def test_gradtape_include_collections(out_dir):
"""
This test ensures that a training script written with GradientTape
handles the case where hook config contains all collections mentioned
through include collections
"""
include_collections = [
CollectionKeys.WEIGHTS,
CollectionKeys.BIASES,
CollectionKeys.GRADIENTS,
CollectionKeys.LOSSES,
CollectionKeys.OUTPUTS,
CollectionKeys.METRICS,
CollectionKeys.OPTIMIZER_VARIABLES,
]
save_config = SaveConfig(save_interval=3)
hook = smd.KerasHook(
out_dir,
save_config=save_config,
include_collections=include_collections,
reduction_config=ReductionConfig(norms=ALLOWED_NORMS,
reductions=ALLOWED_REDUCTIONS),
)
helper_keras_gradtape(out_dir, hook=hook)
trial = smd.create_trial(path=out_dir)
# can't save gradients in TF 2.x
assert len(trial.tensor_names()) == (16 if is_tf_2_2() else 15)
assert len(trial.tensor_names(collection=CollectionKeys.GRADIENTS)) == 4
assert len(
trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 1
def test_subclassed_model(out_dir):
# Download and load MNIST dataset.
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data("MNIST-data")
x_train, x_test = x_train / 255.0, x_test / 255.0
# Add a channels dimension
x_train = x_train[..., tf.newaxis]
x_test = x_test[..., tf.newaxis]
# Create an instance of the model
model = MyModel()
train_ds = (
tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(10000, seed=123).batch(2)
)
MyModel.hook = smd.KerasHook(
out_dir,
save_all=True,
save_config=smd.SaveConfig(save_steps=[x for x in range(10)], save_interval=1),
)
MyModel.hook.register_model(model)
model.compile(optimizer="Adam", loss="mse", run_eagerly=True)
model.fit(train_ds, epochs=1, steps_per_epoch=10, callbacks=[MyModel.hook])
trial = smd.create_trial(out_dir)
assert len(trial.tensor_names(collection=smd.CollectionKeys.LAYERS)) == 8
assert trial.tensor_names(collection=smd.CollectionKeys.INPUTS) == ["model_input"]
assert trial.tensor_names(collection=smd.CollectionKeys.OUTPUTS) == ["labels", "predictions"]
assert trial.tensor_names(collection=smd.CollectionKeys.LOSSES) == ["loss"]
assert len(trial.tensor_names(collection=smd.CollectionKeys.GRADIENTS)) == 6
def test_keras_fit(out_dir, tf_eager_mode, saveall):
hook = smd.KerasHook(out_dir=out_dir, save_all=saveall)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
# can't save gradients in TF 2.x eager mode
if saveall: # save losses, metrics, weights, biases
if tf_eager_mode:
assert len(trial.tensor_names()) == (12 if is_tf_2_2() else 13)
else:
assert len(trial.tensor_names()) == 21
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(
trial.tensor_names(
collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
assert (
len(
trial.tensor_names(
collection=CollectionKeys.OPTIMIZER_VARIABLES,
mode=ModeKeys.EVAL)) == 0,
"No Optimizer Variables Should be Saved in EVAL Mode",
)
else: # save the default losses and metrics
assert len(trial.tensor_names()) == (3 if is_tf_2_2() and tf_eager_mode
else 4)
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == (
2 if is_tf_2_2() and tf_eager_mode else 3)
def main():
parser = argparse.ArgumentParser(description="Train resnet50 cifar10")
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--epoch", type=int, default=3)
parser.add_argument("--model_dir",
type=str,
default="./model_keras_resnet")
parser.add_argument("--out_dir", type=str)
parser.add_argument("--save_interval", type=int, default=500)
opt = parser.parse_args()
model = ResNet50(weights=None, input_shape=(32, 32, 3), classes=10)
##### Enabling SageMaker Debugger ###########
# creating hook
hook = smd.KerasHook(
out_dir=opt.out_dir,
include_collections=["weights", "gradients", "losses"],
save_config=smd.SaveConfig(save_interval=opt.save_interval),
)
optimizer = tf.keras.optimizers.Adam()
##### Enabling SageMaker Debugger ###########
# wrap the optimizer so the hook can identify the gradients
optimizer = hook.wrap_optimizer(optimizer)
model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=["accuracy"])
# start the training.
train(opt.batch_size, opt.epoch, model, hook)
def test_regex_filtering_for_default_collections(out_dir):
hook = smd.KerasHook(
out_dir,
save_config=SaveConfig(save_interval=9),
include_collections=[CollectionKeys.LAYERS, CollectionKeys.GRADIENTS],
)
hook.get_collection(CollectionKeys.LAYERS).include("^dense")
hook.get_collection(CollectionKeys.GRADIENTS).include("gradients/dense")
helper_keras_fit(
out_dir,
hook=hook,
save_config=SaveConfig(save_interval=10),
steps=["train"],
run_eagerly=True,
)
tr = create_trial_fast_refresh(out_dir)
layer_tnames = tr.tensor_names(collection=CollectionKeys.LAYERS)
gradient_tnames = tr.tensor_names(collection=CollectionKeys.GRADIENTS)
assert len(layer_tnames) == (4 if is_tf_2_2() else 0)
assert len(gradient_tnames) == (4 if is_tf_2_2() else 0)
layer_pattern = r"^(dense)(_\d+)?\/(inputs|outputs)"
gradient_pattern = r"gradients/dense"
for tname in layer_tnames:
assert tr.tensor(tname).value(0) is not None
assert re.match(pattern=layer_pattern, string=tname) is not None
for tname in gradient_tnames:
assert tr.tensor(tname).value(0) is not None
assert re.match(pattern=gradient_pattern, string=tname) is not None
def test_model_inputs_and_outputs(out_dir, tf_eager_mode):
# explicitly save INPUTS and OUTPUTS
include_collections = [CollectionKeys.INPUTS, CollectionKeys.OUTPUTS]
hook = smd.KerasHook(out_dir=out_dir,
include_collections=include_collections)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
assert len(trial.steps(mode=ModeKeys.TRAIN)) == 3
assert len(trial.tensor_names(collection=CollectionKeys.OUTPUTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.INPUTS)) == 1
for tname in trial.tensor_names(collection=CollectionKeys.OUTPUTS):
output = trial.tensor(tname)
assert tname in ["y", "y_pred"]
assert output.value(0) is not None
# Check the shape of output tensors
assert trial.tensor("y").value(0).shape[1] == 1 # label
assert trial.tensor("y_pred").value(
0).shape[1] == 10 # Output probability for each class
def test_keras_gradients(script_mode, tf_optimizer):
""" Works as intended. """
smd.del_hook()
tf.reset_default_graph()
tf.keras.backend.clear_session()
json_file_contents = """
{
"S3OutputPath": "s3://sagemaker-test",
"LocalPath": "/opt/ml/output/tensors",
"CollectionConfigurations": [
{
"CollectionName": "gradients"
},
{
"CollectionName": "optimizer_variables"
},
{
"CollectionName": "losses"
}
]
}
"""
with SagemakerSimulator(json_file_contents=json_file_contents) as sim:
model = get_keras_model_v1()
(x_train, y_train), (x_test, y_test) = get_keras_data()
if tf_optimizer:
opt = tf.train.RMSPropOptimizer(0.1)
else:
opt = tf.keras.optimizers.RMSprop()
if script_mode:
hook = smd.KerasHook(
out_dir=sim.out_dir,
include_collections=["gradients", "optimizer_variables", "losses"],
)
opt = hook.wrap_optimizer(opt)
model.compile(
loss="sparse_categorical_crossentropy", optimizer=opt, metrics=["accuracy"]
)
history = model.fit(
x_train, y_train, batch_size=16, epochs=5, validation_split=0.2, callbacks=[hook]
)
test_scores = model.evaluate(x_test, y_test, verbose=2, callbacks=[hook])
else:
model.compile(
loss="sparse_categorical_crossentropy", optimizer=opt, metrics=["accuracy"]
)
history = model.fit(x_train, y_train, batch_size=16, epochs=5, validation_split=0.2)
test_scores = model.evaluate(x_test, y_test, verbose=2)
# Check that hook created and tensors saved
trial = smd.create_trial(path=sim.out_dir)
assert smd.get_hook() is not None, "Hook was not created."
assert len(trial.steps()) > 0, "Nothing saved at any step."
assert len(trial.tensor_names()) > 0, "Tensors were not saved."
assert len(trial.tensor_names(collection="gradients")) > 0
if not tf_optimizer:
# as this is only supported for keras optimizers currently
assert len(trial.tensor_names(collection="optimizer_variables")) > 0
def test_keras_fit(out_dir, tf_eager_mode, saveall):
hook = smd.KerasHook(out_dir=out_dir, save_all=saveall)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
# can't save gradients in TF 2.x eager mode
if saveall: # save losses, metrics, weights, biases
if tf_eager_mode:
assert len(trial.tensor_names()) == 7 if is_tf_2_2() else 8
else:
assert len(trial.tensor_names()) == 21
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
else: # save the default losses and metrics
assert len(trial.tensor_names()) == 3 if is_tf_2_2() and tf_eager_mode else 4
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert (
len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 2
if is_tf_2_2() and tf_eager_mode
else 3
)
def test_keras_v1(script_mode):
""" Works as intended. """
smd.del_hook()
tf.reset_default_graph()
tf.keras.backend.clear_session()
with SagemakerSimulator() as sim:
model = get_keras_model_v1()
(x_train, y_train), (x_test, y_test) = get_keras_data()
model.compile(
loss="sparse_categorical_crossentropy",
optimizer=tf.keras.optimizers.RMSprop(),
metrics=["accuracy"],
)
if script_mode:
hook = smd.KerasHook(out_dir=sim.out_dir)
history = model.fit(
x_train, y_train, batch_size=64, epochs=5, validation_split=0.2, callbacks=[hook]
)
test_scores = model.evaluate(x_test, y_test, verbose=2, callbacks=[hook])
else:
history = model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.2)
test_scores = model.evaluate(x_test, y_test, verbose=2)
# Check that hook created and tensors saved
trial = smd.create_trial(path=sim.out_dir)
assert smd.get_hook() is not None, "Hook was not created."
assert len(trial.steps()) > 0, "Nothing saved at any step."
assert len(trial.tensor_names()) > 0, "Tensors were not saved."
def test_include_collections(out_dir, tf_eager_mode):
include_collections = [
CollectionKeys.WEIGHTS,
CollectionKeys.BIASES,
CollectionKeys.GRADIENTS,
CollectionKeys.LOSSES,
CollectionKeys.OUTPUTS,
CollectionKeys.METRICS,
CollectionKeys.OPTIMIZER_VARIABLES,
]
save_config = SaveConfig(save_interval=3)
hook = smd.KerasHook(
out_dir,
save_config=save_config,
include_collections=include_collections,
reduction_config=ReductionConfig(norms=ALLOWED_NORMS, reductions=ALLOWED_REDUCTIONS),
)
helper_keras_fit(out_dir, hook=hook, steps=["train", "eval", "predict"], eager=tf_eager_mode)
trial = smd.create_trial(path=out_dir)
# can't save gradients in TF 2.x
if tf_eager_mode:
assert len(trial.tensor_names()) == 7 if is_tf_2_2() else 8
else:
assert len(trial.tensor_names()) == 18
assert len(trial.tensor_names(collection=CollectionKeys.GRADIENTS)) == 4
assert len(trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert (
len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 2
if is_tf_2_2() and tf_eager_mode
else 3
)
def test_save_layer_inputs_and_outputs(out_dir, tf_eager_mode):
# explicitly save INPUTS and OUTPUTS
include_collections = [CollectionKeys.INPUTS, CollectionKeys.OUTPUTS]
hook = smd.KerasHook(out_dir=out_dir,
include_collections=include_collections)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
assert len(trial.tensor_names(collection=CollectionKeys.INPUTS)) == 4
assert len(trial.tensor_names(collection=CollectionKeys.OUTPUTS)) == 4
# Check that output of layer is equal to the input of the next
boolean_matrix = trial.tensor("flatten/outputs").value(0) == trial.tensor(
"dense/inputs").value(0)
assert boolean_matrix.all()
boolean_matrix = trial.tensor("dense/outputs").value(0) == trial.tensor(
"dropout/inputs").value(0)
assert boolean_matrix.all()
boolean_matrix = trial.tensor("dropout/outputs").value(0) == trial.tensor(
"dense_1/inputs").value(0)
assert boolean_matrix.all()
def test_keras_fit_pure_eager(out_dir, tf_eager_mode):
"""
Test save all and save default collection in fit() pure eager mode
"""
hook = smd.KerasHook(out_dir=out_dir,
save_all=True,
save_config=SaveConfig(save_interval=3))
helper_keras_fit(trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
run_eagerly=True)
trial = smd.create_trial(path=out_dir)
if is_tf_2_2():
assert len(trial.tensor_names()) == 27
else:
assert len(trial.tensor_names()) == (20 if is_tf_2_3() else 21)
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(
trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
assert len(trial.tensor_names(
collection=CollectionKeys.INPUTS)) == (1 if is_tf_2_2() else 0)
assert len(trial.tensor_names(
collection=CollectionKeys.OUTPUTS)) == (2 if is_tf_2_2() else 0)
def helper_test_keras_v2(script_mode: bool = False, eager_mode: bool = True):
""" Test the default ZCC behavior of saving losses and metrics in eager and non-eager modes."""
smd.del_hook()
tf.keras.backend.clear_session()
if not eager_mode:
tf.compat.v1.disable_eager_execution()
with SagemakerSimulator() as sim:
model = get_keras_model_v2()
(x_train, y_train), (x_test, y_test) = get_keras_data()
x_train, x_test = x_train / 255, x_test / 255
opt = tf.keras.optimizers.RMSprop()
if script_mode:
hook = smd.KerasHook(out_dir=sim.out_dir, export_tensorboard=True)
opt = hook.wrap_optimizer(opt)
model.compile(loss="sparse_categorical_crossentropy",
optimizer=opt,
metrics=["accuracy"])
history = model.fit(x_train,
y_train,
batch_size=64,
epochs=2,
validation_split=0.2,
callbacks=[hook])
test_scores = model.evaluate(x_test,
y_test,
verbose=2,
callbacks=[hook])
else:
model.compile(loss="sparse_categorical_crossentropy",
optimizer=opt,
metrics=["accuracy"])
history = model.fit(x_train,
y_train,
batch_size=64,
epochs=2,
validation_split=0.2)
test_scores = model.evaluate(x_test, y_test, verbose=2)
hook = smd.get_hook()
assert hook
hook.close()
# Check that hook created and tensors saved
trial = smd.create_trial(path=sim.out_dir)
assert len(trial.steps()) > 0, "Nothing saved at any step."
assert len(trial.tensor_names()) > 0, "Tensors were not saved."
# DEFAULT TENSORS SAVED
assert len(trial.tensor_names(
collection=CollectionKeys.LOSSES)) > 0, "No Losses Saved"
assert len(trial.tensor_names(
collection=CollectionKeys.METRICS)) > 0, "No Metrics Saved"
assert (len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 0
), "Weights were not expected to be saved by default"
assert (len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 0
), "Biases were not expected to be saved by default"
def test_keras_fit_shapes(out_dir):
hook = smd.KerasHook(
out_dir=out_dir,
save_all=True,
save_config=SaveConfig(save_steps=[0]),
reduction_config=ReductionConfig(save_shape=True),
)
helper_keras_fit(trial_dir=out_dir, hook=hook)
print(create_trial_fast_refresh(out_dir).tensor_names(step=0))
verify_shapes(out_dir, 0)
def test_keras_gradtape_shapes(out_dir):
hook = smd.KerasHook(
out_dir=out_dir,
save_all=True,
save_config=SaveConfig(save_steps=[0]),
reduction_config=ReductionConfig(save_shape=True),
)
helper_keras_gradtape(trial_dir=out_dir, hook=hook)
verify_shapes(out_dir, 0)
verify_shapes(out_dir, 500)
def test_should_save_tensor_behavior_without_prepare_collections(out_dir):
"""Always return false if an attempt to save a tensor is made before the collections are prepared.
This can happen if the fn is called before callbacks are init."""
hook = smd.KerasHook(out_dir,
save_config=SaveConfig(save_interval=3),
save_all=True)
assert not hook.should_save_tensor_or_collection("dummy",
CollectionKeys.GRADIENTS)
assert not hook.should_save_tensor_or_collection("dummy",
CollectionKeys.LAYERS)
def helper_create_hook(out_dir, collections, include_regex=None):
hook = smd.KerasHook(out_dir,
save_config=SaveConfig(save_interval=3),
include_collections=collections)
if include_regex:
for collection in collections:
hook.get_collection(collection).include(include_regex)
hook.register_model(model)
hook.on_train_begin()
return hook
def test_layer_names_gradient_tape(out_dir):
hook = smd.KerasHook(
out_dir,
save_config=SaveConfig(save_interval=9),
include_collections=[CollectionKeys.LAYERS],
)
helper_keras_gradtape(out_dir, hook=hook, save_config=SaveConfig(save_interval=9))
tr = create_trial_fast_refresh(out_dir)
tnames = tr.tensor_names(collection=CollectionKeys.LAYERS)
pattern = r"^(flatten|dense|dropout)(_\d+)?\/(inputs|outputs)"
for tname in tnames:
assert re.match(pattern=pattern, string=tname) is not None
def test_mixed_precision_training(out_dir):
from tensorflow.keras.mixed_precision import experimental as mixed_precision
hook = smd.KerasHook(out_dir=out_dir, save_all=True)
policy = mixed_precision.Policy("mixed_float16")
mixed_precision.set_policy(policy)
inputs = keras.Input(shape=(784,), name="digits")
if tf.config.list_physical_devices("GPU"):
# The model will run with 4096 units on a GPU
num_units = 4096
else:
# Use fewer units on CPUs so the model finishes in a reasonable amount of time
# The model will run with 64 units on a CPU
num_units = 64
dense1 = layers.Dense(num_units, activation="relu", name="dense_1")
x = dense1(inputs)
dense2 = layers.Dense(num_units, activation="relu", name="dense_2")
x = dense2(x)
# CORRECT: softmax and model output are float32
x = layers.Dense(10, name="dense_logits")(x)
outputs = layers.Activation("softmax", dtype="float32", name="predictions")(x)
# The linear activation is an identity function. So this simply casts 'outputs'
# to float32. In this particular case, 'outputs' is already float32 so this is a
# no-op.
outputs = layers.Activation("linear", dtype="float32")(outputs)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(
loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.RMSprop(),
metrics=["accuracy"],
)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape(60000, 784).astype("float32") / 255
x_test = x_test.reshape(10000, 784).astype("float32") / 255
initial_weights = model.get_weights()
hooks = [hook]
history = model.fit(
x_train, y_train, batch_size=8192, epochs=5, callbacks=hooks, validation_split=0.2
)
test_scores = model.evaluate(x_test, y_test, verbose=2)
trial = create_trial(out_dir)
assert len(trial.tensor_names()) == 30
def test_keras_fit_false(out_dir, tf_eager_mode=False):
test_include_collections = [
CollectionKeys.LOSSES,
CollectionKeys.METRICS,
CollectionKeys.WEIGHTS,
CollectionKeys.BIASES,
CollectionKeys.GRADIENTS,
CollectionKeys.INPUTS,
CollectionKeys.OUTPUTS,
CollectionKeys.LAYERS,
CollectionKeys.OPTIMIZER_VARIABLES,
]
hook = smd.KerasHook(out_dir=out_dir,
include_collections=test_include_collections)
helper_keras_fit(
include_collections=test_include_collections,
trial_dir=out_dir,
hook=hook,
run_eagerly=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
# We first assert that none of the collections we requested for are empty
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.GRADIENTS)) == 4
assert len(trial.tensor_names(
collection=CollectionKeys.INPUTS)) == 1 # 1 Model Input
assert len(trial.tensor_names(
collection=CollectionKeys.OUTPUTS)) == 2 # 2 Model outputs
assert len(
trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
# We assert that all the tensors saved have a valid value
for tname in trial.tensor_names():
assert trial.tensor(tname).value(0) is not None
# We then analyse Layer Inputs and Layer Outputs
# Check that output of layer is equal to the input of the next
boolean_matrix = trial.tensor("flatten_1/outputs").value(
0) == trial.tensor("dense_2/inputs").value(0)
assert boolean_matrix.all()
boolean_matrix = trial.tensor("dense_2/outputs").value(0) == trial.tensor(
"dropout_1/inputs").value(0)
assert boolean_matrix.all()
boolean_matrix = trial.tensor("dropout_1/outputs").value(
0) == trial.tensor("dense_3/inputs").value(0)
assert boolean_matrix.all()
def test_keras_fit(out_dir, tf_eager_mode, saveall):
hook = smd.KerasHook(out_dir=out_dir, save_all=saveall)
ts = time.time()
hook.save_scalar("foobar", 1, sm_metric=True, timestamp=ts)
scalars_to_be_saved = dict()
scalars_to_be_saved["scalar/foobar"] = (ts, 0)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
run_eagerly=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
# can't save gradients in TF 2.x eager mode
if saveall: # save losses, metrics, weights, biases, scalar
if tf_eager_mode:
if is_tf_2_2():
assert len(trial.tensor_names()) == 28
else:
assert len(trial.tensor_names()) == (21 if is_tf_2_3() else 14)
assert len(trial.tensor_names(collection=CollectionKeys.INPUTS)) == (
1 if is_tf_2_2() else 0
)
assert len(trial.tensor_names(collection=CollectionKeys.OUTPUTS)) == (
2 if is_tf_2_2() else 0
)
else:
assert len(trial.tensor_names()) == 21
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
assert (
len(
trial.tensor_names(
collection=CollectionKeys.OPTIMIZER_VARIABLES, mode=ModeKeys.EVAL
)
)
== 0,
"No Optimizer Variables Should be Saved in EVAL Mode",
)
else: # save the default losses and metrics
assert len(trial.tensor_names()) == (
4 if (is_tf_2_2() or is_tf_2_3()) and tf_eager_mode else 5
)
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == (
2 if (is_tf_2_2() or is_tf_2_3()) and tf_eager_mode else 3
)
for tname in trial.tensor_names():
assert trial.tensor(tname).value(0) is not None
def helper_create_hook(out_dir, collections, include_regex=None):
hook = smd.KerasHook(out_dir,
save_config=SaveConfig(save_interval=3),
include_collections=collections)
if include_regex:
for collection in collections:
hook.get_collection(collection).include(include_regex)
hook.register_model(model)
hook.set_mode(ModeKeys.TRAIN)
hook._prepare_collections_for_tf2()
hook._increment_step()
hook.on_train_begin()
return hook
def test_include_only_custom_collection(out_dir, tf_eager_mode):
include_collections = ["custom_optimizer_variables"]
save_config = SaveConfig(save_interval=3)
hook = smd.KerasHook(
out_dir,
save_config=save_config,
include_collections=include_collections,
reduction_config=ReductionConfig(norms=ALLOWED_NORMS, reductions=ALLOWED_REDUCTIONS),
)
hook.get_collection("custom_optimizer_variables").include("Adam")
helper_keras_fit(out_dir, hook=hook, steps=["train", "eval", "predict"], eager=tf_eager_mode)
trial = smd.create_trial(path=out_dir)
assert len(trial.tensor_names()) == (8 if is_tf_2_2() and tf_eager_mode else 9)
assert len(trial.tensor_names(collection="custom_optimizer_variables")) == 5
def test_hook_timeline_file_write(
set_up_smprofiler_config_path, set_up_resource_config, out_dir, tf_eager_mode
):
hook = smd.KerasHook(out_dir=out_dir, save_all=False)
helper_keras_fit(trial_dir=out_dir, hook=hook, eager=tf_eager_mode, steps=["train", "eval"])
files = []
for path in Path(out_dir + "/" + DEFAULT_PREFIX).rglob("*.json"):
files.append(path)
assert len(files) == 1
with open(files[0]) as timeline_file:
events_dict = json.load(timeline_file)
assert events_dict
def test_gradtape_include_regex(out_dir):
"""
Test custom collection with regex
"""
hook = smd.KerasHook(
out_dir, save_config=SaveConfig(save_interval=9), include_collections=["custom_coll"]
)
hook.get_collection("custom_coll").include("dense")
helper_keras_gradtape(out_dir, hook=hook, save_config=SaveConfig(save_interval=9))
tr = create_trial_fast_refresh(out_dir)
tnames = tr.tensor_names(collection="custom_coll")
assert len(tnames) == (12 if is_tf_2_2() else 8)
for tname in tnames:
assert tr.tensor(tname).value(0) is not None
def test_gradtape_persistent(out_dir, saveall):
"""
Test save all and save default collection
"""
hook = smd.KerasHook(out_dir=out_dir, save_all=saveall, save_config=SaveConfig(save_interval=3))
helper_keras_gradtape(trial_dir=out_dir, hook=hook, persistent=True)
trial = smd.create_trial(path=out_dir)
if saveall: # save losses, metrics, weights, biases
assert len(trial.tensor_names()) == 10
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
else: # save the default losses and metrics
assert len(trial.tensor_names()) == 2
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 1
def test_multiple_inputs(out_dir):
my_model = MyModel()
hook = smd.KerasHook(
out_dir, save_all=True, save_config=smd.SaveConfig(save_steps=[0], save_interval=1)
)
hook.register_model(my_model)
x_train = np.random.random((1000, 20))
y_train = np.random.random((1000, 1))
my_model.compile(optimizer="Adam", loss="mse", run_eagerly=True)
my_model.fit(x_train, y_train, epochs=1, steps_per_epoch=1, callbacks=[hook])
trial = create_trial(path=out_dir)
tnames = sorted(trial.tensor_names(collection=smd.CollectionKeys.LAYERS))
assert "concatenate" in tnames[0]
assert len(trial.tensor(tnames[0]).value(0)) == 2
assert trial.tensor(tnames[0]).shape(0) == (2, 1000, 20)
def test_save_gradients(out_dir, tf_eager_mode):
# explicitly save INPUTS and OUTPUTS
include_collections = [CollectionKeys.GRADIENTS]
hook = smd.KerasHook(out_dir=out_dir, include_collections=include_collections)
helper_keras_fit(
trial_dir=out_dir,
hook=hook,
eager=tf_eager_mode,
steps=["train", "eval", "predict", "train"],
)
trial = smd.create_trial(path=out_dir)
assert len(trial.tensor_names(collection=CollectionKeys.GRADIENTS)) == 4
for tname in trial.tensor_names(collection=CollectionKeys.GRADIENTS):
output = trial.tensor(tname)
assert output.value(0) is not None
def test_keras_gradtape(out_dir, saveall):
"""
Test save all and save default collection
"""
hook = smd.KerasHook(out_dir=out_dir, save_all=saveall, save_config=SaveConfig(save_interval=3))
helper_keras_gradtape(trial_dir=out_dir, hook=hook)
trial = smd.create_trial(path=out_dir)
if saveall: # save losses, metrics, weights, biases
assert len(trial.tensor_names()) == (25 if is_tf_2_2() else 15)
assert len(trial.tensor_names(collection=CollectionKeys.BIASES)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.WEIGHTS)) == 2
assert len(trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES)) == 5
else: # save the default losses and metrics
assert len(trial.tensor_names()) == 2
assert len(trial.tensor_names(collection=CollectionKeys.LOSSES)) == 1
assert len(trial.tensor_names(collection=CollectionKeys.METRICS)) == 1
def test_include_regex(out_dir, tf_eager_mode):
hook = smd.KerasHook(out_dir,
save_config=SaveConfig(save_interval=9),
include_collections=["custom_coll"])
hook.get_collection("custom_coll").include("dense")
helper_keras_fit(
out_dir,
hook=hook,
save_config=SaveConfig(save_interval=9),
steps=["train"],
run_eagerly=tf_eager_mode,
)
tr = create_trial_fast_refresh(out_dir)
tnames = tr.tensor_names(collection="custom_coll")
assert len(tnames) == 12
for tname in tnames:
assert tr.tensor(tname).value(0) is not None
