def run_simple(args: RunConfiguration, with_continue: bool = False):
hf_config = AutoConfig.from_pretrained(args.hf_pretrained_model_name_or_path)
model_cache_path = replace_none(
args.model_cache_path, default=os.path.join(args.exp_dir, "models")
)
with distributed.only_first_process(local_rank=args.local_rank):
# === Step 1: Write task configs based on templates === #
full_task_name_list = sorted(list(set(args.train_tasks + args.val_tasks + args.test_tasks)))
task_config_path_dict = {}
if args.create_config:
task_config_path_dict = create_and_write_task_configs(
task_name_list=full_task_name_list,
data_dir=args.data_dir,
task_config_base_path=os.path.join(args.data_dir, "configs"),
)
else:
for task_name in full_task_name_list:
task_config_path_dict[task_name] = os.path.join(
args.data_dir, "configs", f"{task_name}_config.json"
)
# === Step 2: Download models === #
# if not os.path.exists(os.path.join(model_cache_path, hf_config.model_type)):
# print("Downloading model")
# export_model.export_model(
# hf_pretrained_model_name_or_path=args.hf_pretrained_model_name_or_path,
# output_base_path=os.path.join(model_cache_path, hf_config.model_type),
# )
# === Step 3: Tokenize and cache === #
phase_task_dict = {
"train": args.train_tasks,
"val": args.val_tasks,
"test": args.test_tasks,
}
for task_name in full_task_name_list:
phases_to_do = []
for phase, phase_task_list in phase_task_dict.items():
if task_name in phase_task_list and not os.path.exists(
os.path.join(args.exp_dir, "cache", hf_config.model_type, task_name, phase)
):
config = read_json(task_config_path_dict[task_name])
if phase in config["paths"]:
phases_to_do.append(phase)
else:
phase_task_list.remove(task_name)
if not phases_to_do:
continue
print(f"Tokenizing Task '{task_name}' for phases '{','.join(phases_to_do)}'")
tokenize_and_cache.main(
tokenize_and_cache.RunConfiguration(
task_config_path=task_config_path_dict[task_name],
hf_pretrained_model_name_or_path=args.hf_pretrained_model_name_or_path,
output_dir=os.path.join(args.exp_dir, "cache", hf_config.model_type, task_name),
phases=phases_to_do,
# TODO: Need a strategy for task-specific max_seq_length issues (issue #1176)
max_seq_length=args.max_seq_length,
smart_truncate=True,
do_iter=True,
)
)
# === Step 4: Generate jiant_task_container_config === #
# We'll do this with a configurator. Creating a jiant_task_config has a surprising
# number of moving parts.
jiant_task_container_config = configurator.SimpleAPIMultiTaskConfigurator(
task_config_base_path=os.path.join(args.data_dir, "configs"),
task_cache_base_path=os.path.join(args.exp_dir, "cache", hf_config.model_type),
train_task_name_list=args.train_tasks,
val_task_name_list=args.val_tasks,
test_task_name_list=args.test_tasks,
train_batch_size=args.train_batch_size,
eval_batch_multiplier=2,
epochs=args.num_train_epochs,
num_gpus=torch.cuda.device_count(),
train_examples_cap=args.train_examples_cap,
).create_config()
os.makedirs(os.path.join(args.exp_dir, "run_configs"), exist_ok=True)
jiant_task_container_config_path = os.path.join(
args.exp_dir, "run_configs", f"{args.run_name}_config.json"
)
py_io.write_json(jiant_task_container_config, path=jiant_task_container_config_path)
# === Step 5: Train/Eval! === #
if args.model_weights_path:
model_load_mode = "partial"
model_weights_path = args.model_weights_path
else:
# From Transformers
if any(task_name.startswith("mlm_") for task_name in full_task_name_list):
model_load_mode = "from_transformers_with_mlm"
else:
model_load_mode = "from_transformers"
model_weights_path = os.path.join(
model_cache_path, hf_config.model_type, "model", "model.p"
)
run_output_dir = os.path.join(args.exp_dir, "runs", args.run_name)
if (
args.save_checkpoint_every_steps
and os.path.exists(os.path.join(run_output_dir, "checkpoint.p"))
and with_continue
):
print("Resuming")
checkpoint = torch.load(os.path.join(run_output_dir, "checkpoint.p"))
run_args = runscript.RunConfiguration.from_dict(checkpoint["metadata"]["args"])
else:
print("Running from start")
run_args = runscript.RunConfiguration(
# === Required parameters === #
jiant_task_container_config_path=jiant_task_container_config_path,
output_dir=run_output_dir,
# === Model parameters === #
hf_pretrained_model_name_or_path=args.hf_pretrained_model_name_or_path,
model_path=model_weights_path,
model_config_path=os.path.join(
model_cache_path, hf_config.model_type, "model", "config.json",
),
model_load_mode=model_load_mode,
# === Running Setup === #
do_train=bool(args.train_tasks),
do_val=bool(args.val_tasks),
do_save=args.do_save,
do_save_best=args.do_save_best,
do_save_last=args.do_save_last,
write_val_preds=args.write_val_preds,
write_test_preds=args.write_test_preds,
eval_every_steps=args.eval_every_steps,
save_every_steps=args.save_every_steps,
save_checkpoint_every_steps=args.save_checkpoint_every_steps,
no_improvements_for_n_evals=args.no_improvements_for_n_evals,
keep_checkpoint_when_done=args.keep_checkpoint_when_done,
force_overwrite=args.force_overwrite,
seed=args.seed,
# === Training Learning Parameters === #
learning_rate=args.learning_rate,
adam_epsilon=args.adam_epsilon,
max_grad_norm=args.max_grad_norm,
optimizer_type=args.optimizer_type,
# === Specialized config === #
no_cuda=args.no_cuda,
fp16=args.fp16,
fp16_opt_level=args.fp16_opt_level,
local_rank=args.local_rank,
server_ip=args.server_ip,
server_port=args.server_port,
)
checkpoint = None
runscript.run_loop(args=run_args, checkpoint=checkpoint)
py_io.write_file(args.to_json(), os.path.join(run_output_dir, "simple_run_config.json"))
def setup_runner(
args: RunConfiguration,
jiant_task_container: container_setup.JiantTaskContainer,
quick_init_out,
verbose: bool = True,
) -> jiant_runner.JiantRunner:
"""Setup jiant model, optimizer, and runner, and return runner.
Args:
args (RunConfiguration): configuration carrying command line args specifying run params.
jiant_task_container (container_setup.JiantTaskContainer): task and sampler configs.
quick_init_out (QuickInitContainer): device (GPU/CPU) and logging configuration.
verbose: If True, enables printing configuration info (to standard out).
Returns:
jiant_runner.JiantRunner
"""
# TODO document why the distributed.only_first_process() context manager is being used here.
with distributed.only_first_process(local_rank=args.local_rank):
# load the model
jiant_model = jiant_model_setup.setup_jiant_model(
hf_pretrained_model_name_or_path=args.
hf_pretrained_model_name_or_path,
model_config_path=args.model_config_path,
task_dict=jiant_task_container.task_dict,
taskmodels_config=jiant_task_container.taskmodels_config,
)
jiant_model_setup.delegate_load_from_path(
jiant_model=jiant_model,
weights_path=args.model_path,
load_mode=args.model_load_mode)
jiant_model.to(quick_init_out.device)
optimizer_scheduler = model_setup.create_optimizer(
model=jiant_model,
learning_rate=args.learning_rate,
t_total=jiant_task_container.global_train_config.max_steps,
warmup_steps=jiant_task_container.global_train_config.warmup_steps,
warmup_proportion=None,
optimizer_type=args.optimizer_type,
verbose=verbose,
)
jiant_model, optimizer = model_setup.raw_special_model_setup(
model=jiant_model,
optimizer=optimizer_scheduler.optimizer,
fp16=args.fp16,
fp16_opt_level=args.fp16_opt_level,
n_gpu=quick_init_out.n_gpu,
local_rank=args.local_rank,
)
print("-----------------", jiant_model)
optimizer_scheduler.optimizer = optimizer
rparams = jiant_runner.RunnerParameters(
local_rank=args.local_rank,
n_gpu=quick_init_out.n_gpu,
fp16=args.fp16,
max_grad_norm=args.max_grad_norm,
)
runner = jiant_runner.JiantRunner(
jiant_task_container=jiant_task_container,
jiant_model=jiant_model,
optimizer_scheduler=optimizer_scheduler,
device=quick_init_out.device,
rparams=rparams,
log_writer=quick_init_out.log_writer,
)
return runner
