def predict_tags(
text: Optional[
str] = "Transfer learning with BERT for self-supervised learning",
run_id: str = open(Path(config.MODEL_DIR, "run_id.txt")).read(),
) -> Dict:
"""Predict tags for a give input text using a trained model.
Warning:
Make sure that you have a trained model first!
Args:
text (str, optional): Input text to predict tags for.
Defaults to "Transfer learning with BERT for self-supervised learning".
run_id (str): ID of the model run to load artifacts. Defaults to run ID in config.MODEL_DIR.
Raises:
ValueError: Run id doesn't exist in experiment.
Returns:
Predicted tags for input text.
"""
# Predict
artifacts = main.load_artifacts(run_id=run_id)
prediction = predict.predict(texts=[text], artifacts=artifacts)
logger.info(json.dumps(prediction, indent=2))
return prediction
def predict_tags(
text: Optional[
str] = "Transfer learning with BERT for self-supervised learning",
model_dir: Path = config.MODEL_DIR,
) -> Dict:
"""Predict tags for a give input text using a trained model.
Warning:
Make sure that you have a trained model first!
Args:
text (str, optional): Input text to predict tags for.
Defaults to "Transfer learning with BERT for self-supervised learning".
model_dir (Path): location of model artifacts. Defaults to config.MODEL_DIR.
Raises:
ValueError: Run id doesn't exist in experiment.
Returns:
Predicted tags for input text.
"""
# Predict
artifacts = main.load_artifacts(model_dir=model_dir)
prediction = predict.predict(texts=[text], artifacts=artifacts)
logger.info(json.dumps(prediction, indent=2))
return prediction
def predict_tags(
text: str = "Transfer learning with BERT for self-supervised learning",
run_id: str = "",
) -> Dict:
"""Predict tags for a give input text using a trained model.
Warning:
Make sure that you have a trained model first!
Args:
text (str, optional): Input text to predict tags for.
Defaults to "Transfer learning with BERT for self-supervised learning".
run_id (str, optional): ID of the run to load model artifacts from.
Defaults to model with lowest `best_val_loss` from the `best` experiment.
Returns:
Predicted tags for input text.
"""
# Get best run
if not run_id:
experiment_id = mlflow.get_experiment_by_name("best").experiment_id
all_runs = mlflow.search_runs(
experiment_ids=experiment_id,
order_by=["metrics.best_val_loss ASC"],
)
run_id = all_runs.iloc[0].run_id
# Predict
prediction = predict.predict(texts=[text], run_id=run_id)
logger.info(json.dumps(prediction, indent=2))
return prediction
def train_model(
params_fp: Path = Path(config.CONFIG_DIR, "params.json"),
model_dir: Optional[Path] = Path(config.MODEL_DIR),
experiment_name: Optional[str] = "best",
run_name: Optional[str] = "model",
) -> None:
"""Train a model using the specified parameters.
Args:
params_fp (Path, optional): Parameters to use for training. Defaults to `config/params.json`.
model_dir (Path): location of model artifacts. Defaults to config.MODEL_DIR.
experiment_name (str, optional): Name of the experiment to save the run to. Defaults to `best`.
run_name (str, optional): Name of the run. Defaults to `model`.
"""
# Set experiment and start run
params = Namespace(**utils.load_dict(filepath=params_fp))
# Start run
mlflow.set_experiment(experiment_name=experiment_name)
with mlflow.start_run(run_name=run_name):
run_id = mlflow.active_run().info.run_id
# Train
artifacts = main.run(params=params)
# Set tags
tags = {}
mlflow.set_tags(tags)
# Log metrics
performance = artifacts["performance"]
logger.info(json.dumps(performance["overall"], indent=2))
metrics = {
"precision": performance["overall"]["precision"],
"recall": performance["overall"]["recall"],
"f1": performance["overall"]["f1"],
"best_val_loss": artifacts["loss"],
"behavioral_score": performance["behavioral"]["score"],
"slices_f1": performance["slices"]["overall"]["f1"],
}
mlflow.log_metrics(metrics)
# Log artifacts
with tempfile.TemporaryDirectory() as dp:
utils.save_dict(vars(artifacts["params"]),
Path(dp, "params.json"),
cls=NumpyEncoder)
utils.save_dict(performance, Path(dp, "performance.json"))
artifacts["label_encoder"].save(Path(dp, "label_encoder.json"))
artifacts["tokenizer"].save(Path(dp, "tokenizer.json"))
torch.save(artifacts["model"].state_dict(), Path(dp, "model.pt"))
mlflow.log_artifacts(dp)
mlflow.log_params(vars(artifacts["params"]))
# Save for repo
open(Path(model_dir, "run_id.txt"), "w").write(run_id)
utils.save_dict(vars(params),
Path(model_dir, "params.json"),
cls=NumpyEncoder)
utils.save_dict(performance, Path(model_dir, "performance.json"))
def health_check():
response = requests.get(
f"http://0.0.0.0:{os.environ.get('PORT', 5000)}/",
headers=headers,
)
results = json.loads(response.text)
logger.info(json.dumps(results, indent=4))
def load_best_artifacts():
global runs, run_ids, best_artifacts, best_run_id
runs = utils.get_sorted_runs(experiment_name="best",
order_by=["metrics.f1 DESC"])
run_ids = [run["run_id"] for run in runs]
best_run_id = run_ids[0]
best_artifacts = predict.load_artifacts(run_id=best_run_id)
logger.info(
"Loaded trained model and other required artifacts for inference!")
def set_artifact_metadata():
"""Set the artifact URI for all experiments and runs.
Used when transferring experiments from other locations (ex. Colab).
Note:
check out the [optimize.ipynb](https://colab.research.google.com/github/GokuMohandas/applied-ml/blob/main/notebooks/optimize.ipynb){:target="_blank"} notebook for how to train on Google Colab and transfer to local.
"""
def set_artifact_location(var, fp):
"""Set variable's yaml value on file at fp."""
with open(fp) as f:
metadata = yaml.load(f)
# Set new value
experiment_id = metadata[var].split("/")[-1]
artifact_location = Path("file://", config.EXPERIMENTS_DIR,
experiment_id)
metadata[var] = str(artifact_location)
with open(fp, "w") as f:
yaml.dump(metadata, f)
def set_artifact_uri(var, fp):
"""Set variable's yaml value on file at fp."""
with open(fp) as f:
metadata = yaml.load(f)
# Set new value
experiment_id = metadata[var].split("/")[-3]
run_id = metadata[var].split("/")[-2]
artifact_uri = Path(
"file://",
config.EXPERIMENTS_DIR,
experiment_id,
run_id,
"artifacts",
)
metadata[var] = str(artifact_uri)
with open(fp, "w") as f:
yaml.dump(metadata, f)
# Get artifact location
experiment_meta_yamls = list(
Path(config.EXPERIMENTS_DIR).glob("*/meta.yaml"))
for meta_yaml in experiment_meta_yamls:
set_artifact_location(var="artifact_location", fp=meta_yaml)
logger.info(f"Set artfifact location for {meta_yaml}")
# Change artifact URI
run_meta_yamls = list(Path(config.EXPERIMENTS_DIR).glob("*/*/meta.yaml"))
for meta_yaml in run_meta_yamls:
set_artifact_uri(var="artifact_uri", fp=meta_yaml)
logger.info(f"Set artfifact URI for {meta_yaml}")
def update_behavioral_report(run_id):
with mlflow.start_run(run_id=run_id):
# Generate behavioral report
artifacts = main.load_artifacts(run_id=run_id)
behavioral_report = eval.get_behavioral_report(artifacts=artifacts)
mlflow.log_metric("behavioral_score", behavioral_report["score"])
# Log artifacts
with tempfile.TemporaryDirectory() as dp:
utils.save_dict(behavioral_report,
Path(dp, "behavioral_report.json"))
mlflow.log_artifacts(dp)
logger.info(f"Updated behavioral report for run_id {run_id}")
def performance(
author: str = config.AUTHOR,
repo: str = config.REPO,
tag: str = "workspace",
verbose: bool = True,
):
if tag == "workspace":
performance = utils.load_dict(filepath=Path(config.MODEL_DIR, "performance.json"))
else:
url = f"https://raw.githubusercontent.com/{author}/{repo}/{tag}/model/performance.json"
performance = utils.load_json_from_url(url=url)
if verbose:
logger.info(json.dumps(performance, indent=2))
return performance
def train_model(
args_fp: Path = Path(config.CONFIG_DIR, "args.json"),
experiment_name: Optional[str] = "best",
run_name: Optional[str] = "model",
) -> None:
"""Train a model using the specified parameters.
Args:
args_fp (Path, optional): Location of arguments to use for training. Defaults to `config/args.json`.
experiment_name (str, optional): Name of the experiment to save the run to. Defaults to `best`.
run_name (str, optional): Name of the run. Defaults to `model`.
"""
# Set experiment and start run
args = Namespace(**utils.load_dict(filepath=args_fp))
# Start run
mlflow.set_experiment(experiment_name=experiment_name)
with mlflow.start_run(run_name=run_name
) as run: # NOQA: F841 (assigned to but never used)
# Train
artifacts = main.run(args=args)
# Set tags
tags = {"data_version": artifacts["data_version"]}
mlflow.set_tags(tags)
# Log metrics
performance = artifacts["performance"]
logger.info(json.dumps(performance["overall"], indent=2))
metrics = {
"precision": performance["overall"]["precision"],
"recall": performance["overall"]["recall"],
"f1": performance["overall"]["f1"],
"best_val_loss": artifacts["loss"],
"behavioral_score": artifacts["behavioral_report"]["score"],
"slices_f1": performance["slices"]["f1"],
}
mlflow.log_metrics(metrics)
# Log artifacts
with tempfile.TemporaryDirectory() as dp:
artifacts["label_encoder"].save(Path(dp, "label_encoder.json"))
artifacts["tokenizer"].save(Path(dp, "tokenizer.json"))
torch.save(artifacts["model"].state_dict(), Path(dp, "model.pt"))
utils.save_dict(performance, Path(dp, "performance.json"))
utils.save_dict(artifacts["behavioral_report"],
Path(dp, "behavioral_report.json"))
mlflow.log_artifacts(dp)
mlflow.log_params(vars(artifacts["args"]))
def predict():
data = {
"run_id": "",
"texts": [
{"text": "Transfer learning with transformers for self-supervised learning."},
{"text": "Generative adversarial networks in both PyTorch and TensorFlow."},
],
}
response = requests.post(
f"http://0.0.0.0:{os.environ.get('PORT', 5000)}/predict",
headers=headers,
data=json.dumps(data),
)
results = json.loads(response.text)
logger.info(json.dumps(results, indent=4))
def predict_tags(
text: Optional[
str] = "Transfer learning with BERT for self-supervised learning",
experiment_name: Optional[str] = "best",
run_id: Optional[str] = "",
) -> Dict:
"""Predict tags for a give input text using a trained model.
Warning:
Make sure that you have a trained model first!
Args:
text (str, optional): Input text to predict tags for.
Defaults to "Transfer learning with BERT for self-supervised learning".
experiment_name (str, optional): Name of the experiment to fetch run from.
run_id (str, optional): ID of the run to load model artifacts from.
Defaults to run with highest F1 score.
Raises:
ValueError: Run id doesn't exist in experiment.
Returns:
Predicted tags for input text.
"""
# Get sorted runs
runs = utils.get_sorted_runs(
experiment_name=experiment_name,
order_by=["metrics.f1 DESC"],
)
run_ids = [run["run_id"] for run in runs]
# Get best run
if not run_id:
run_id = run_ids[0]
# Validate run id
if run_id not in run_ids: # pragma: no cover, simple value check
raise ValueError(
f"Run_id {run_id} does not exist in experiment {experiment_name}")
# Predict
artifacts = main.load_artifacts(run_id=run_id)
prediction = predict.predict(texts=[text], artifacts=artifacts)
logger.info(json.dumps(prediction, indent=2))
return prediction
def download_data():
"""Download data from online to local drive.
Note:
We could've just copied files from `datasets` but
we'll use this later on with other data sources.
"""
# Download data
projects_url = "https://raw.githubusercontent.com/GokuMohandas/applied-ml/main/datasets/projects.json"
tags_url = "https://raw.githubusercontent.com/GokuMohandas/applied-ml/main/datasets/tags.json"
projects = utils.load_json_from_url(url=projects_url)
tags = utils.load_json_from_url(url=tags_url)
# Save data
projects_fp = Path(config.DATA_DIR, "projects.json")
tags_fp = Path(config.DATA_DIR, "tags.json")
utils.save_dict(d=projects, filepath=projects_fp)
utils.save_dict(d=tags, filepath=tags_fp)
logger.info("✅ Data downloaded!")
def optimize(
params_fp: Path = Path(config.CONFIG_DIR, "params.json"),
study_name: Optional[str] = "optimization",
num_trials: int = 100,
) -> None:
"""Optimize a subset of hyperparameters towards an objective.
This saves the best trial's parameters into `config/params.json`.
Args:
params_fp (Path, optional): Location of parameters (just using num_samples,
num_epochs, etc.) to use for training.
Defaults to `config/params.json`.
study_name (str, optional): Name of the study to save trial runs under. Defaults to `optimization`.
num_trials (int, optional): Number of trials to run. Defaults to 100.
"""
# Starting parameters (not actually used but needed for set up)
params = Namespace(**utils.load_dict(filepath=params_fp))
# Optimize
pruner = optuna.pruners.MedianPruner(n_startup_trials=5, n_warmup_steps=5)
study = optuna.create_study(study_name=study_name,
direction="maximize",
pruner=pruner)
mlflow_callback = MLflowCallback(tracking_uri=mlflow.get_tracking_uri(),
metric_name="f1")
study.optimize(
lambda trial: main.objective(params, trial),
n_trials=num_trials,
callbacks=[mlflow_callback],
)
# All trials
trials_df = study.trials_dataframe()
trials_df = trials_df.sort_values(["value"], ascending=False)
# Best trial
logger.info(f"Best value (f1): {study.best_trial.value}")
params = {**params.__dict__, **study.best_trial.params}
params["threshold"] = study.best_trial.user_attrs["threshold"]
with open(params_fp, "w") as fp:
json.dump(params, fp=fp, indent=2, cls=NumpyEncoder)
logger.info(json.dumps(params, indent=2, cls=NumpyEncoder))
def optimize(num_trials: int = 100) -> None:
"""Optimize a subset of hyperparameters towards an objective.
This saves the best trial's arguments into `config/args.json`.
Args:
num_trials (int, optional): Number of trials to run. Defaults to 100.
"""
# Starting arguments (not actually used but needed for set up)
args_fp = Path(config.CONFIG_DIR, "args.json")
args = Namespace(**utils.load_dict(filepath=args_fp))
# Optimize
pruner = optuna.pruners.MedianPruner(n_startup_trials=5, n_warmup_steps=5)
study = optuna.create_study(
study_name="optimization", direction="maximize", pruner=pruner
)
mlflow_callback = MLflowCallback(
tracking_uri=mlflow.get_tracking_uri(), metric_name="f1"
)
study.optimize(
lambda trial: train.objective(args, trial),
n_trials=num_trials,
callbacks=[mlflow_callback],
)
# All trials
trials_df = study.trials_dataframe()
trials_df = trials_df.sort_values(
["value"], ascending=False
) # sort by metric
trials_df.to_csv(
Path(config.EXPERIMENTS_DIR, "trials.csv"), index=False
) # save
# Best trial
logger.info(f"Best value (f1): {study.best_trial.value}")
params = {**args.__dict__, **study.best_trial.params}
params["threshold"] = study.best_trial.user_attrs["threshold"]
with open(Path(config.CONFIG_DIR, "args.json"), "w") as fp:
json.dump(params, fp=fp, indent=2, cls=NumpyEncoder)
logger.info(json.dumps(params, indent=2, cls=NumpyEncoder))
def objective(args: Namespace, trial: optuna.trial._trial.Trial) -> float:
"""Objective function for optimization trials.
Args:
args (Namespace): Input arguments for each trial (see `config/args.json`) for argument names.
trial (optuna.trial._trial.Trial): Optuna optimization trial.
Returns:
F1 score from evaluating the trained model on the test data split.
"""
# Paramters (to tune)
args.embedding_dim = trial.suggest_int("embedding_dim", 128, 512)
args.num_filters = trial.suggest_int("num_filters", 128, 512)
args.hidden_dim = trial.suggest_int("hidden_dim", 128, 512)
args.dropout_p = trial.suggest_uniform("dropout_p", 0.3, 0.8)
args.lr = trial.suggest_loguniform("lr", 5e-5, 5e-4)
# Train (can move some of these outside for efficiency)
logger.info(f"\nTrial {trial.number}:")
logger.info(json.dumps(trial.params, indent=2))
artifacts = run(args=args, trial=trial)
# Set additional attributes
args = artifacts["args"]
performance = artifacts["performance"]
logger.info(json.dumps(performance["overall"], indent=2))
trial.set_user_attr("threshold", args.threshold)
trial.set_user_attr("precision", performance["overall"]["precision"])
trial.set_user_attr("recall", performance["overall"]["recall"])
trial.set_user_attr("f1", performance["overall"]["f1"])
return performance["overall"]["f1"]
def diff(
author: str = config.AUTHOR,
repo: str = config.REPO,
tag_a: str = "workspace",
tag_b: str = "",
): # pragma: no cover, can't be certain what diffs will exist
# Tag b
if tag_b == "":
tags_url = f"https://api.github.com/repos/{author}/{repo}/tags"
tag_b = utils.load_json_from_url(url=tags_url)[0]["name"]
logger.info(f"Comparing {tag_a} with {tag_b}:")
# Params
params_a = params(author=author, repo=repo, tag=tag_a, verbose=False)
params_b = params(author=author, repo=repo, tag=tag_b, verbose=False)
params_diff = utils.dict_diff(d_a=params_a, d_b=params_b, d_a_name=tag_a, d_b_name=tag_b)
logger.info(f"Parameter differences: {json.dumps(params_diff, indent=2)}")
# Performance
performance_a = performance(author=author, repo=repo, tag=tag_a, verbose=False)
performance_b = performance(author=author, repo=repo, tag=tag_b, verbose=False)
performance_diff = utils.dict_diff(
d_a=performance_a, d_b=performance_b, d_a_name=tag_a, d_b_name=tag_b
)
logger.info(f"Performance differences: {json.dumps(performance_diff, indent=2)}")
return params_diff, performance_diff
def train_model(args_fp: Path = Path(config.CONFIG_DIR, "args.json")) -> None:
"""Train a model using the specified parameters.
Args:
args_fp (Path, optional): Location of arguments to use for training.
Defaults to `config/args.json`.
"""
# Set experiment and start run
args = Namespace(**utils.load_dict(filepath=args_fp))
# Start run
mlflow.set_experiment(experiment_name="best")
with mlflow.start_run(
run_name="cnn"
) as run: # NOQA: F841 (assigned to but never used)
# Train
artifacts = train.run(args=args)
# Log key metrics
performance = artifacts["performance"]
loss = artifacts["loss"]
mlflow.log_metrics({"precision": performance["overall"]["precision"]})
mlflow.log_metrics({"recall": performance["overall"]["recall"]})
mlflow.log_metrics({"f1": performance["overall"]["f1"]})
mlflow.log_metrics({"best_val_loss": loss})
# Log artifacts
args = artifacts["args"]
model = artifacts["model"]
label_encoder = artifacts["label_encoder"]
tokenizer = artifacts["tokenizer"]
with tempfile.TemporaryDirectory() as fp:
label_encoder.save(Path(fp, "label_encoder.json"))
tokenizer.save(Path(fp, "tokenizer.json"))
torch.save(model.state_dict(), Path(fp, "model.pt"))
utils.save_dict(performance, Path(fp, "performance.json"))
mlflow.log_artifacts(fp)
mlflow.log_params(vars(args))
logger.info(json.dumps(performance["overall"], indent=2))
def train(
self,
num_epochs: int,
patience: int,
train_dataloader: torch.utils.data.DataLoader,
val_dataloader: torch.utils.data.DataLoader,
) -> Tuple:
"""Training loop.
Args:
num_epochs (int): Maximum number of epochs to train for (can stop earlier based on performance).
patience (int): Number of acceptable epochs for continuous degrading performance.
train_dataloader (torch.utils.data.DataLoader): Dataloader object with training data split.
val_dataloader (torch.utils.data.DataLoader): Dataloader object with validation data split.
Raises:
optuna.TrialPruned: Early stopping of the optimization trial if poor performance.
Returns:
The best validation loss and the trained model from that point.
"""
best_val_loss = np.inf
best_model = None
_patience = patience
for epoch in range(num_epochs):
# Steps
train_loss = self.train_step(dataloader=train_dataloader)
val_loss, _, _ = self.eval_step(dataloader=val_dataloader)
self.scheduler.step(val_loss)
# Pruning based on the intermediate value
if self.trial:
self.trial.report(val_loss, epoch)
if self.trial.should_prune(): # pragma: no cover, optuna pruning
logger.info("Unpromising trial pruned!")
raise optuna.TrialPruned()
# Early stopping
if val_loss < best_val_loss:
best_val_loss = val_loss
best_model = self.model
_patience = patience # reset _patience
else: # pragma: no cover, simple subtraction
_patience -= 1
if not _patience: # pragma: no cover, simple break
logger.info("Stopping early!")
break
# Logging
logger.info(
f"Epoch: {epoch+1} | "
f"train_loss: {train_loss:.5f}, "
f"val_loss: {val_loss:.5f}, "
f"lr: {self.optimizer.param_groups[0]['lr']:.2E}, "
f"_patience: {_patience}"
)
return best_val_loss, best_model
def clean_experiments(experiments_to_keep: str = "best"):
"""Removes all experiments besides the
ones specified in `experiments_to_keep`.
Args:
experiments_to_keep (str): comma separated string of experiments to keep.
"""
# Get experiments to keep
experiments_to_keep = list(
set([exp.strip() for exp in experiments_to_keep.split(",")]))
if not len(experiments_to_keep) or not experiments_to_keep[0]:
raise ValueError("You must keep at least one experiment.")
# Filter and delete
client = mlflow.tracking.MlflowClient()
for experiment in client.list_experiments():
if experiment.name not in experiments_to_keep: # pragma: no cover, mlflow function
logger.info(f"Deleting Experiment {experiment.name}")
client.delete_experiment(experiment_id=experiment.experiment_id)
# Delete MLFlow trash
shutil.rmtree(Path(config.EXPERIMENTS_DIR, ".trash"))
logger.info(f"Cleared experiments besides {experiments_to_keep}")
max_filter_size=args.max_filter_size)
val_dataset = data.CNNTextDataset(X=X_val,
y=y_val,
max_filter_size=args.max_filter_size)
test_dataset = data.CNNTextDataset(X=X_test,
y=y_test,
max_filter_size=args.max_filter_size)
train_dataloader = train_dataset.create_dataloader(
batch_size=args.batch_size)
val_dataloader = val_dataset.create_dataloader(batch_size=args.batch_size)
test_dataloader = test_dataset.create_dataloader(
batch_size=args.batch_size)
# Load artifacts
runs = utils.get_sorted_runs(experiment_name="best",
order_by=["metrics.f1 DESC"])
run_ids = [run["run_id"] for run in runs]
artifacts = main.load_artifacts(run_id=run_ids[0],
device=torch.device("cpu"))
# Evaluation
device = torch.device("cpu")
performance, behavioral_report = evaluate(
artifacts=artifacts,
dataloader=test_dataloader,
df=test_df,
device=device,
)
logger.info(json.dumps(performance, indent=2))
logger.info(json.dumps(behavioral_report, indent=2))
def run(args: Namespace, trial: optuna.trial._trial.Trial = None) -> Dict:
"""Operations for training.
1. Set seed
2. Set device
3. Load data
4. Clean data
5. Preprocess data
6. Encode labels
7. Split data
8. Tokenize inputs
9. Create dataloaders
10. Initialize model
11. Train model
12. Evaluate model
Args:
args (Namespace): Input arguments for operations.
trial (optuna.trial._trial.Trail, optional): Optuna optimization trial. Defaults to None.
Returns:
Artifacts to save and load for later.
"""
# 1. Set seed
utils.set_seed(seed=args.seed)
# 2. Set device
device = utils.set_device(cuda=args.cuda)
# 3. Load data
df, projects_dict, tags_dict = data.load(
shuffle=args.shuffle, num_samples=args.num_samples
)
# 4. Clean data
df, tags_dict, tags_above_frequency = data.clean(
df=df, tags_dict=tags_dict, min_tag_freq=args.min_tag_freq
)
# 5. Preprocess data
df.text = df.text.apply(data.preprocess, lower=args.lower, stem=args.stem)
# 6. Encode labels
y, class_weights, label_encoder = data.encode_labels(labels=df.tags)
# 7. Split data
utils.set_seed(seed=args.seed) # needed for skmultilearn
X_train, X_val, X_test, y_train, y_val, y_test = data.split(
X=df.text.to_numpy(), y=y, train_size=args.train_size
)
# 8. Tokenize inputs
X_train, tokenizer = data.tokenize_text(
X=X_train, char_level=args.char_level
)
X_val, _ = data.tokenize_text(
X=X_val, char_level=args.char_level, tokenizer=tokenizer
)
X_test, _ = data.tokenize_text(
X=X_test, char_level=args.char_level, tokenizer=tokenizer
)
# 9. Create dataloaders
train_dataloader = data.get_dataloader(
data=[X_train, y_train],
max_filter_size=args.max_filter_size,
batch_size=args.batch_size,
)
val_dataloader = data.get_dataloader(
data=[X_val, y_val],
max_filter_size=args.max_filter_size,
batch_size=args.batch_size,
)
test_dataloader = data.get_dataloader(
data=[X_test, y_test],
max_filter_size=args.max_filter_size,
batch_size=args.batch_size,
)
# 10. Initialize model
model = models.initialize_model(
args=args,
vocab_size=len(tokenizer),
num_classes=len(label_encoder),
device=device,
)
# 11. Train model
logger.info(
f"Arguments: {json.dumps(args.__dict__, indent=2, cls=NumpyEncoder)}"
)
args, model, loss = train(
args=args,
train_dataloader=train_dataloader,
val_dataloader=val_dataloader,
model=model,
device=device,
class_weights=class_weights,
trial=trial,
)
# 12. Evaluate model
device = torch.device("cpu")
performance = evaluate(
dataloader=test_dataloader,
model=model.to(device),
device=device,
threshold=args.threshold,
classes=label_encoder.classes,
)
return {
"args": args,
"label_encoder": label_encoder,
"tokenizer": tokenizer,
"model": model,
"loss": loss,
"performance": performance,
}
def diff(commit_a: str = "workspace",
commit_b: str = "head"): # pragma: no cover
"""Compare relevant differences (params, metrics) between commits.
Inspired by DVC's `dvc metrics diff`but repurposed to
display diffs pertinent to our experiments.
Args:
commit_a (str, optional): Primary commit. Defaults to "workspace".
commit_b (str, optional): Commit to compare to. Defaults to "head".
Raises:
ValueError: Invalid commit.
"""
diffs = {}
commits = ["a", "b"]
if commit_a.lower() in ("head", "current"):
commit_a = "main"
if commit_b.lower() in ("head", "current"):
commit_b = "main"
# Get params
params = {"a": {}, "b": {}}
for i, commit in enumerate([commit_a, commit_b]):
if commit == "workspace":
params[commits[i]] = utils.load_dict(
filepath=Path(config.CONFIG_DIR, "params.json"))
continue
params_url = (
f"https://raw.githubusercontent.com/GokuMohandas/applied-ml/{commit}/model/params.json"
)
params[commits[i]] = utils.load_json_from_url(url=params_url)
# Parameter differences
diffs["params"] = {}
for arg in params["a"]:
a = params["a"][arg]
b = params["b"][arg]
if a != b:
diffs["params"][arg] = {commit_a: a, commit_b: b}
logger.info(
f"Parameter differences:\n{json.dumps(diffs['params'], indent=2)}")
# Get metrics
metrics = {"a": {}, "b": {}}
for i, commit in enumerate([commit_a, commit_b]):
if commit == "workspace":
metrics[commits[i]] = utils.load_dict(
filepath=Path(config.MODEL_DIR, "performance.json"))
continue
metrics_url = f"https://raw.githubusercontent.com/GokuMohandas/applied-ml/{commit}/model/performance.json"
metrics[commits[i]] = utils.load_json_from_url(url=metrics_url)
# Recursively flatten
metrics_a = pd.json_normalize(metrics["a"],
sep=".").to_dict(orient="records")[0]
metrics_b = pd.json_normalize(metrics["b"],
sep=".").to_dict(orient="records")[0]
if metrics_a.keys() != metrics_b.keys():
raise Exception(
"Cannot compare these commits because they have different metrics."
)
# Metric differences
diffs["metrics"] = {}
diffs["metrics"]["improvements"] = {}
diffs["metrics"]["regressions"] = {}
for metric in metrics_a:
if ((metric in metrics_b) and (metrics_a[metric] != metrics_b[metric])
and (isinstance(metrics_a[metric], numbers.Number))
and (metric.split(".")[-1] != "num_samples")):
item = {
commit_a: metrics_a[metric],
commit_b: metrics_b[metric],
"diff": metrics_a[metric] - metrics_b[metric],
}
if item["diff"] >= 0.0:
diffs["metrics"]["improvements"][metric] = item
else:
diffs["metrics"]["regressions"][metric] = item
logger.info(
f"Metric differences:\n{json.dumps(diffs['metrics'], indent=2)}")
return diffs
def load_artifacts():
global artifacts
run_id = open(Path(config.MODEL_DIR, "run_id.txt")).read()
artifacts = main.load_artifacts(run_id=run_id)
logger.info("Ready for inference!")
def load_artifacts():
global artifacts
artifacts = main.load_artifacts(model_dir=config.MODEL_DIR)
logger.info("Ready for inference!")
def run(params: Namespace, trial: optuna.trial._trial.Trial = None) -> Dict:
"""Operations for training.
Args:
params (Namespace): Input parameters for operations.
trial (optuna.trial._trial.Trail, optional): Optuna optimization trial. Defaults to None.
Returns:
Artifacts to save and load for later.
"""
# 1. Set seed
utils.set_seed(seed=params.seed)
# 2. Set device
device = utils.set_device(cuda=params.cuda)
# 3. Load data
projects_fp = Path(config.DATA_DIR, "projects.json")
tags_fp = Path(config.DATA_DIR, "tags.json")
projects = utils.load_dict(filepath=projects_fp)
tags_dict = utils.list_to_dict(utils.load_dict(filepath=tags_fp),
key="tag")
df = pd.DataFrame(projects)
if params.shuffle:
df = df.sample(frac=1).reset_index(drop=True)
df = df[:params.subset] # None = all samples
# 4. Prepare data (feature engineering, filter, clean)
df, tags_above_freq, tags_below_freq = data.prepare(
df=df,
include=list(tags_dict.keys()),
exclude=config.EXCLUDED_TAGS,
min_tag_freq=params.min_tag_freq,
)
params.num_samples = len(df)
# 5. Preprocess data
df.text = df.text.apply(data.preprocess,
lower=params.lower,
stem=params.stem)
# 6. Encode labels
labels = df.tags
label_encoder = data.MultiLabelLabelEncoder()
label_encoder.fit(labels)
y = label_encoder.encode(labels)
# Class weights
all_tags = list(itertools.chain.from_iterable(labels.values))
counts = np.bincount(
[label_encoder.class_to_index[class_] for class_ in all_tags])
class_weights = {i: 1.0 / count for i, count in enumerate(counts)}
# 7. Split data
utils.set_seed(seed=params.seed) # needed for skmultilearn
X = df.text.to_numpy()
X_train, X_, y_train, y_ = data.iterative_train_test_split(
X=X, y=y, train_size=params.train_size)
X_val, X_test, y_val, y_test = data.iterative_train_test_split(
X=X_, y=y_, train_size=0.5)
test_df = pd.DataFrame({
"text": X_test,
"tags": label_encoder.decode(y_test)
})
# 8. Tokenize inputs
tokenizer = data.Tokenizer(char_level=params.char_level)
tokenizer.fit_on_texts(texts=X_train)
X_train = np.array(tokenizer.texts_to_sequences(X_train), dtype=object)
X_val = np.array(tokenizer.texts_to_sequences(X_val), dtype=object)
X_test = np.array(tokenizer.texts_to_sequences(X_test), dtype=object)
# 9. Create dataloaders
train_dataset = data.CNNTextDataset(X=X_train,
y=y_train,
max_filter_size=params.max_filter_size)
val_dataset = data.CNNTextDataset(X=X_val,
y=y_val,
max_filter_size=params.max_filter_size)
train_dataloader = train_dataset.create_dataloader(
batch_size=params.batch_size)
val_dataloader = val_dataset.create_dataloader(
batch_size=params.batch_size)
# 10. Initialize model
model = models.initialize_model(
params=params,
vocab_size=len(tokenizer),
num_classes=len(label_encoder),
device=device,
)
# 11. Train model
logger.info(
f"Parameters: {json.dumps(params.__dict__, indent=2, cls=NumpyEncoder)}"
)
params, model, loss = train.train(
params=params,
train_dataloader=train_dataloader,
val_dataloader=val_dataloader,
model=model,
device=device,
class_weights=class_weights,
trial=trial,
)
# 12. Evaluate model
artifacts = {
"params": params,
"label_encoder": label_encoder,
"tokenizer": tokenizer,
"model": model,
"loss": loss,
}
device = torch.device("cpu")
y_true, y_pred, performance = eval.evaluate(df=test_df,
artifacts=artifacts)
artifacts["performance"] = performance
return artifacts