_config=reduction_config,
save_config=save_config,
include_regex=include_regex,
include_collections=include_collections,
include_workers=include_workers,
save_all=save_all,
)
add_collections_to_manager(hook.collection_manager, params_dict,
hook_params)
return hook
def test_hook():
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
run_id = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir = "/tmp/newlogsRunTest/" + run_id
hook = t_hook(out_dir=out_dir, save_config=save_config)
run_mnist_gluon_model(
hook=hook,
num_steps_train=10,
num_steps_eval=10,
register_to_loss_block=True,
save_custom_tensor=True,
)
trial = create_trial(out_dir)
custom_tensors = trial.tensor_names(collection=CollectionKeys.DEFAULT)
all_tensors = trial.tensor_names()
assert len(custom_tensors) == 2
assert len(all_tensors) == 4
shutil.rmtree(out_dir)
def test_hook_save_all(tmpdir):
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
out_dir = os.path.join(tmpdir, str(uuid.uuid4()))
hook = Hook(out_dir=out_dir, save_config=save_config, save_all=True)
run_xgboost_model(hook=hook)
trial = create_trial(out_dir)
collections = trial.collections()
tensors = trial.tensor_names()
assert len(tensors) > 0
assert len(trial.steps()) == 4
assert "all" in collections
assert "metrics" in collections
assert "feature_importance" in collections
assert "train-rmse" in tensors
assert any(t.startswith("feature_importance/") for t in tensors)
assert any(t.startswith("trees/") for t in tensors)
assert len(collections["all"].tensor_names) == len(tensors)
def test_hook(tmpdir):
save_config = SaveConfig(save_steps=[0, 1, 2, 3])
out_dir = os.path.join(tmpdir, str(uuid.uuid4()))
hook = Hook(out_dir=out_dir, save_config=save_config)
assert has_training_ended(out_dir) is False
run_xgboost_model(hook=hook)
def test_spot_hook():
os.environ[
CHECKPOINT_CONFIG_FILE_PATH_ENV_VAR] = "./tests/mxnet/test_json_configs/checkpointconfig.json"
checkpoint_path = "/tmp/savedParams"
if not os.path.exists(checkpoint_path):
os.mkdir(checkpoint_path)
save_config = SaveConfig(
save_steps=[10, 11, 12, 13, 14, 40, 50, 60, 70, 80])
"""
Run the training for 2 epochs and save the parameter after every epoch.
We expect that steps 0 to 14 will be written.
"""
run_id_1 = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir_1 = "/tmp/newlogsRunTest/" + run_id_1
hook = t_hook(out_dir=out_dir_1,
save_config=save_config,
include_collections=["weights", "gradients"])
assert has_training_ended(out_dir_1) == False
run_mnist(
hook=hook,
num_steps_train=10,
num_steps_eval=10,
epochs=2,
save_interval=1,
save_path=checkpoint_path,
)
"""
Run the training again for 4 epochs and save the parameter after every epoch.
We DONOT expect that steps 0 to 14 are written.
We expect to read steps 40, 50, 60, 70 and 80
"""
run_id_2 = "trial_" + datetime.now().strftime("%Y%m%d-%H%M%S%f")
out_dir_2 = "/tmp/newlogsRunTest/" + run_id_2
hook = t_hook(out_dir=out_dir_2,
save_config=save_config,
include_collections=["weights", "gradients"])
assert has_training_ended(out_dir_2) == False
run_mnist(
hook=hook,
num_steps_train=10,
num_steps_eval=10,
epochs=4,
save_interval=1,
save_path=checkpoint_path,
)
# Unset the environ variable before validation so that it won't affect the other scripts in py test environment.
del os.environ[CHECKPOINT_CONFIG_FILE_PATH_ENV_VAR]
# Validation
print("Created the trial with out_dir {0} for the first training".format(
out_dir_1))
tr = create_trial(out_dir_1)
assert tr
available_steps_1 = tr.steps()
assert 40 not in available_steps_1
assert 80 not in available_steps_1
print(available_steps_1)
print("Created the trial with out_dir {0} for the second training".format(
out_dir_2))
tr = create_trial(out_dir_2)
assert tr
available_steps_2 = tr.steps()
assert 40 in available_steps_2
assert 50 in available_steps_2
assert 60 in available_steps_2
assert 70 in available_steps_2
assert 80 in available_steps_2
assert 0 not in available_steps_2
assert 10 not in available_steps_2
assert 11 not in available_steps_2
assert 12 not in available_steps_2
print(available_steps_2)
print("Cleaning up.")
shutil.rmtree(os.path.dirname(out_dir_1))
shutil.rmtree(checkpoint_path, ignore_errors=True)