def _evaluate(
self,
model: EmmentalModel,
dataloaders: List[EmmentalDataLoader],
split: Union[List[str], str],
) -> Dict[str, float]:
r"""Evaluate the model.
Args:
model(EmmentalModel): The model to evaluate.
dataloaders(List[EmmentalDataLoader]): The data to evaluate.
split(str): The split to evaluate.
Returns:
dict: The score dict.
"""
if not isinstance(split, list):
valid_split = [split]
else:
valid_split = split
valid_dataloaders = [
dataloader for dataloader in dataloaders if dataloader.split in valid_split
]
return model.score(valid_dataloaders)
def main(args):
# Initialize Emmental
config = parse_args_to_config(args)
emmental.init(log_dir=config["meta_config"]["log_path"], config=config)
# Log configuration into files
cmd_msg = " ".join(sys.argv)
logger.info(f"COMMAND: {cmd_msg}")
write_to_file(f"{emmental.Meta.log_path}/cmd.txt", cmd_msg)
logger.info(f"Config: {emmental.Meta.config}")
write_to_file(f"{emmental.Meta.log_path}/config.txt", emmental.Meta.config)
# Create dataloaders
dataloaders = get_dataloaders(args)
# Assign transforms to dataloaders
aug_dataloaders = []
if args.augment_policy:
for idx in range(len(dataloaders)):
if dataloaders[idx].split in args.train_split:
dataloaders[idx].dataset.transform_cls = Augmentation(
args=args)
config["learner_config"]["task_scheduler_config"][
"task_scheduler"] = AugScheduler(augment_k=args.augment_k,
enlarge=args.augment_enlarge)
emmental.Meta.config["learner_config"]["task_scheduler_config"][
"task_scheduler"] = config["learner_config"]["task_scheduler_config"][
"task_scheduler"]
# Create tasks
model = EmmentalModel(name=f"{args.task}_task")
model.add_task(create_task(args))
# Set cudnn benchmark
cudnn.benchmark = True
# Load the best model from the pretrained model
if config["model_config"]["model_path"] is not None:
model.load(config["model_config"]["model_path"])
if args.train:
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, dataloaders + aug_dataloaders)
# Remove all extra augmentation policy
for idx in range(len(dataloaders)):
dataloaders[idx].dataset.transform_cls = None
scores = model.score(dataloaders)
# Save metrics and models
logger.info(f"Metrics: {scores}")
scores["log_path"] = emmental.Meta.log_path
write_to_json_file(f"{emmental.Meta.log_path}/metrics.txt", scores)
model.save(f"{emmental.Meta.log_path}/last_model.pth")
def main(args):
# Ensure that global state is fresh
Meta.reset()
# Initialize Emmental
config = parse_arg_to_config(args)
emmental.init(config["meta_config"]["log_path"], config=config)
# Save command line argument into file
cmd_msg = " ".join(sys.argv)
logger.info(f"COMMAND: {cmd_msg}")
write_to_file(Meta.log_path, "cmd.txt", cmd_msg)
# Save Emmental config into file
logger.info(f"Config: {Meta.config}")
write_to_file(Meta.log_path, "config.txt", Meta.config)
Meta.config["learner_config"]["global_evaluation_metric_dict"] = {
f"model/SuperGLUE/{split}/score": partial(superglue_scorer,
split=split)
for split in ["val"]
}
# Construct dataloaders and tasks and load slices
dataloaders = []
tasks = []
for task_name in args.task:
task_dataloaders = get_dataloaders(
data_dir=args.data_dir,
task_name=task_name,
splits=["train", "val", "test"],
max_sequence_length=args.max_sequence_length,
max_data_samples=args.max_data_samples,
tokenizer_name=args.bert_model,
batch_size=args.batch_size,
augment=args.augmentations,
)
task = models.model[task_name](
args.bert_model,
last_hidden_dropout_prob=args.last_hidden_dropout_prob)
if args.slices:
logger.info("Initializing task-specific slices")
slice_func_dict = slicing.slice_func_dict[task_name]
# Include general purpose slices
if args.general_slices:
logger.info("Including general slices")
slice_func_dict.update(slicing.slice_func_dict["general"])
task_dataloaders = slicing.add_slice_labels(
task_name, task_dataloaders, slice_func_dict)
slice_tasks = slicing.add_slice_tasks(task_name, task,
slice_func_dict,
args.slice_hidden_dim)
tasks.extend(slice_tasks)
else:
tasks.append(task)
dataloaders.extend(task_dataloaders)
# Build Emmental model
model = EmmentalModel(name="SuperGLUE", tasks=tasks)
# Load pretrained model if necessary
if Meta.config["model_config"]["model_path"]:
model.load(Meta.config["model_config"]["model_path"])
# Training
if args.train:
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, dataloaders)
# If model is slice-aware, slice scores will be calculated from slice heads
# If model is not slice-aware, manually calculate performance on slices
if not args.slices:
slice_func_dict = {}
slice_keys = args.task
if args.general_slices:
slice_keys.append("general")
for k in slice_keys:
slice_func_dict.update(slicing.slice_func_dict[k])
scores = slicing.score_slices(model, dataloaders, args.task,
slice_func_dict)
else:
scores = model.score(dataloaders)
# Save metrics into file
logger.info(f"Metrics: {scores}")
write_to_file(Meta.log_path, "metrics.txt", scores)
# Save best metrics into file
if args.train:
logger.info(
f"Best metrics: "
f"{emmental_learner.logging_manager.checkpointer.best_metric_dict}"
)
write_to_file(
Meta.log_path,
"best_metrics.txt",
emmental_learner.logging_manager.checkpointer.best_metric_dict,
)
# Save submission file
for task_name in args.task:
dataloaders = [d for d in dataloaders if d.split == "test"]
assert len(dataloaders) == 1
filepath = os.path.join(Meta.log_path, f"{task_name}.jsonl")
make_submission_file(model, dataloaders[0], task_name, filepath)
def main(args):
# Initialize Emmental
config = parse_args_to_config(args)
emmental.init(log_dir=config["meta_config"]["log_path"], config=config)
# Log configuration into files
cmd_msg = " ".join(sys.argv)
logger.info(f"COMMAND: {cmd_msg}")
write_to_file(f"{emmental.Meta.log_path}/cmd.txt", cmd_msg)
logger.info(f"Config: {emmental.Meta.config}")
write_to_file(f"{emmental.Meta.log_path}/config.txt", emmental.Meta.config)
# Create dataloaders
dataloaders = get_dataloaders(args)
config["learner_config"]["task_scheduler_config"][
"task_scheduler"] = AugScheduler(augment_k=args.augment_k,
enlarge=args.augment_enlarge)
emmental.Meta.config["learner_config"]["task_scheduler_config"][
"task_scheduler"] = config["learner_config"]["task_scheduler_config"][
"task_scheduler"]
# Specify parameter group for Adam BERT
def grouped_parameters(model):
no_decay = ["bias", "LayerNorm.weight"]
return [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
emmental.Meta.config["learner_config"]["optimizer_config"]
["l2"],
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
emmental.Meta.config["learner_config"]["optimizer_config"][
"parameters"] = grouped_parameters
# Create tasks
model = EmmentalModel(name=f"{args.task}_task")
model.add_task(create_task(args))
# Load the best model from the pretrained model
if config["model_config"]["model_path"] is not None:
model.load(config["model_config"]["model_path"])
if args.train:
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, dataloaders)
# Remove all extra augmentation policy
for idx in range(len(dataloaders)):
dataloaders[idx].dataset.transform_cls = None
dataloaders[idx].dataset.k = 1
scores = model.score(dataloaders)
# Save metrics and models
logger.info(f"Metrics: {scores}")
scores["log_path"] = emmental.Meta.log_path
write_to_json_file(f"{emmental.Meta.log_path}/metrics.txt", scores)
model.save(f"{emmental.Meta.log_path}/last_model.pth")
task_name: create_task(
task_name, args, datasets[task_name]["nclasses"], emb_layer
)
for task_name in args.task
}
model = EmmentalModel(name="TC_task")
if Meta.config["model_config"]["model_path"]:
model.load(Meta.config["model_config"]["model_path"])
else:
for task_name, task in tasks.items():
model.add_task(task)
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, dataloaders)
scores = model.score(dataloaders)
logger.info(f"Metrics: {scores}")
write_to_json_file(f"{Meta.log_path}/metrics.txt", scores)
if args.checkpointing:
logger.info(
f"Best metrics: "
f"{emmental_learner.logging_manager.checkpointer.best_metric_dict}"
)
write_to_file(
f"{Meta.log_path}/best_metrics.txt",
emmental_learner.logging_manager.checkpointer.best_metric_dict,
)
def test_e2e(caplog):
"""Run an end-to-end test."""
caplog.set_level(logging.INFO)
dirpath = "temp_test_e2e"
use_exact_log_path = False
Meta.reset()
emmental.init(dirpath, use_exact_log_path=use_exact_log_path)
config = {
"meta_config": {
"seed": 0
},
"learner_config": {
"n_epochs": 3,
"optimizer_config": {
"lr": 0.01,
"grad_clip": 100
},
},
"logging_config": {
"counter_unit": "epoch",
"evaluation_freq": 1,
"writer_config": {
"writer": "tensorboard",
"verbose": True
},
"checkpointing": True,
"checkpointer_config": {
"checkpoint_path": None,
"checkpoint_freq": 1,
"checkpoint_metric": {
"model/all/train/loss": "min"
},
"checkpoint_task_metrics": None,
"checkpoint_runway": 1,
"checkpoint_all": False,
"clear_intermediate_checkpoints": True,
"clear_all_checkpoints": True,
},
},
}
emmental.Meta.update_config(config)
# Generate synthetic data
N = 500
X = np.random.random((N, 2)) * 2 - 1
Y1 = (X[:, 0] > X[:, 1] + 0.25).astype(int)
Y2 = (X[:, 0] > X[:, 1] + 0.2).astype(int)
X = [torch.Tensor(X[i]) for i in range(N)]
# Create dataset and dataloader
X_train, X_dev, X_test = (
X[:int(0.8 * N)],
X[int(0.8 * N):int(0.9 * N)],
X[int(0.9 * N):],
)
Y1_train, Y1_dev, Y1_test = (
torch.tensor(Y1[:int(0.8 * N)]),
torch.tensor(Y1[int(0.8 * N):int(0.9 * N)]),
torch.tensor(Y1[int(0.9 * N):]),
)
Y2_train, Y2_dev, Y2_test = (
torch.tensor(Y2[:int(0.8 * N)]),
torch.tensor(Y2[int(0.8 * N):int(0.9 * N)]),
torch.tensor(Y2[int(0.9 * N):]),
)
train_dataset1 = EmmentalDataset(name="synthetic",
X_dict={"data": X_train},
Y_dict={"label1": Y1_train})
train_dataset2 = EmmentalDataset(name="synthetic",
X_dict={"data": X_train},
Y_dict={"label2": Y2_train})
dev_dataset1 = EmmentalDataset(name="synthetic",
X_dict={"data": X_dev},
Y_dict={"label1": Y1_dev})
dev_dataset2 = EmmentalDataset(name="synthetic",
X_dict={"data": X_dev},
Y_dict={"label2": Y2_dev})
test_dataset1 = EmmentalDataset(name="synthetic",
X_dict={"data": X_test},
Y_dict={"label1": Y1_test})
test_dataset2 = EmmentalDataset(name="synthetic",
X_dict={"data": X_test},
Y_dict={"label2": Y2_test})
task_to_label_dict = {"task1": "label1"}
train_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=train_dataset1,
split="train",
batch_size=10,
)
dev_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=dev_dataset1,
split="valid",
batch_size=10,
)
test_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=test_dataset1,
split="test",
batch_size=10,
)
task_to_label_dict = {"task2": "label2"}
train_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=train_dataset2,
split="train",
batch_size=10,
)
dev_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=dev_dataset2,
split="valid",
batch_size=10,
)
test_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=test_dataset2,
split="test",
batch_size=10,
)
# Create task
def ce_loss(task_name, immediate_ouput_dict, Y, active):
module_name = f"{task_name}_pred_head"
return F.cross_entropy(immediate_ouput_dict[module_name][0][active],
(Y.view(-1))[active])
def output(task_name, immediate_ouput_dict):
module_name = f"{task_name}_pred_head"
return F.softmax(immediate_ouput_dict[module_name][0], dim=1)
task_metrics = {"task1": ["accuracy"], "task2": ["accuracy", "roc_auc"]}
tasks = [
EmmentalTask(
name=task_name,
module_pool=nn.ModuleDict({
"input_module":
nn.Linear(2, 8),
f"{task_name}_pred_head":
nn.Linear(8, 2),
}),
task_flow=[
{
"name": "input",
"module": "input_module",
"inputs": [("_input_", "data")],
},
{
"name": f"{task_name}_pred_head",
"module": f"{task_name}_pred_head",
"inputs": [("input", 0)],
},
],
loss_func=partial(ce_loss, task_name),
output_func=partial(output, task_name),
scorer=Scorer(metrics=task_metrics[task_name]),
) for task_name in ["task1", "task2"]
]
# Build model
mtl_model = EmmentalModel(name="all", tasks=tasks)
# Create learner
emmental_learner = EmmentalLearner()
# Learning
emmental_learner.learn(
mtl_model,
[
train_dataloader1, train_dataloader2, dev_dataloader1,
dev_dataloader2
],
)
test1_score = mtl_model.score(test_dataloader1)
test2_score = mtl_model.score(test_dataloader2)
assert test1_score["task1/synthetic/test/accuracy"] >= 0.7
assert (test1_score["model/all/test/macro_average"] ==
test1_score["task1/synthetic/test/accuracy"])
assert test2_score["task2/synthetic/test/accuracy"] >= 0.7
assert test2_score["task2/synthetic/test/roc_auc"] >= 0.7
shutil.rmtree(dirpath)
def main(
conn_string,
max_docs=float("inf"),
parse=False,
first_time=False,
gpu=None,
parallel=4,
log_dir=None,
verbose=False,
):
if not log_dir:
log_dir = "logs"
if verbose:
level = logging.INFO
else:
level = logging.WARNING
dirname = os.path.dirname(os.path.abspath(__file__))
init_logging(log_dir=os.path.join(dirname, log_dir), level=level)
session = Meta.init(conn_string).Session()
# Parsing
logger.info(f"Starting parsing...")
start = timer()
docs, train_docs, dev_docs, test_docs = parse_dataset(
session, dirname, first_time=first_time, parallel=parallel, max_docs=max_docs
)
end = timer()
logger.warning(f"Parse Time (min): {((end - start) / 60.0):.1f}")
logger.info(f"# of train Documents: {len(train_docs)}")
logger.info(f"# of dev Documents: {len(dev_docs)}")
logger.info(f"# of test Documents: {len(test_docs)}")
logger.info(f"Documents: {session.query(Document).count()}")
logger.info(f"Sections: {session.query(Section).count()}")
logger.info(f"Paragraphs: {session.query(Paragraph).count()}")
logger.info(f"Sentences: {session.query(Sentence).count()}")
logger.info(f"Figures: {session.query(Figure).count()}")
start = timer()
Thumbnails = mention_subclass("Thumbnails")
thumbnails_img = MentionFigures()
class HasFigures(_Matcher):
def _f(self, m):
file_path = ""
for prefix in [
f"{dirname}/data/train/html/",
f"{dirname}/data/dev/html/",
f"{dirname}/data/test/html/",
]:
if os.path.exists(prefix + m.figure.url):
file_path = prefix + m.figure.url
if file_path == "":
return False
img = Image.open(file_path)
width, height = img.size
min_value = min(width, height)
return min_value > 50
mention_extractor = MentionExtractor(
session, [Thumbnails], [thumbnails_img], [HasFigures()], parallelism=parallel
)
if first_time:
mention_extractor.apply(docs)
logger.info("Total Mentions: {}".format(session.query(Mention).count()))
ThumbnailLabel = candidate_subclass("ThumbnailLabel", [Thumbnails])
candidate_extractor = CandidateExtractor(
session, [ThumbnailLabel], throttlers=[None], parallelism=parallel
)
if first_time:
candidate_extractor.apply(train_docs, split=0)
candidate_extractor.apply(dev_docs, split=1)
candidate_extractor.apply(test_docs, split=2)
train_cands = candidate_extractor.get_candidates(split=0)
# Sort the dev_cands, which are used for training, for deterministic behavior
dev_cands = candidate_extractor.get_candidates(split=1, sort=True)
test_cands = candidate_extractor.get_candidates(split=2)
end = timer()
logger.warning(f"Candidate Extraction Time (min): {((end - start) / 60.0):.1f}")
logger.info("Total train candidate:\t{}".format(len(train_cands[0])))
logger.info("Total dev candidate:\t{}".format(len(dev_cands[0])))
logger.info("Total test candidate:\t{}".format(len(test_cands[0])))
fin = open(f"{dirname}/data/ground_truth.txt", "r")
gt = set()
for line in fin:
gt.add("::".join(line.lower().split()))
fin.close()
# Labeling
start = timer()
def LF_gt_label(c):
doc_file_id = (
f"{c[0].context.figure.document.name.lower()}.pdf::"
f"{os.path.basename(c[0].context.figure.url.lower())}"
)
return TRUE if doc_file_id in gt else FALSE
gt_dev = [LF_gt_label(cand) for cand in dev_cands[0]]
gt_test = [LF_gt_label(cand) for cand in test_cands[0]]
end = timer()
logger.warning(f"Supervision Time (min): {((end - start) / 60.0):.1f}")
batch_size = 64
input_size = 224
K = 2
emmental.init(log_dir=Meta.log_path, config=emmental_config)
emmental.Meta.config["learner_config"]["task_scheduler_config"][
"task_scheduler"
] = DauphinScheduler(augment_k=K, enlarge=1)
train_dataset = ThumbnailDataset(
"Thumbnail",
dev_cands[0],
gt_dev,
"train",
prob_label=True,
prefix=f"{dirname}/data/dev/html/",
input_size=input_size,
transform_cls=Augmentation(2),
k=K,
)
val_dataset = ThumbnailDataset(
"Thumbnail",
dev_cands[0],
gt_dev,
"valid",
prob_label=False,
prefix=f"{dirname}/data/dev/html/",
input_size=input_size,
k=1,
)
test_dataset = ThumbnailDataset(
"Thumbnail",
test_cands[0],
gt_test,
"test",
prob_label=False,
prefix=f"{dirname}/data/test/html/",
input_size=input_size,
k=1,
)
dataloaders = []
dataloaders.append(
EmmentalDataLoader(
task_to_label_dict={"Thumbnail": "labels"},
dataset=train_dataset,
split="train",
shuffle=True,
batch_size=batch_size,
num_workers=1,
)
)
dataloaders.append(
EmmentalDataLoader(
task_to_label_dict={"Thumbnail": "labels"},
dataset=val_dataset,
split="valid",
shuffle=False,
batch_size=batch_size,
num_workers=1,
)
)
dataloaders.append(
EmmentalDataLoader(
task_to_label_dict={"Thumbnail": "labels"},
dataset=test_dataset,
split="test",
shuffle=False,
batch_size=batch_size,
num_workers=1,
)
)
model = EmmentalModel(name=f"Thumbnail")
model.add_task(
create_task("Thumbnail", n_class=2, model="resnet18", pretrained=True)
)
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, dataloaders)
scores = model.score(dataloaders)
logger.warning("Model Score:")
logger.warning(f"precision: {scores['Thumbnail/Thumbnail/test/precision']:.3f}")
logger.warning(f"recall: {scores['Thumbnail/Thumbnail/test/recall']:.3f}")
logger.warning(f"f1: {scores['Thumbnail/Thumbnail/test/f1']:.3f}")
def test_e2e(caplog):
"""Run an end-to-end test."""
caplog.set_level(logging.INFO)
dirpath = "temp_test_e2e"
Meta.reset()
emmental.init(dirpath)
# Generate synthetic data
N = 50
X = np.random.random((N, 2)) * 2 - 1
Y1 = (X[:, 0] > X[:, 1] + 0.25).astype(int) + 1
Y2 = (-X[:, 0] > X[:, 1] + 0.25).astype(int) + 1
# Create dataset and dataloader
splits = [0.8, 0.1, 0.1]
X_train, X_dev, X_test = [], [], []
Y1_train, Y1_dev, Y1_test = [], [], []
Y2_train, Y2_dev, Y2_test = [], [], []
for i in range(N):
if i <= N * splits[0]:
X_train.append(torch.Tensor(X[i]))
Y1_train.append(Y1[i])
Y2_train.append(Y2[i])
elif i < N * (splits[0] + splits[1]):
X_dev.append(torch.Tensor(X[i]))
Y1_dev.append(Y1[i])
Y2_dev.append(Y2[i])
else:
X_test.append(torch.Tensor(X[i]))
Y1_test.append(Y1[i])
Y2_test.append(Y2[i])
Y1_train = torch.from_numpy(np.array(Y1_train))
Y1_dev = torch.from_numpy(np.array(Y1_dev))
Y1_test = torch.from_numpy(np.array(Y1_test))
Y2_train = torch.from_numpy(np.array(Y1_train))
Y2_dev = torch.from_numpy(np.array(Y2_dev))
Y2_test = torch.from_numpy(np.array(Y2_test))
train_dataset1 = EmmentalDataset(
name="synthetic", X_dict={"data": X_train}, Y_dict={"label1": Y1_train}
)
train_dataset2 = EmmentalDataset(
name="synthetic", X_dict={"data": X_train}, Y_dict={"label2": Y2_train}
)
dev_dataset1 = EmmentalDataset(
name="synthetic", X_dict={"data": X_dev}, Y_dict={"label1": Y1_dev}
)
dev_dataset2 = EmmentalDataset(
name="synthetic", X_dict={"data": X_dev}, Y_dict={"label2": Y2_dev}
)
test_dataset1 = EmmentalDataset(
name="synthetic", X_dict={"data": X_test}, Y_dict={"label1": Y2_test}
)
test_dataset2 = EmmentalDataset(
name="synthetic", X_dict={"data": X_test}, Y_dict={"label2": Y2_test}
)
task_to_label_dict = {"task1": "label1"}
train_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=train_dataset1,
split="train",
batch_size=10,
)
dev_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=dev_dataset1,
split="valid",
batch_size=10,
)
test_dataloader1 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=test_dataset1,
split="test",
batch_size=10,
)
task_to_label_dict = {"task2": "label2"}
train_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=train_dataset2,
split="train",
batch_size=10,
)
dev_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=dev_dataset2,
split="valid",
batch_size=10,
)
test_dataloader2 = EmmentalDataLoader(
task_to_label_dict=task_to_label_dict,
dataset=test_dataset2,
split="test",
batch_size=10,
)
# Create task
def ce_loss(task_name, immediate_ouput_dict, Y, active):
module_name = f"{task_name}_pred_head"
return F.cross_entropy(
immediate_ouput_dict[module_name][0][active], (Y.view(-1) - 1)[active]
)
def output(task_name, immediate_ouput_dict):
module_name = f"{task_name}_pred_head"
return F.softmax(immediate_ouput_dict[module_name][0], dim=1)
task_name = "task1"
task1 = EmmentalTask(
name=task_name,
module_pool=nn.ModuleDict(
{"input_module": nn.Linear(2, 8), f"{task_name}_pred_head": nn.Linear(8, 2)}
),
task_flow=[
{
"name": "input",
"module": "input_module",
"inputs": [("_input_", "data")],
},
{
"name": f"{task_name}_pred_head",
"module": f"{task_name}_pred_head",
"inputs": [("input", 0)],
},
],
loss_func=partial(ce_loss, task_name),
output_func=partial(output, task_name),
scorer=Scorer(metrics=["accuracy", "roc_auc"]),
)
task_name = "task2"
task2 = EmmentalTask(
name=task_name,
module_pool=nn.ModuleDict(
{"input_module": nn.Linear(2, 8), f"{task_name}_pred_head": nn.Linear(8, 2)}
),
task_flow=[
{
"name": "input",
"module": "input_module",
"inputs": [("_input_", "data")],
},
{
"name": f"{task_name}_pred_head",
"module": f"{task_name}_pred_head",
"inputs": [("input", 0)],
},
],
loss_func=partial(ce_loss, task_name),
output_func=partial(output, task_name),
scorer=Scorer(metrics=["accuracy", "roc_auc"]),
)
# Build model
mtl_model = EmmentalModel(name="all", tasks=[task1, task2])
# Create learner
emmental_learner = EmmentalLearner()
# Update learning config
Meta.update_config(
config={"learner_config": {"n_epochs": 10, "optimizer_config": {"lr": 0.01}}}
)
# Learning
emmental_learner.learn(
mtl_model,
[train_dataloader1, train_dataloader2, dev_dataloader1, dev_dataloader2],
)
test1_score = mtl_model.score(test_dataloader1)
test2_score = mtl_model.score(test_dataloader2)
assert test1_score["task1/synthetic/test/accuracy"] >= 0.5
assert test1_score["task1/synthetic/test/roc_auc"] >= 0.6
assert test2_score["task2/synthetic/test/accuracy"] >= 0.5
assert test2_score["task2/synthetic/test/roc_auc"] >= 0.6
shutil.rmtree(dirpath)
def run_model(mode, config, run_config_path=None):
"""
Main run method for Emmental Bootleg models.
Args:
mode: run mode (train, eval, dump_preds, dump_embs)
config: parsed model config
run_config_path: original config path (for saving)
Returns:
"""
# Set up distributed backend and save configuration files
setup(config, run_config_path)
# Load entity symbols
log_rank_0_info(logger, f"Loading entity symbols...")
entity_symbols = EntitySymbols.load_from_cache(
load_dir=os.path.join(config.data_config.entity_dir,
config.data_config.entity_map_dir),
alias_cand_map_file=config.data_config.alias_cand_map,
alias_idx_file=config.data_config.alias_idx_map,
)
# Create tasks
tasks = [NED_TASK]
if config.data_config.type_prediction.use_type_pred is True:
tasks.append(TYPE_PRED_TASK)
# Create splits for data loaders
data_splits = [TRAIN_SPLIT, DEV_SPLIT, TEST_SPLIT]
# Slices are for eval so we only split on test/dev
slice_splits = [DEV_SPLIT, TEST_SPLIT]
# If doing eval, only run on test data
if mode in ["eval", "dump_preds", "dump_embs"]:
data_splits = [TEST_SPLIT]
slice_splits = [TEST_SPLIT]
# We only do dumping if weak labels is True
if mode in ["dump_preds", "dump_embs"]:
if config.data_config[
f"{TEST_SPLIT}_dataset"].use_weak_label is False:
raise ValueError(
f"When calling dump_preds or dump_embs, we require use_weak_label to be True."
)
# Gets embeddings that need to be prepped during data prep or in the __get_item__ method
batch_on_the_fly_kg_adj = get_dataloader_embeddings(config, entity_symbols)
# Gets dataloaders
dataloaders = get_dataloaders(
config,
tasks,
data_splits,
entity_symbols,
batch_on_the_fly_kg_adj,
)
slice_datasets = get_slicedatasets(config, slice_splits, entity_symbols)
configure_optimizer(config)
# Create models and add tasks
if config.model_config.attn_class == "BERTNED":
log_rank_0_info(logger, f"Starting NED-Base Model")
assert (config.data_config.type_prediction.use_type_pred is
False), f"NED-Base does not support type prediction"
assert (
config.data_config.word_embedding.use_sent_proj is False
), f"NED-Base requires word_embeddings.use_sent_proj to be False"
model = EmmentalModel(name="NED-Base")
model.add_tasks(
ned_task.create_task(config, entity_symbols, slice_datasets))
else:
log_rank_0_info(logger, f"Starting Bootleg Model")
model = EmmentalModel(name="Bootleg")
# TODO: make this more general for other tasks -- iterate through list of tasks
# and add task for each
model.add_task(
ned_task.create_task(config, entity_symbols, slice_datasets))
if TYPE_PRED_TASK in tasks:
model.add_task(
type_pred_task.create_task(config, entity_symbols,
slice_datasets))
# Add the mention type embedding to the embedding payload
type_pred_task.update_ned_task(model)
# Print param counts
if mode == "train":
log_rank_0_debug(logger, "PARAMS WITH GRAD\n" + "=" * 30)
total_params = count_parameters(model,
requires_grad=True,
logger=logger)
log_rank_0_info(logger, f"===> Total Params With Grad: {total_params}")
log_rank_0_debug(logger, "PARAMS WITHOUT GRAD\n" + "=" * 30)
total_params = count_parameters(model,
requires_grad=False,
logger=logger)
log_rank_0_info(logger,
f"===> Total Params Without Grad: {total_params}")
# Load the best model from the pretrained model
if config["model_config"]["model_path"] is not None:
model.load(config["model_config"]["model_path"])
# Barrier
if config["learner_config"]["local_rank"] == 0:
torch.distributed.barrier()
# Train model
if mode == "train":
emmental_learner = EmmentalLearner()
emmental_learner._set_optimizer(model)
emmental_learner.learn(model, dataloaders)
if config.learner_config.local_rank in [0, -1]:
model.save(f"{emmental.Meta.log_path}/last_model.pth")
# Multi-gpu DataParallel eval (NOT distributed)
if mode in ["eval", "dump_embs", "dump_preds"]:
# This happens inside EmmentalLearner for training
if (config["learner_config"]["local_rank"] == -1
and config["model_config"]["dataparallel"]):
model._to_dataparallel()
# If just finished training a model or in eval mode, run eval
if mode in ["train", "eval"]:
scores = model.score(dataloaders)
# Save metrics and models
log_rank_0_info(logger, f"Saving metrics to {emmental.Meta.log_path}")
log_rank_0_info(logger, f"Metrics: {scores}")
scores["log_path"] = emmental.Meta.log_path
if config.learner_config.local_rank in [0, -1]:
write_to_file(f"{emmental.Meta.log_path}/{mode}_metrics.txt",
scores)
eval_utils.write_disambig_metrics_to_csv(
f"{emmental.Meta.log_path}/{mode}_disambig_metrics.csv",
scores)
return scores
# If you want detailed dumps, save model outputs
assert mode in [
"dump_preds",
"dump_embs",
], 'Mode must be "dump_preds" or "dump_embs"'
dump_embs = False if mode != "dump_embs" else True
assert (
len(dataloaders) == 1
), f"We should only have length 1 dataloaders for dump_embs and dump_preds!"
final_result_file, final_out_emb_file = None, None
if config.learner_config.local_rank in [0, -1]:
# Setup files/folders
filename = os.path.basename(dataloaders[0].dataset.raw_filename)
log_rank_0_debug(
logger,
f"Collecting sentence to mention map {os.path.join(config.data_config.data_dir, filename)}",
)
sentidx2num_mentions, sent_idx2row = eval_utils.get_sent_idx2num_mens(
os.path.join(config.data_config.data_dir, filename))
log_rank_0_debug(logger, f"Done collecting sentence to mention map")
eval_folder = eval_utils.get_eval_folder(filename)
subeval_folder = os.path.join(eval_folder, "batch_results")
utils.ensure_dir(subeval_folder)
# Will keep track of sentences dumped already. These will only be ones with mentions
all_dumped_sentences = set()
number_dumped_batches = 0
total_mentions_seen = 0
all_result_files = []
all_out_emb_files = []
# Iterating over batches of predictions
for res_i, res_dict in enumerate(
eval_utils.batched_pred_iter(
model,
dataloaders[0],
config.run_config.eval_accumulation_steps,
sentidx2num_mentions,
)):
(
result_file,
out_emb_file,
final_sent_idxs,
mentions_seen,
) = eval_utils.disambig_dump_preds(
res_i,
total_mentions_seen,
config,
res_dict,
sentidx2num_mentions,
sent_idx2row,
subeval_folder,
entity_symbols,
dump_embs,
NED_TASK,
)
all_dumped_sentences.update(final_sent_idxs)
all_result_files.append(result_file)
all_out_emb_files.append(out_emb_file)
total_mentions_seen += mentions_seen
number_dumped_batches += 1
# Dump the sentences that had no mentions and were not already dumped
# Assert all remaining sentences have no mentions
assert all(
v == 0 for k, v in sentidx2num_mentions.items()
if k not in all_dumped_sentences
), (f"Sentences with mentions were not dumped: "
f"{[k for k, v in sentidx2num_mentions.items() if k not in all_dumped_sentences]}"
)
empty_sentidx2row = {
k: v
for k, v in sent_idx2row.items() if k not in all_dumped_sentences
}
empty_resultfile = eval_utils.get_result_file(number_dumped_batches,
subeval_folder)
all_result_files.append(empty_resultfile)
# Dump the outputs
eval_utils.write_data_labels_single(
sentidx2row=empty_sentidx2row,
output_file=empty_resultfile,
filt_emb_data=None,
sental2embid={},
alias_cand_map=entity_symbols.get_alias2qids(),
qid2eid=entity_symbols.get_qid2eid(),
result_alias_offset=total_mentions_seen,
train_in_cands=config.data_config.train_in_candidates,
max_cands=entity_symbols.max_candidates,
dump_embs=dump_embs,
)
log_rank_0_info(
logger,
f"Finished dumping. Merging results across accumulation steps.")
# Final result files for labels and embeddings
final_result_file = os.path.join(eval_folder,
config.run_config.result_label_file)
# Copy labels
output = open(final_result_file, "wb")
for file in all_result_files:
shutil.copyfileobj(open(file, "rb"), output)
output.close()
log_rank_0_info(logger, f"Bootleg labels saved at {final_result_file}")
# Try to copy embeddings
if dump_embs:
final_out_emb_file = os.path.join(
eval_folder, config.run_config.result_emb_file)
log_rank_0_info(
logger,
f"Trying to merge numpy embedding arrays. "
f"If your machine is limited in memory, this may cause OOM errors. "
f"Is that happens, result files should be saved in {subeval_folder}.",
)
all_arrays = []
for i, npfile in enumerate(all_out_emb_files):
all_arrays.append(np.load(npfile))
np.save(final_out_emb_file, np.concatenate(all_arrays))
log_rank_0_info(
logger, f"Bootleg embeddings saved at {final_out_emb_file}")
# Cleanup
try_rmtree(subeval_folder)
return final_result_file, final_out_emb_file
