def model_builder(hp):
# config["parameters"]["combined_graph_layer"]["hidden_dim"] = hp.Choice("hidden_dim", values=[128])
# config["parameters"]["combined_graph_layer"]["distance_dim"] = hp.Choice("distance_dim", values=[128])
config["parameters"]["combined_graph_layer"][
"num_node_messages"] = hp.Choice("num_node_messages", [1, 2])
config["parameters"]["num_graph_layers_common"] = hp.Choice(
"num_graph_layers_common", [1, 2, 3])
config["parameters"]["num_graph_layers_energy"] = hp.Choice(
"num_graph_layers_energy", [1, 2, 3])
config["parameters"]["combined_graph_layer"]["dropout"] = hp.Choice(
"cg_dropout", values=[0.0, 0.1, 0.2])
config["parameters"]["output_decoding"]["dropout"] = hp.Choice(
"output_dropout", values=[0.0, 0.1, 0.2])
config["parameters"]["combined_graph_layer"]["bin_size"] = hp.Choice(
"bin_size", values=[160, 320, 640])
config["parameters"]["combined_graph_layer"][
"ffn_dist_hidden_dim"] = hp.Choice("ffn_dist_hidden_dim",
values=[64, 128, 256])
config["parameters"]["combined_graph_layer"][
"ffn_dist_num_layers"] = hp.Choice("ffn_dist_num_layers",
values=[1, 2])
config["parameters"]["combined_graph_layer"]["kernel"][
"dist_mult"] = hp.Choice("dist_mult", values=[0.01, 0.1, 1.0])
config["parameters"]["combined_graph_layer"]["node_message"][
"output_dim"] = hp.Choice("output_dim", values=[128, 256, 512])
config["parameters"]["combined_graph_layer"]["node_message"][
"normalize_degrees"] = hp.Choice("normalize_degrees",
values=[True, False])
# config["setup"]["lr"] = hp.Choice("lr", values=[1e-4, 3e-4])
# config["setup"]["lr_schedule"] = hp.Choice("lr_schedule", values=["exponentialdecay"])
# config["setup"]["optimizer"] = hp.Choice("optimizer", values=["adam"])
model = make_model(config, dtype="float32")
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
lr_schedule, _ = get_lr_schedule(config, steps=total_steps)
opt = get_optimizer(config, lr_schedule)
loss_dict, loss_weights = get_loss_dict(config)
model.compile(
loss=loss_dict,
optimizer=opt,
sample_weight_mode="temporal",
loss_weights=loss_weights,
metrics={
"cls": [
FlattenedCategoricalAccuracy(name="acc_unweighted",
dtype=tf.float64),
FlattenedCategoricalAccuracy(use_weights=True,
name="acc_weighted",
dtype=tf.float64),
]
},
)
return model
def build_model_and_train(config,
checkpoint_dir=None,
full_config=None,
ntrain=None,
ntest=None,
name=None,
seeds=False):
from ray import tune
from ray.tune.integration.keras import TuneReportCheckpointCallback
from raytune.search_space import set_raytune_search_parameters
if seeds:
# Set seeds for reproducibility
random.seed(1234)
np.random.seed(1234)
tf.random.set_seed(1234)
full_config, config_file_stem = parse_config(full_config)
if config is not None:
full_config = set_raytune_search_parameters(search_space=config,
config=full_config)
strategy, num_gpus = get_strategy()
ds_train, num_train_steps = get_datasets(
full_config["train_test_datasets"], full_config, num_gpus, "train")
ds_test, num_test_steps = get_datasets(full_config["train_test_datasets"],
full_config, num_gpus, "test")
ds_val, ds_info = get_heptfds_dataset(
full_config["validation_datasets"][0],
full_config,
num_gpus,
"test",
full_config["setup"]["num_events_validation"],
supervised=False,
)
ds_val = ds_val.batch(5)
if ntrain:
ds_train = ds_train.take(ntrain)
num_train_steps = ntrain
if ntest:
ds_test = ds_test.take(ntest)
num_test_steps = ntest
print("num_train_steps", num_train_steps)
print("num_test_steps", num_test_steps)
total_steps = num_train_steps * full_config["setup"]["num_epochs"]
print("total_steps", total_steps)
callbacks = prepare_callbacks(
full_config,
tune.get_trial_dir(),
ds_val,
)
callbacks = callbacks[:
-1] # remove the CustomCallback at the end of the list
with strategy.scope():
lr_schedule, optim_callbacks = get_lr_schedule(full_config,
steps=total_steps)
callbacks.append(optim_callbacks)
opt = get_optimizer(full_config, lr_schedule)
model = make_model(full_config, dtype=tf.dtypes.float32)
# Run model once to build the layers
model.build((1, full_config["dataset"]["padded_num_elem_size"],
full_config["dataset"]["num_input_features"]))
full_config = set_config_loss(full_config,
full_config["setup"]["trainable"])
configure_model_weights(model, full_config["setup"]["trainable"])
model.build((1, full_config["dataset"]["padded_num_elem_size"],
full_config["dataset"]["num_input_features"]))
loss_dict, loss_weights = get_loss_dict(full_config)
model.compile(
loss=loss_dict,
optimizer=opt,
sample_weight_mode="temporal",
loss_weights=loss_weights,
metrics={
"cls": [
FlattenedCategoricalAccuracy(name="acc_unweighted",
dtype=tf.float64),
FlattenedCategoricalAccuracy(use_weights=True,
name="acc_weighted",
dtype=tf.float64),
]
},
)
model.summary()
callbacks.append(
TuneReportCheckpointCallback(metrics=[
"adam_beta_1",
"charge_loss",
"cls_acc_unweighted",
"cls_loss",
"cos_phi_loss",
"energy_loss",
"eta_loss",
"learning_rate",
"loss",
"pt_loss",
"sin_phi_loss",
"val_charge_loss",
"val_cls_acc_unweighted",
"val_cls_acc_weighted",
"val_cls_loss",
"val_cos_phi_loss",
"val_energy_loss",
"val_eta_loss",
"val_loss",
"val_pt_loss",
"val_sin_phi_loss",
], ), )
try:
model.fit(
ds_train.repeat(),
validation_data=ds_test.repeat(),
epochs=full_config["setup"]["num_epochs"],
callbacks=callbacks,
steps_per_epoch=num_train_steps,
validation_steps=num_test_steps,
)
except tf.errors.ResourceExhaustedError:
logging.warning(
"Resource exhausted, skipping this hyperparameter configuration."
)
skiplog_file_path = Path(full_config["raytune"]["local_dir"]
) / name / "skipped_configurations.txt"
lines = [
"{}: {}\n".format(item[0], item[1]) for item in config.items()
]
with open(skiplog_file_path, "a") as f:
f.write("#" * 80 + "\n")
for line in lines:
f.write(line)
logging.warning(line[:-1])
f.write("#" * 80 + "\n\n")
def model_scope(config, total_steps, weights, horovod_enabled=False):
lr_schedule, optim_callbacks, lr = get_lr_schedule(config,
steps=total_steps)
opt = get_optimizer(config, lr_schedule)
if config["setup"]["dtype"] == "float16":
model_dtype = tf.dtypes.float16
policy = mixed_precision.Policy("mixed_float16")
mixed_precision.set_global_policy(policy)
opt = mixed_precision.LossScaleOptimizer(opt)
else:
model_dtype = tf.dtypes.float32
model = make_model(config, model_dtype)
# Build the layers after the element and feature dimensions are specified
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
initial_epoch = 0
loaded_opt = None
if weights:
if lr_schedule:
raise Exception(
"Restoring the optimizer state with a learning rate schedule is currently not supported"
)
# We need to load the weights in the same trainable configuration as the model was set up
configure_model_weights(model,
config["setup"].get("weights_config", "all"))
model.load_weights(weights, by_name=True)
opt_weight_file = weights.replace("hdf5",
"pkl").replace("/weights-", "/opt-")
if os.path.isfile(opt_weight_file):
loaded_opt = pickle.load(open(opt_weight_file, "rb"))
initial_epoch = int(weights.split("/")[-1].split("-")[1])
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
config = set_config_loss(config, config["setup"]["trainable"])
configure_model_weights(model, config["setup"]["trainable"])
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
print("model weights")
tw_names = [m.name for m in model.trainable_weights]
for w in model.weights:
print("layer={} trainable={} shape={} num_weights={}".format(
w.name, w.name in tw_names, w.shape, np.prod(w.shape)))
loss_dict, loss_weights = get_loss_dict(config)
model.compile(
loss=loss_dict,
optimizer=opt,
sample_weight_mode="temporal",
loss_weights=loss_weights,
metrics={
"cls": [
FlattenedCategoricalAccuracy(name="acc_unweighted",
dtype=tf.float64),
FlattenedCategoricalAccuracy(
use_weights=True, name="acc_weighted", dtype=tf.float64),
] + [
SingleClassRecall(
icls, name="rec_cls{}".format(icls), dtype=tf.float64)
for icls in range(config["dataset"]["num_output_classes"])
]
},
)
model.summary()
# Set the optimizer weights
if loaded_opt:
def model_weight_setting():
grad_vars = model.trainable_weights
zero_grads = [tf.zeros_like(w) for w in grad_vars]
model.optimizer.apply_gradients(zip(zero_grads, grad_vars))
if model.optimizer.__class__.__module__ == "keras.optimizers.optimizer_v1":
model.optimizer.optimizer.optimizer.set_weights(
loaded_opt["weights"])
else:
model.optimizer.set_weights(loaded_opt["weights"])
# FIXME: check that this still works with multiple GPUs
strategy = tf.distribute.get_strategy()
strategy.run(model_weight_setting)
return model, optim_callbacks, initial_epoch
def train(config, weights, ntrain, ntest, nepochs, recreate, prefix, plot_freq,
customize):
try:
from comet_ml import Experiment
experiment = Experiment(
project_name="particleflow-tf",
auto_metric_logging=True,
auto_param_logging=True,
auto_histogram_weight_logging=True,
auto_histogram_gradient_logging=False,
auto_histogram_activation_logging=False,
)
except Exception as e:
print("Failed to initialize comet-ml dashboard")
experiment = None
"""Train a model defined by config"""
config_file_path = config
config, config_file_stem = parse_config(config,
nepochs=nepochs,
weights=weights)
if plot_freq:
config["callbacks"]["plot_freq"] = plot_freq
if customize:
config = customization_functions[customize](config)
if recreate or (weights is None):
outdir = create_experiment_dir(prefix=prefix + config_file_stem + "_",
suffix=platform.node())
else:
outdir = str(Path(weights).parent)
# Decide tf.distribute.strategy depending on number of available GPUs
strategy, num_gpus = get_strategy()
#if "CPU" not in strategy.extended.worker_devices[0]:
# nvidia_smi_call = "nvidia-smi --query-gpu=timestamp,name,pci.bus_id,pstate,power.draw,temperature.gpu,utilization.gpu,utilization.memory,memory.total,memory.free,memory.used --format=csv -l 1 -f {}/nvidia_smi_log.csv".format(outdir)
# p = subprocess.Popen(shlex.split(nvidia_smi_call))
ds_train, num_train_steps = get_datasets(config["train_test_datasets"],
config, num_gpus, "train")
ds_test, num_test_steps = get_datasets(config["train_test_datasets"],
config, num_gpus, "test")
ds_val, ds_info = get_heptfds_dataset(
config["validation_dataset"], config, num_gpus, "test",
config["setup"]["num_events_validation"])
ds_val = ds_val.batch(5)
if ntrain:
ds_train = ds_train.take(ntrain)
num_train_steps = ntrain
if ntest:
ds_test = ds_test.take(ntest)
num_test_steps = ntest
print("num_train_steps", num_train_steps)
print("num_test_steps", num_test_steps)
total_steps = num_train_steps * config["setup"]["num_epochs"]
print("total_steps", total_steps)
if experiment:
experiment.set_name(outdir)
experiment.log_code("mlpf/tfmodel/model.py")
experiment.log_code("mlpf/tfmodel/utils.py")
experiment.log_code(config_file_path)
shutil.copy(config_file_path, outdir + "/config.yaml"
) # Copy the config file to the train dir for later reference
with strategy.scope():
lr_schedule, optim_callbacks = get_lr_schedule(config,
steps=total_steps)
opt = get_optimizer(config, lr_schedule)
if config["setup"]["dtype"] == "float16":
model_dtype = tf.dtypes.float16
policy = mixed_precision.Policy("mixed_float16")
mixed_precision.set_global_policy(policy)
opt = mixed_precision.LossScaleOptimizer(opt)
else:
model_dtype = tf.dtypes.float32
model = make_model(config, model_dtype)
# Build the layers after the element and feature dimensions are specified
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
initial_epoch = 0
if weights:
# We need to load the weights in the same trainable configuration as the model was set up
configure_model_weights(
model, config["setup"].get("weights_config", "all"))
model.load_weights(weights, by_name=True)
initial_epoch = int(weights.split("/")[-1].split("-")[1])
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
config = set_config_loss(config, config["setup"]["trainable"])
configure_model_weights(model, config["setup"]["trainable"])
model.build((1, config["dataset"]["padded_num_elem_size"],
config["dataset"]["num_input_features"]))
print("model weights")
tw_names = [m.name for m in model.trainable_weights]
for w in model.weights:
print("layer={} trainable={} shape={} num_weights={}".format(
w.name, w.name in tw_names, w.shape, np.prod(w.shape)))
loss_dict, loss_weights = get_loss_dict(config)
model.compile(
loss=loss_dict,
optimizer=opt,
sample_weight_mode="temporal",
loss_weights=loss_weights,
metrics={
"cls": [
FlattenedCategoricalAccuracy(name="acc_unweighted",
dtype=tf.float64),
FlattenedCategoricalAccuracy(use_weights=True,
name="acc_weighted",
dtype=tf.float64),
] + [
SingleClassRecall(
icls, name="rec_cls{}".format(icls), dtype=tf.float64)
for icls in range(config["dataset"]["num_output_classes"])
]
},
)
model.summary()
callbacks = prepare_callbacks(config["callbacks"],
outdir,
ds_val,
ds_info,
comet_experiment=experiment)
callbacks.append(optim_callbacks)
fit_result = model.fit(
ds_train.repeat(),
validation_data=ds_test.repeat(),
epochs=initial_epoch + config["setup"]["num_epochs"],
callbacks=callbacks,
steps_per_epoch=num_train_steps,
validation_steps=num_test_steps,
initial_epoch=initial_epoch,
)
history_path = Path(outdir) / "history"
history_path = str(history_path)
with open("{}/history.json".format(history_path), "w") as fi:
json.dump(fit_result.history, fi)
weights = get_best_checkpoint(outdir)
print("Loading best weights that could be found from {}".format(weights))
model.load_weights(weights, by_name=True)
model.save(outdir + "/model_full", save_format="tf")
print("Training done.")