summaryrefslogtreecommitdiff
path: root/training
diff options
context:
space:
mode:
Diffstat (limited to 'training')
-rw-r--r--training/experiments/default_config_emnist.yml70
-rw-r--r--training/experiments/embedding_experiment.yml64
-rw-r--r--training/experiments/sample_experiment.yml99
-rw-r--r--training/gpu_manager.py62
-rw-r--r--training/prepare_experiments.py34
-rw-r--r--training/run_experiment.py382
-rw-r--r--training/run_sweep.py92
-rw-r--r--training/sweep_emnist.yml26
-rw-r--r--training/sweep_emnist_resnet.yml50
-rw-r--r--training/trainer/__init__.py2
-rw-r--r--training/trainer/callbacks/__init__.py29
-rw-r--r--training/trainer/callbacks/base.py188
-rw-r--r--training/trainer/callbacks/checkpoint.py95
-rw-r--r--training/trainer/callbacks/early_stopping.py108
-rw-r--r--training/trainer/callbacks/lr_schedulers.py77
-rw-r--r--training/trainer/callbacks/progress_bar.py65
-rw-r--r--training/trainer/callbacks/wandb_callbacks.py261
-rw-r--r--training/trainer/train.py325
-rw-r--r--training/trainer/util.py28
19 files changed, 2057 insertions, 0 deletions
diff --git a/training/experiments/default_config_emnist.yml b/training/experiments/default_config_emnist.yml
new file mode 100644
index 0000000..bf2ed0a
--- /dev/null
+++ b/training/experiments/default_config_emnist.yml
@@ -0,0 +1,70 @@
+dataset: EmnistDataset
+dataset_args:
+ sample_to_balance: true
+ subsample_fraction: 0.33
+ transform: null
+ target_transform: null
+ seed: 4711
+
+data_loader_args:
+ splits: [train, val]
+ shuffle: true
+ num_workers: 8
+ cuda: true
+
+model: CharacterModel
+metrics: [accuracy]
+
+network_args:
+ in_channels: 1
+ num_classes: 80
+ depths: [2]
+ block_sizes: [256]
+
+train_args:
+ batch_size: 256
+ epochs: 5
+
+criterion: CrossEntropyLoss
+criterion_args:
+ weight: null
+ ignore_index: -100
+ reduction: mean
+
+optimizer: AdamW
+optimizer_args:
+ lr: 1.e-03
+ betas: [0.9, 0.999]
+ eps: 1.e-08
+ # weight_decay: 5.e-4
+ amsgrad: false
+
+lr_scheduler: OneCycleLR
+lr_scheduler_args:
+ max_lr: 1.e-03
+ epochs: 5
+ anneal_strategy: linear
+
+
+callbacks: [Checkpoint, ProgressBar, EarlyStopping, WandbCallback, WandbImageLogger, OneCycleLR]
+callback_args:
+ Checkpoint:
+ monitor: val_accuracy
+ ProgressBar:
+ epochs: 5
+ log_batch_frequency: 100
+ EarlyStopping:
+ monitor: val_loss
+ min_delta: 0.0
+ patience: 3
+ mode: min
+ WandbCallback:
+ log_batch_frequency: 10
+ WandbImageLogger:
+ num_examples: 4
+ OneCycleLR:
+ null
+verbosity: 1 # 0, 1, 2
+resume_experiment: null
+train: true
+validation_metric: val_accuracy
diff --git a/training/experiments/embedding_experiment.yml b/training/experiments/embedding_experiment.yml
new file mode 100644
index 0000000..1e5f941
--- /dev/null
+++ b/training/experiments/embedding_experiment.yml
@@ -0,0 +1,64 @@
+experiment_group: Embedding Experiments
+experiments:
+ - train_args:
+ transformer_model: false
+ batch_size: &batch_size 256
+ max_epochs: &max_epochs 32
+ input_shape: [[1, 28, 28]]
+ dataset:
+ type: EmnistDataset
+ args:
+ sample_to_balance: true
+ subsample_fraction: null
+ transform: null
+ target_transform: null
+ seed: 4711
+ train_args:
+ num_workers: 8
+ train_fraction: 0.85
+ batch_size: *batch_size
+ model: CharacterModel
+ metrics: []
+ network:
+ type: DenseNet
+ args:
+ growth_rate: 4
+ block_config: [4, 4]
+ in_channels: 1
+ base_channels: 24
+ num_classes: 128
+ bn_size: 4
+ dropout_rate: 0.1
+ classifier: true
+ activation: elu
+ criterion:
+ type: EmbeddingLoss
+ args:
+ margin: 0.2
+ type_of_triplets: semihard
+ optimizer:
+ type: AdamW
+ args:
+ lr: 1.e-02
+ betas: [0.9, 0.999]
+ eps: 1.e-08
+ weight_decay: 5.e-4
+ amsgrad: false
+ lr_scheduler:
+ type: CosineAnnealingLR
+ args:
+ T_max: *max_epochs
+ callbacks: [Checkpoint, ProgressBar, WandbCallback]
+ callback_args:
+ Checkpoint:
+ monitor: val_loss
+ mode: min
+ ProgressBar:
+ epochs: *max_epochs
+ WandbCallback:
+ log_batch_frequency: 10
+ verbosity: 1 # 0, 1, 2
+ resume_experiment: null
+ train: true
+ test: true
+ test_metric: mean_average_precision_at_r
diff --git a/training/experiments/sample_experiment.yml b/training/experiments/sample_experiment.yml
new file mode 100644
index 0000000..8f94475
--- /dev/null
+++ b/training/experiments/sample_experiment.yml
@@ -0,0 +1,99 @@
+experiment_group: Sample Experiments
+experiments:
+ - train_args:
+ batch_size: 256
+ max_epochs: &max_epochs 32
+ dataset:
+ type: EmnistDataset
+ args:
+ sample_to_balance: true
+ subsample_fraction: null
+ transform: null
+ target_transform: null
+ seed: 4711
+ train_args:
+ num_workers: 6
+ train_fraction: 0.8
+
+ model: CharacterModel
+ metrics: [accuracy]
+ # network: MLP
+ # network_args:
+ # input_size: 784
+ # hidden_size: 512
+ # output_size: 80
+ # num_layers: 5
+ # dropout_rate: 0.2
+ # activation_fn: SELU
+ network:
+ type: ResidualNetwork
+ args:
+ in_channels: 1
+ num_classes: 80
+ depths: [2, 2]
+ block_sizes: [64, 64]
+ activation: leaky_relu
+ # network:
+ # type: WideResidualNetwork
+ # args:
+ # in_channels: 1
+ # num_classes: 80
+ # depth: 10
+ # num_layers: 3
+ # width_factor: 4
+ # dropout_rate: 0.2
+ # activation: SELU
+ # network: LeNet
+ # network_args:
+ # output_size: 62
+ # activation_fn: GELU
+ criterion:
+ type: CrossEntropyLoss
+ args:
+ weight: null
+ ignore_index: -100
+ reduction: mean
+ optimizer:
+ type: AdamW
+ args:
+ lr: 1.e-02
+ betas: [0.9, 0.999]
+ eps: 1.e-08
+ # weight_decay: 5.e-4
+ amsgrad: false
+ # lr_scheduler:
+ # type: OneCycleLR
+ # args:
+ # max_lr: 1.e-03
+ # epochs: *max_epochs
+ # anneal_strategy: linear
+ lr_scheduler:
+ type: CosineAnnealingLR
+ args:
+ T_max: *max_epochs
+ interval: epoch
+ swa_args:
+ start: 2
+ lr: 5.e-2
+ callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, EarlyStopping]
+ callback_args:
+ Checkpoint:
+ monitor: val_accuracy
+ ProgressBar:
+ epochs: null
+ log_batch_frequency: 100
+ EarlyStopping:
+ monitor: val_loss
+ min_delta: 0.0
+ patience: 5
+ mode: min
+ WandbCallback:
+ log_batch_frequency: 10
+ WandbImageLogger:
+ num_examples: 4
+ use_transpose: true
+ verbosity: 0 # 0, 1, 2
+ resume_experiment: null
+ train: true
+ test: true
+ test_metric: test_accuracy
diff --git a/training/gpu_manager.py b/training/gpu_manager.py
new file mode 100644
index 0000000..ce1b3dd
--- /dev/null
+++ b/training/gpu_manager.py
@@ -0,0 +1,62 @@
+"""GPUManager class."""
+import os
+import time
+from typing import Optional
+
+import gpustat
+from loguru import logger
+import numpy as np
+from redlock import Redlock
+
+
+GPU_LOCK_TIMEOUT = 5000 # ms
+
+
+class GPUManager:
+ """Class for allocating GPUs."""
+
+ def __init__(self, verbose: bool = False) -> None:
+ """Initializes Redlock manager."""
+ self.lock_manager = Redlock([{"host": "localhost", "port": 6379, "db": 0}])
+ self.verbose = verbose
+
+ def get_free_gpu(self) -> int:
+ """Gets a free GPU.
+
+ If some GPUs are available, try reserving one by checking out an exclusive redis lock.
+ If none available or can not get lock, sleep and check again.
+
+ Returns:
+ int: The gpu index.
+
+ """
+ while True:
+ gpu_index = self._get_free_gpu()
+ if gpu_index is not None:
+ return gpu_index
+
+ if self.verbose:
+ logger.debug(f"pid {os.getpid()} sleeping")
+ time.sleep(GPU_LOCK_TIMEOUT / 1000)
+
+ def _get_free_gpu(self) -> Optional[int]:
+ """Fetches an available GPU index."""
+ try:
+ available_gpu_indices = [
+ gpu.index
+ for gpu in gpustat.GPUStatCollection.new_query()
+ if gpu.memory_used < 0.5 * gpu.memory_total
+ ]
+ except Exception as e:
+ logger.debug(f"Got the following exception: {e}")
+ return None
+
+ if available_gpu_indices:
+ gpu_index = np.random.choice(available_gpu_indices)
+ if self.verbose:
+ logger.debug(f"pid {os.getpid()} picking gpu {gpu_index}")
+ if self.lock_manager.lock(f"gpu_{gpu_index}", GPU_LOCK_TIMEOUT):
+ return int(gpu_index)
+ if self.verbose:
+ logger.debug(f"pid {os.getpid()} could not get lock.")
+ return None
diff --git a/training/prepare_experiments.py b/training/prepare_experiments.py
new file mode 100644
index 0000000..21997af
--- /dev/null
+++ b/training/prepare_experiments.py
@@ -0,0 +1,34 @@
+"""Run a experiment from a config file."""
+import json
+
+import click
+import yaml
+
+
+def run_experiments(experiments_filename: str) -> None:
+ """Run experiment from file."""
+ with open(experiments_filename, "r") as f:
+ experiments_config = yaml.safe_load(f)
+
+ num_experiments = len(experiments_config["experiments"])
+ for index in range(num_experiments):
+ experiment_config = experiments_config["experiments"][index]
+ experiment_config["experiment_group"] = experiments_config["experiment_group"]
+ cmd = f"poetry run run-experiment --gpu=-1 --save '{json.dumps(experiment_config)}'"
+ print(cmd)
+
+
+@click.command()
+@click.option(
+ "--experiments_filename",
+ required=True,
+ type=str,
+ help="Filename of Yaml file of experiments to run.",
+)
+def run_cli(experiments_filename: str) -> None:
+ """Parse command-line arguments and run experiments from provided file."""
+ run_experiments(experiments_filename)
+
+
+if __name__ == "__main__":
+ run_cli()
diff --git a/training/run_experiment.py b/training/run_experiment.py
new file mode 100644
index 0000000..faafea6
--- /dev/null
+++ b/training/run_experiment.py
@@ -0,0 +1,382 @@
+"""Script to run experiments."""
+from datetime import datetime
+from glob import glob
+import importlib
+import json
+import os
+from pathlib import Path
+import re
+from typing import Callable, Dict, List, Optional, Tuple, Type
+import warnings
+
+import click
+from loguru import logger
+import numpy as np
+import torch
+from torchsummary import summary
+from tqdm import tqdm
+from training.gpu_manager import GPUManager
+from training.trainer.callbacks import CallbackList
+from training.trainer.train import Trainer
+import wandb
+import yaml
+
+import text_recognizer.models
+from text_recognizer.models import Model
+import text_recognizer.networks
+from text_recognizer.networks.loss import loss as custom_loss_module
+
+EXPERIMENTS_DIRNAME = Path(__file__).parents[0].resolve() / "experiments"
+
+
+def _get_level(verbose: int) -> int:
+ """Sets the logger level."""
+ levels = {0: 40, 1: 20, 2: 10}
+ verbose = verbose if verbose <= 2 else 2
+ return levels[verbose]
+
+
+def _create_experiment_dir(
+ experiment_config: Dict, checkpoint: Optional[str] = None
+) -> Path:
+ """Create new experiment."""
+ EXPERIMENTS_DIRNAME.mkdir(parents=True, exist_ok=True)
+ experiment_dir = EXPERIMENTS_DIRNAME / (
+ f"{experiment_config['model']}_"
+ + f"{experiment_config['dataset']['type']}_"
+ + f"{experiment_config['network']['type']}"
+ )
+
+ if checkpoint is None:
+ experiment = datetime.now().strftime("%m%d_%H%M%S")
+ logger.debug(f"Creating a new experiment called {experiment}")
+ else:
+ available_experiments = glob(str(experiment_dir) + "/*")
+ available_experiments.sort()
+ if checkpoint == "last":
+ experiment = available_experiments[-1]
+ logger.debug(f"Resuming the latest experiment {experiment}")
+ else:
+ experiment = checkpoint
+ if not str(experiment_dir / experiment) in available_experiments:
+ raise FileNotFoundError("Experiment does not exist.")
+ logger.debug(f"Resuming the from experiment {checkpoint}")
+
+ experiment_dir = experiment_dir / experiment
+
+ # Create log and model directories.
+ log_dir = experiment_dir / "log"
+ model_dir = experiment_dir / "model"
+
+ return experiment_dir, log_dir, model_dir
+
+
+def _load_modules_and_arguments(experiment_config: Dict,) -> Tuple[Callable, Dict]:
+ """Loads all modules and arguments."""
+ # Load the dataset module.
+ dataset_args = experiment_config.get("dataset", {})
+ dataset_ = dataset_args["type"]
+
+ # Import the model module and model arguments.
+ model_class_ = getattr(text_recognizer.models, experiment_config["model"])
+
+ # Import metrics.
+ metric_fns_ = (
+ {
+ metric: getattr(text_recognizer.networks, metric)
+ for metric in experiment_config["metrics"]
+ }
+ if experiment_config["metrics"] is not None
+ else None
+ )
+
+ # Import network module and arguments.
+ network_fn_ = experiment_config["network"]["type"]
+ network_args = experiment_config["network"].get("args", {})
+
+ # Criterion
+ if experiment_config["criterion"]["type"] in custom_loss_module.__all__:
+ criterion_ = getattr(custom_loss_module, experiment_config["criterion"]["type"])
+ else:
+ criterion_ = getattr(torch.nn, experiment_config["criterion"]["type"])
+ criterion_args = experiment_config["criterion"].get("args", {}) or {}
+
+ # Optimizers
+ optimizer_ = getattr(torch.optim, experiment_config["optimizer"]["type"])
+ optimizer_args = experiment_config["optimizer"].get("args", {})
+
+ # Learning rate scheduler
+ lr_scheduler_ = None
+ lr_scheduler_args = None
+ if "lr_scheduler" in experiment_config:
+ lr_scheduler_ = getattr(
+ torch.optim.lr_scheduler, experiment_config["lr_scheduler"]["type"]
+ )
+ lr_scheduler_args = experiment_config["lr_scheduler"].get("args", {}) or {}
+
+ # SWA scheduler.
+ if "swa_args" in experiment_config:
+ swa_args = experiment_config.get("swa_args", {}) or {}
+ else:
+ swa_args = None
+
+ model_args = {
+ "dataset": dataset_,
+ "dataset_args": dataset_args,
+ "metrics": metric_fns_,
+ "network_fn": network_fn_,
+ "network_args": network_args,
+ "criterion": criterion_,
+ "criterion_args": criterion_args,
+ "optimizer": optimizer_,
+ "optimizer_args": optimizer_args,
+ "lr_scheduler": lr_scheduler_,
+ "lr_scheduler_args": lr_scheduler_args,
+ "swa_args": swa_args,
+ }
+
+ return model_class_, model_args
+
+
+def _configure_callbacks(experiment_config: Dict, model_dir: Path) -> CallbackList:
+ """Configure a callback list for trainer."""
+ if "Checkpoint" in experiment_config["callback_args"]:
+ experiment_config["callback_args"]["Checkpoint"]["checkpoint_path"] = str(
+ model_dir
+ )
+
+ # Initializes callbacks.
+ callback_modules = importlib.import_module("training.trainer.callbacks")
+ callbacks = []
+ for callback in experiment_config["callbacks"]:
+ args = experiment_config["callback_args"][callback] or {}
+ callbacks.append(getattr(callback_modules, callback)(**args))
+
+ return callbacks
+
+
+def _configure_logger(log_dir: Path, verbose: int = 0) -> None:
+ """Configure the loguru logger for output to terminal and disk."""
+ # Have to remove default logger to get tqdm to work properly.
+ logger.remove()
+
+ # Fetch verbosity level.
+ level = _get_level(verbose)
+
+ logger.add(lambda msg: tqdm.write(msg, end=""), colorize=True, level=level)
+ logger.add(
+ str(log_dir / "train.log"),
+ format="{time:YYYY-MM-DD at HH:mm:ss} | {level} | {message}",
+ )
+
+
+def _save_config(experiment_dir: Path, experiment_config: Dict) -> None:
+ """Copy config to experiment directory."""
+ config_path = experiment_dir / "config.yml"
+ with open(str(config_path), "w") as f:
+ yaml.dump(experiment_config, f)
+
+
+def _load_from_checkpoint(
+ model: Type[Model], model_dir: Path, pretrained_weights: str = None,
+) -> None:
+ """If checkpoint exists, load model weights and optimizers from checkpoint."""
+ # Get checkpoint path.
+ if pretrained_weights is not None:
+ logger.info(f"Loading weights from {pretrained_weights}.")
+ checkpoint_path = (
+ EXPERIMENTS_DIRNAME / Path(pretrained_weights) / "model" / "best.pt"
+ )
+ else:
+ logger.info(f"Loading weights from {model_dir}.")
+ checkpoint_path = model_dir / "last.pt"
+ if checkpoint_path.exists():
+ logger.info("Loading and resuming training from checkpoint.")
+ model.load_from_checkpoint(checkpoint_path)
+
+
+def evaluate_embedding(model: Type[Model]) -> Dict:
+ """Evaluates the embedding space."""
+ from pytorch_metric_learning import testers
+ from pytorch_metric_learning.utils.accuracy_calculator import AccuracyCalculator
+
+ accuracy_calculator = AccuracyCalculator(
+ include=("mean_average_precision_at_r",), k=10
+ )
+
+ def get_all_embeddings(model: Type[Model]) -> Tuple:
+ tester = testers.BaseTester()
+ return tester.get_all_embeddings(model.test_dataset, model.network)
+
+ embeddings, labels = get_all_embeddings(model)
+ logger.info("Computing embedding accuracy")
+ accuracies = accuracy_calculator.get_accuracy(
+ embeddings, embeddings, np.squeeze(labels), np.squeeze(labels), True
+ )
+ logger.info(
+ f"Test set accuracy (MAP@10) = {accuracies['mean_average_precision_at_r']}"
+ )
+ return accuracies
+
+
+def run_experiment(
+ experiment_config: Dict,
+ save_weights: bool,
+ device: str,
+ use_wandb: bool,
+ train: bool,
+ test: bool,
+ verbose: int = 0,
+ checkpoint: Optional[str] = None,
+ pretrained_weights: Optional[str] = None,
+) -> None:
+ """Runs an experiment."""
+ logger.info(f"Experiment config: {json.dumps(experiment_config)}")
+
+ # Create new experiment.
+ experiment_dir, log_dir, model_dir = _create_experiment_dir(
+ experiment_config, checkpoint
+ )
+
+ # Make sure the log/model directory exists.
+ log_dir.mkdir(parents=True, exist_ok=True)
+ model_dir.mkdir(parents=True, exist_ok=True)
+
+ # Load the modules and model arguments.
+ model_class_, model_args = _load_modules_and_arguments(experiment_config)
+
+ # Initializes the model with experiment config.
+ model = model_class_(**model_args, device=device)
+
+ callbacks = _configure_callbacks(experiment_config, model_dir)
+
+ # Setup logger.
+ _configure_logger(log_dir, verbose)
+
+ # Load from checkpoint if resuming an experiment.
+ resume = False
+ if checkpoint is not None or pretrained_weights is not None:
+ # resume = True
+ _load_from_checkpoint(model, model_dir, pretrained_weights)
+
+ logger.info(f"The class mapping is {model.mapping}")
+
+ # Initializes Weights & Biases
+ if use_wandb:
+ wandb.init(project="text-recognizer", config=experiment_config, resume=resume)
+
+ # Lets W&B save the model and track the gradients and optional parameters.
+ wandb.watch(model.network)
+
+ experiment_config["experiment_group"] = experiment_config.get(
+ "experiment_group", None
+ )
+
+ experiment_config["device"] = device
+
+ # Save the config used in the experiment folder.
+ _save_config(experiment_dir, experiment_config)
+
+ # Prints a summary of the network in terminal.
+ model.summary(experiment_config["train_args"]["input_shape"])
+
+ # Load trainer.
+ trainer = Trainer(
+ max_epochs=experiment_config["train_args"]["max_epochs"],
+ callbacks=callbacks,
+ transformer_model=experiment_config["train_args"]["transformer_model"],
+ max_norm=experiment_config["train_args"]["max_norm"],
+ freeze_backbone=experiment_config["train_args"]["freeze_backbone"],
+ )
+
+ # Train the model.
+ if train:
+ trainer.fit(model)
+
+ # Run inference over test set.
+ if test:
+ logger.info("Loading checkpoint with the best weights.")
+ if "checkpoint" in experiment_config["train_args"]:
+ model.load_from_checkpoint(
+ model_dir / experiment_config["train_args"]["checkpoint"]
+ )
+ else:
+ model.load_from_checkpoint(model_dir / "best.pt")
+
+ logger.info("Running inference on test set.")
+ if experiment_config["criterion"]["type"] == "EmbeddingLoss":
+ logger.info("Evaluating embedding.")
+ score = evaluate_embedding(model)
+ else:
+ score = trainer.test(model)
+
+ logger.info(f"Test set evaluation: {score}")
+
+ if use_wandb:
+ wandb.log(
+ {
+ experiment_config["test_metric"]: score[
+ experiment_config["test_metric"]
+ ]
+ }
+ )
+
+ if save_weights:
+ model.save_weights(model_dir)
+
+
+@click.command()
+@click.argument("experiment_config",)
+@click.option("--gpu", type=int, default=0, help="Provide the index of the GPU to use.")
+@click.option(
+ "--save",
+ is_flag=True,
+ help="If set, the final weights will be saved to a canonical, version-controlled location.",
+)
+@click.option(
+ "--nowandb", is_flag=False, help="If true, do not use wandb for this run."
+)
+@click.option("--test", is_flag=True, help="If true, test the model.")
+@click.option("-v", "--verbose", count=True)
+@click.option("--checkpoint", type=str, help="Path to the experiment.")
+@click.option(
+ "--pretrained_weights", type=str, help="Path to pretrained model weights."
+)
+@click.option(
+ "--notrain", is_flag=False, help="Do not train the model.",
+)
+def run_cli(
+ experiment_config: str,
+ gpu: int,
+ save: bool,
+ nowandb: bool,
+ notrain: bool,
+ test: bool,
+ verbose: int,
+ checkpoint: Optional[str] = None,
+ pretrained_weights: Optional[str] = None,
+) -> None:
+ """Run experiment."""
+ if gpu < 0:
+ gpu_manager = GPUManager(True)
+ gpu = gpu_manager.get_free_gpu()
+ device = "cuda:" + str(gpu)
+
+ experiment_config = json.loads(experiment_config)
+ os.environ["CUDA_VISIBLE_DEVICES"] = f"{gpu}"
+
+ run_experiment(
+ experiment_config,
+ save,
+ device,
+ use_wandb=not nowandb,
+ train=not notrain,
+ test=test,
+ verbose=verbose,
+ checkpoint=checkpoint,
+ pretrained_weights=pretrained_weights,
+ )
+
+
+if __name__ == "__main__":
+ run_cli()
diff --git a/training/run_sweep.py b/training/run_sweep.py
new file mode 100644
index 0000000..a578592
--- /dev/null
+++ b/training/run_sweep.py
@@ -0,0 +1,92 @@
+"""W&B Sweep Functionality."""
+from ast import literal_eval
+import json
+import os
+from pathlib import Path
+import signal
+import subprocess # nosec
+import sys
+from typing import Dict, List, Tuple
+
+import click
+import yaml
+
+EXPERIMENTS_DIRNAME = Path(__file__).parents[0].resolve() / "experiments"
+
+
+def load_config() -> Dict:
+ """Load base hyperparameter config."""
+ with open(str(EXPERIMENTS_DIRNAME / "default_config_emnist.yml"), "r") as f:
+ default_config = yaml.safe_load(f)
+ return default_config
+
+
+def args_to_json(
+ default_config: dict, preserve_args: tuple = ("gpu", "save")
+) -> Tuple[dict, list]:
+ """Convert command line arguments to nested config values.
+
+ i.e. run_sweep.py --dataset_args.foo=1.7
+ {
+ "dataset_args": {
+ "foo": 1.7
+ }
+ }
+
+ Args:
+ default_config (dict): The base config used for every experiment.
+ preserve_args (tuple): Arguments preserved for all runs. Defaults to ("gpu", "save").
+
+ Returns:
+ Tuple[dict, list]: Tuple of config dictionary and list of arguments.
+
+ """
+
+ args = []
+ config = default_config.copy()
+ key, val = None, None
+ for arg in sys.argv[1:]:
+ if "=" in arg:
+ key, val = arg.split("=")
+ elif key:
+ val = arg
+ else:
+ key = arg
+ if key and val:
+ parsed_key = key.lstrip("-").split(".")
+ if parsed_key[0] in preserve_args:
+ args.append("--{}={}".format(parsed_key[0], val))
+ else:
+ nested = config
+ for level in parsed_key[:-1]:
+ nested[level] = config.get(level, {})
+ nested = nested[level]
+ try:
+ # Convert numerics to floats / ints
+ val = literal_eval(val)
+ except ValueError:
+ pass
+ nested[parsed_key[-1]] = val
+ key, val = None, None
+ return config, args
+
+
+def main() -> None:
+ """Runs a W&B sweep."""
+ default_config = load_config()
+ config, args = args_to_json(default_config)
+ env = {
+ k: v for k, v in os.environ.items() if k not in ("WANDB_PROGRAM", "WANDB_ARGS")
+ }
+ # pylint: disable=subprocess-popen-preexec-fn
+ run = subprocess.Popen(
+ ["python", "training/run_experiment.py", *args, json.dumps(config)],
+ env=env,
+ preexec_fn=os.setsid,
+ ) # nosec
+ signal.signal(signal.SIGTERM, lambda *args: run.terminate())
+ run.wait()
+
+
+if __name__ == "__main__":
+ main()
diff --git a/training/sweep_emnist.yml b/training/sweep_emnist.yml
new file mode 100644
index 0000000..48d7261
--- /dev/null
+++ b/training/sweep_emnist.yml
@@ -0,0 +1,26 @@
+program: training/run_sweep.py
+method: bayes
+metric:
+ name: val_loss
+ goal: minimize
+parameters:
+ dataset:
+ value: EmnistDataset
+ model:
+ value: CharacterModel
+ network:
+ value: MLP
+ network_args.hidden_size:
+ values: [128, 256]
+ network_args.dropout_rate:
+ values: [0.2, 0.4]
+ network_args.num_layers:
+ values: [3, 6]
+ optimizer_args.lr:
+ values: [1.e-1, 1.e-5]
+ lr_scheduler_args.max_lr:
+ values: [1.0e-1, 1.0e-5]
+ train_args.batch_size:
+ values: [64, 128]
+ train_args.epochs:
+ value: 5
diff --git a/training/sweep_emnist_resnet.yml b/training/sweep_emnist_resnet.yml
new file mode 100644
index 0000000..19a3040
--- /dev/null
+++ b/training/sweep_emnist_resnet.yml
@@ -0,0 +1,50 @@
+program: training/run_sweep.py
+method: bayes
+metric:
+ name: val_accuracy
+ goal: maximize
+parameters:
+ dataset:
+ value: EmnistDataset
+ model:
+ value: CharacterModel
+ network:
+ value: ResidualNetwork
+ network_args.block_sizes:
+ distribution: q_uniform
+ min: 16
+ max: 256
+ q: 8
+ network_args.depths:
+ distribution: int_uniform
+ min: 1
+ max: 3
+ network_args.levels:
+ distribution: int_uniform
+ min: 1
+ max: 2
+ network_args.activation:
+ distribution: categorical
+ values:
+ - gelu
+ - leaky_relu
+ - relu
+ - selu
+ optimizer_args.lr:
+ distribution: uniform
+ min: 1.e-5
+ max: 1.e-1
+ lr_scheduler_args.max_lr:
+ distribution: uniform
+ min: 1.e-5
+ max: 1.e-1
+ train_args.batch_size:
+ distribution: q_uniform
+ min: 32
+ max: 256
+ q: 8
+ train_args.epochs:
+ value: 5
+early_terminate:
+ type: hyperband
+ min_iter: 2
diff --git a/training/trainer/__init__.py b/training/trainer/__init__.py
new file mode 100644
index 0000000..de41bfb
--- /dev/null
+++ b/training/trainer/__init__.py
@@ -0,0 +1,2 @@
+"""Trainer modules."""
+from .train import Trainer
diff --git a/training/trainer/callbacks/__init__.py b/training/trainer/callbacks/__init__.py
new file mode 100644
index 0000000..80c4177
--- /dev/null
+++ b/training/trainer/callbacks/__init__.py
@@ -0,0 +1,29 @@
+"""The callback modules used in the training script."""
+from .base import Callback, CallbackList
+from .checkpoint import Checkpoint
+from .early_stopping import EarlyStopping
+from .lr_schedulers import (
+ LRScheduler,
+ SWA,
+)
+from .progress_bar import ProgressBar
+from .wandb_callbacks import (
+ WandbCallback,
+ WandbImageLogger,
+ WandbReconstructionLogger,
+ WandbSegmentationLogger,
+)
+
+__all__ = [
+ "Callback",
+ "CallbackList",
+ "Checkpoint",
+ "EarlyStopping",
+ "LRScheduler",
+ "WandbCallback",
+ "WandbImageLogger",
+ "WandbReconstructionLogger",
+ "WandbSegmentationLogger",
+ "ProgressBar",
+ "SWA",
+]
diff --git a/training/trainer/callbacks/base.py b/training/trainer/callbacks/base.py
new file mode 100644
index 0000000..500b642
--- /dev/null
+++ b/training/trainer/callbacks/base.py
@@ -0,0 +1,188 @@
+"""Metaclass for callback functions."""
+
+from enum import Enum
+from typing import Callable, Dict, List, Optional, Type, Union
+
+from loguru import logger
+import numpy as np
+import torch
+
+from text_recognizer.models import Model
+
+
+class ModeKeys:
+ """Mode keys for CallbackList."""
+
+ TRAIN = "train"
+ VALIDATION = "validation"
+
+
+class Callback:
+ """Metaclass for callbacks used in training."""
+
+ def __init__(self) -> None:
+ """Initializes the Callback instance."""
+ self.model = None
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Set the model."""
+ self.model = model
+
+ def on_fit_begin(self) -> None:
+ """Called when fit begins."""
+ pass
+
+ def on_fit_end(self) -> None:
+ """Called when fit ends."""
+ pass
+
+ def on_epoch_begin(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the beginning of an epoch. Only used in training mode."""
+ pass
+
+ def on_epoch_end(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch. Only used in training mode."""
+ pass
+
+ def on_train_batch_begin(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the beginning of an epoch."""
+ pass
+
+ def on_train_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch."""
+ pass
+
+ def on_validation_batch_begin(
+ self, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Called at the beginning of an epoch."""
+ pass
+
+ def on_validation_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch."""
+ pass
+
+ def on_test_begin(self) -> None:
+ """Called at the beginning of test."""
+ pass
+
+ def on_test_end(self) -> None:
+ """Called at the end of test."""
+ pass
+
+
+class CallbackList:
+ """Container for abstracting away callback calls."""
+
+ mode_keys = ModeKeys()
+
+ def __init__(self, model: Type[Model], callbacks: List[Callback] = None) -> None:
+ """Container for `Callback` instances.
+
+ This object wraps a list of `Callback` instances and allows them all to be
+ called via a single end point.
+
+ Args:
+ model (Type[Model]): A `Model` instance.
+ callbacks (List[Callback]): List of `Callback` instances. Defaults to None.
+
+ """
+
+ self._callbacks = callbacks or []
+ if model:
+ self.set_model(model)
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Set the model for all callbacks."""
+ self.model = model
+ for callback in self._callbacks:
+ callback.set_model(model=self.model)
+
+ def append(self, callback: Type[Callback]) -> None:
+ """Append new callback to callback list."""
+ self._callbacks.append(callback)
+
+ def on_fit_begin(self) -> None:
+ """Called when fit begins."""
+ for callback in self._callbacks:
+ callback.on_fit_begin()
+
+ def on_fit_end(self) -> None:
+ """Called when fit ends."""
+ for callback in self._callbacks:
+ callback.on_fit_end()
+
+ def on_test_begin(self) -> None:
+ """Called when test begins."""
+ for callback in self._callbacks:
+ callback.on_test_begin()
+
+ def on_test_end(self) -> None:
+ """Called when test ends."""
+ for callback in self._callbacks:
+ callback.on_test_end()
+
+ def on_epoch_begin(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the beginning of an epoch."""
+ for callback in self._callbacks:
+ callback.on_epoch_begin(epoch, logs)
+
+ def on_epoch_end(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch."""
+ for callback in self._callbacks:
+ callback.on_epoch_end(epoch, logs)
+
+ def _call_batch_hook(
+ self, mode: str, hook: str, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Helper function for all batch_{begin | end} methods."""
+ if hook == "begin":
+ self._call_batch_begin_hook(mode, batch, logs)
+ elif hook == "end":
+ self._call_batch_end_hook(mode, batch, logs)
+ else:
+ raise ValueError(f"Unrecognized hook {hook}.")
+
+ def _call_batch_begin_hook(
+ self, mode: str, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Helper function for all `on_*_batch_begin` methods."""
+ hook_name = f"on_{mode}_batch_begin"
+ self._call_batch_hook_helper(hook_name, batch, logs)
+
+ def _call_batch_end_hook(
+ self, mode: str, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Helper function for all `on_*_batch_end` methods."""
+ hook_name = f"on_{mode}_batch_end"
+ self._call_batch_hook_helper(hook_name, batch, logs)
+
+ def _call_batch_hook_helper(
+ self, hook_name: str, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Helper function for `on_*_batch_begin` methods."""
+ for callback in self._callbacks:
+ hook = getattr(callback, hook_name)
+ hook(batch, logs)
+
+ def on_train_batch_begin(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the beginning of an epoch."""
+ self._call_batch_hook(self.mode_keys.TRAIN, "begin", batch, logs)
+
+ def on_train_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch."""
+ self._call_batch_hook(self.mode_keys.TRAIN, "end", batch, logs)
+
+ def on_validation_batch_begin(
+ self, batch: int, logs: Optional[Dict] = None
+ ) -> None:
+ """Called at the beginning of an epoch."""
+ self._call_batch_hook(self.mode_keys.VALIDATION, "begin", batch, logs)
+
+ def on_validation_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Called at the end of an epoch."""
+ self._call_batch_hook(self.mode_keys.VALIDATION, "end", batch, logs)
+
+ def __iter__(self) -> iter:
+ """Iter function for callback list."""
+ return iter(self._callbacks)
diff --git a/training/trainer/callbacks/checkpoint.py b/training/trainer/callbacks/checkpoint.py
new file mode 100644
index 0000000..a54e0a9
--- /dev/null
+++ b/training/trainer/callbacks/checkpoint.py
@@ -0,0 +1,95 @@
+"""Callback checkpoint for training models."""
+from enum import Enum
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Type, Union
+
+from loguru import logger
+import numpy as np
+import torch
+from training.trainer.callbacks import Callback
+
+from text_recognizer.models import Model
+
+
+class Checkpoint(Callback):
+ """Saving model parameters at the end of each epoch."""
+
+ mode_dict = {
+ "min": torch.lt,
+ "max": torch.gt,
+ }
+
+ def __init__(
+ self,
+ checkpoint_path: Union[str, Path],
+ monitor: str = "accuracy",
+ mode: str = "auto",
+ min_delta: float = 0.0,
+ ) -> None:
+ """Monitors a quantity that will allow us to determine the best model weights.
+
+ Args:
+ checkpoint_path (Union[str, Path]): Path to the experiment with the checkpoint.
+ monitor (str): Name of the quantity to monitor. Defaults to "accuracy".
+ mode (str): Description of parameter `mode`. Defaults to "auto".
+ min_delta (float): Description of parameter `min_delta`. Defaults to 0.0.
+
+ """
+ super().__init__()
+ self.checkpoint_path = Path(checkpoint_path)
+ self.monitor = monitor
+ self.mode = mode
+ self.min_delta = torch.tensor(min_delta)
+
+ if mode not in ["auto", "min", "max"]:
+ logger.warning(f"Checkpoint mode {mode} is unkown, fallback to auto mode.")
+
+ self.mode = "auto"
+
+ if self.mode == "auto":
+ if "accuracy" in self.monitor:
+ self.mode = "max"
+ else:
+ self.mode = "min"
+ logger.debug(
+ f"Checkpoint mode set to {self.mode} for monitoring {self.monitor}."
+ )
+
+ torch_inf = torch.tensor(np.inf)
+ self.min_delta *= 1 if self.monitor_op == torch.gt else -1
+ self.best_score = torch_inf if self.monitor_op == torch.lt else -torch_inf
+
+ @property
+ def monitor_op(self) -> float:
+ """Returns the comparison method."""
+ return self.mode_dict[self.mode]
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Saves a checkpoint for the network parameters.
+
+ Args:
+ epoch (int): The current epoch.
+ logs (Dict): The log containing the monitored metrics.
+
+ """
+ current = self.get_monitor_value(logs)
+ if current is None:
+ return
+ if self.monitor_op(current - self.min_delta, self.best_score):
+ self.best_score = current
+ is_best = True
+ else:
+ is_best = False
+
+ self.model.save_checkpoint(self.checkpoint_path, is_best, epoch, self.monitor)
+
+ def get_monitor_value(self, logs: Dict) -> Union[float, None]:
+ """Extracts the monitored value."""
+ monitor_value = logs.get(self.monitor)
+ if monitor_value is None:
+ logger.warning(
+ f"Checkpoint is conditioned on metric {self.monitor} which is not available. Available"
+ + f" metrics are: {','.join(list(logs.keys()))}"
+ )
+ return None
+ return monitor_value
diff --git a/training/trainer/callbacks/early_stopping.py b/training/trainer/callbacks/early_stopping.py
new file mode 100644
index 0000000..02b431f
--- /dev/null
+++ b/training/trainer/callbacks/early_stopping.py
@@ -0,0 +1,108 @@
+"""Implements Early stopping for PyTorch model."""
+from typing import Dict, Union
+
+from loguru import logger
+import numpy as np
+import torch
+from torch import Tensor
+from training.trainer.callbacks import Callback
+
+
+class EarlyStopping(Callback):
+ """Stops training when a monitored metric stops improving."""
+
+ mode_dict = {
+ "min": torch.lt,
+ "max": torch.gt,
+ }
+
+ def __init__(
+ self,
+ monitor: str = "val_loss",
+ min_delta: float = 0.0,
+ patience: int = 3,
+ mode: str = "auto",
+ ) -> None:
+ """Initializes the EarlyStopping callback.
+
+ Args:
+ monitor (str): Description of parameter `monitor`. Defaults to "val_loss".
+ min_delta (float): Description of parameter `min_delta`. Defaults to 0.0.
+ patience (int): Description of parameter `patience`. Defaults to 3.
+ mode (str): Description of parameter `mode`. Defaults to "auto".
+
+ """
+ super().__init__()
+ self.monitor = monitor
+ self.patience = patience
+ self.min_delta = torch.tensor(min_delta)
+ self.mode = mode
+ self.wait_count = 0
+ self.stopped_epoch = 0
+
+ if mode not in ["auto", "min", "max"]:
+ logger.warning(
+ f"EarlyStopping mode {mode} is unkown, fallback to auto mode."
+ )
+
+ self.mode = "auto"
+
+ if self.mode == "auto":
+ if "accuracy" in self.monitor:
+ self.mode = "max"
+ else:
+ self.mode = "min"
+ logger.debug(
+ f"EarlyStopping mode set to {self.mode} for monitoring {self.monitor}."
+ )
+
+ self.torch_inf = torch.tensor(np.inf)
+ self.min_delta *= 1 if self.monitor_op == torch.gt else -1
+ self.best_score = (
+ self.torch_inf if self.monitor_op == torch.lt else -self.torch_inf
+ )
+
+ @property
+ def monitor_op(self) -> float:
+ """Returns the comparison method."""
+ return self.mode_dict[self.mode]
+
+ def on_fit_begin(self) -> Union[torch.lt, torch.gt]:
+ """Reset the early stopping variables for reuse."""
+ self.wait_count = 0
+ self.stopped_epoch = 0
+ self.best_score = (
+ self.torch_inf if self.monitor_op == torch.lt else -self.torch_inf
+ )
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Computes the early stop criterion."""
+ current = self.get_monitor_value(logs)
+ if current is None:
+ return
+ if self.monitor_op(current - self.min_delta, self.best_score):
+ self.best_score = current
+ self.wait_count = 0
+ else:
+ self.wait_count += 1
+ if self.wait_count >= self.patience:
+ self.stopped_epoch = epoch
+ self.model.stop_training = True
+
+ def on_fit_end(self) -> None:
+ """Logs if early stopping was used."""
+ if self.stopped_epoch > 0:
+ logger.info(
+ f"Stopped training at epoch {self.stopped_epoch + 1} with early stopping."
+ )
+
+ def get_monitor_value(self, logs: Dict) -> Union[Tensor, None]:
+ """Extracts the monitor value."""
+ monitor_value = logs.get(self.monitor)
+ if monitor_value is None:
+ logger.warning(
+ f"Early stopping is conditioned on metric {self.monitor} which is not available. Available"
+ + f"metrics are: {','.join(list(logs.keys()))}"
+ )
+ return None
+ return torch.tensor(monitor_value)
diff --git a/training/trainer/callbacks/lr_schedulers.py b/training/trainer/callbacks/lr_schedulers.py
new file mode 100644
index 0000000..630c434
--- /dev/null
+++ b/training/trainer/callbacks/lr_schedulers.py
@@ -0,0 +1,77 @@
+"""Callbacks for learning rate schedulers."""
+from typing import Callable, Dict, List, Optional, Type
+
+from torch.optim.swa_utils import update_bn
+from training.trainer.callbacks import Callback
+
+from text_recognizer.models import Model
+
+
+class LRScheduler(Callback):
+ """Generic learning rate scheduler callback."""
+
+ def __init__(self) -> None:
+ super().__init__()
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Sets the model and lr scheduler."""
+ self.model = model
+ self.lr_scheduler = self.model.lr_scheduler["lr_scheduler"]
+ self.interval = self.model.lr_scheduler["interval"]
+
+ def on_epoch_end(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Takes a step at the end of every epoch."""
+ if self.interval == "epoch":
+ if "ReduceLROnPlateau" in self.lr_scheduler.__class__.__name__:
+ self.lr_scheduler.step(logs["val_loss"])
+ else:
+ self.lr_scheduler.step()
+
+ def on_train_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Takes a step at the end of every training batch."""
+ if self.interval == "step":
+ self.lr_scheduler.step()
+
+
+class SWA(Callback):
+ """Stochastic Weight Averaging callback."""
+
+ def __init__(self) -> None:
+ """Initializes the callback."""
+ super().__init__()
+ self.lr_scheduler = None
+ self.interval = None
+ self.swa_scheduler = None
+ self.swa_start = None
+ self.current_epoch = 1
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Sets the model and lr scheduler."""
+ self.model = model
+ self.lr_scheduler = self.model.lr_scheduler["lr_scheduler"]
+ self.interval = self.model.lr_scheduler["interval"]
+ self.swa_scheduler = self.model.swa_scheduler["swa_scheduler"]
+ self.swa_start = self.model.swa_scheduler["swa_start"]
+
+ def on_epoch_end(self, epoch: int, logs: Optional[Dict] = None) -> None:
+ """Takes a step at the end of every training batch."""
+ if epoch > self.swa_start:
+ self.model.swa_network.update_parameters(self.model.network)
+ self.swa_scheduler.step()
+ elif self.interval == "epoch":
+ self.lr_scheduler.step()
+ self.current_epoch = epoch
+
+ def on_train_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Takes a step at the end of every training batch."""
+ if self.current_epoch < self.swa_start and self.interval == "step":
+ self.lr_scheduler.step()
+
+ def on_fit_end(self) -> None:
+ """Update batch norm statistics for the swa model at the end of training."""
+ if self.model.swa_network:
+ update_bn(
+ self.model.val_dataloader(),
+ self.model.swa_network,
+ device=self.model.device,
+ )
diff --git a/training/trainer/callbacks/progress_bar.py b/training/trainer/callbacks/progress_bar.py
new file mode 100644
index 0000000..6c4305a
--- /dev/null
+++ b/training/trainer/callbacks/progress_bar.py
@@ -0,0 +1,65 @@
+"""Progress bar callback for the training loop."""
+from typing import Dict, Optional
+
+from tqdm import tqdm
+from training.trainer.callbacks import Callback
+
+
+class ProgressBar(Callback):
+ """A TQDM progress bar for the training loop."""
+
+ def __init__(self, epochs: int, log_batch_frequency: int = None) -> None:
+ """Initializes the tqdm callback."""
+ self.epochs = epochs
+ print(epochs, type(epochs))
+ self.log_batch_frequency = log_batch_frequency
+ self.progress_bar = None
+ self.val_metrics = {}
+
+ def _configure_progress_bar(self) -> None:
+ """Configures the tqdm progress bar with custom bar format."""
+ self.progress_bar = tqdm(
+ total=len(self.model.train_dataloader()),
+ leave=False,
+ unit="steps",
+ mininterval=self.log_batch_frequency,
+ bar_format="{desc} |{bar:32}| {n_fmt}/{total_fmt} ETA: {remaining} {rate_fmt}{postfix}",
+ )
+
+ def _key_abbreviations(self, logs: Dict) -> Dict:
+ """Changes the length of keys, so that the progress bar fits better."""
+
+ def rename(key: str) -> str:
+ """Renames accuracy to acc."""
+ return key.replace("accuracy", "acc")
+
+ return {rename(key): value for key, value in logs.items()}
+
+ # def on_fit_begin(self) -> None:
+ # """Creates a tqdm progress bar."""
+ # self._configure_progress_bar()
+
+ def on_epoch_begin(self, epoch: int, logs: Optional[Dict]) -> None:
+ """Updates the description with the current epoch."""
+ if epoch == 1:
+ self._configure_progress_bar()
+ else:
+ self.progress_bar.reset()
+ self.progress_bar.set_description(f"Epoch {epoch}/{self.epochs}")
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """At the end of each epoch, the validation metrics are updated to the progress bar."""
+ self.val_metrics = logs
+ self.progress_bar.set_postfix(**self._key_abbreviations(logs))
+ self.progress_bar.update()
+
+ def on_train_batch_end(self, batch: int, logs: Dict) -> None:
+ """Updates the progress bar for each training step."""
+ if self.val_metrics:
+ logs.update(self.val_metrics)
+ self.progress_bar.set_postfix(**self._key_abbreviations(logs))
+ self.progress_bar.update()
+
+ def on_fit_end(self) -> None:
+ """Closes the tqdm progress bar."""
+ self.progress_bar.close()
diff --git a/training/trainer/callbacks/wandb_callbacks.py b/training/trainer/callbacks/wandb_callbacks.py
new file mode 100644
index 0000000..552a4f4
--- /dev/null
+++ b/training/trainer/callbacks/wandb_callbacks.py
@@ -0,0 +1,261 @@
+"""Callback for W&B."""
+from typing import Callable, Dict, List, Optional, Type
+
+import numpy as np
+from training.trainer.callbacks import Callback
+import wandb
+
+import text_recognizer.datasets.transforms as transforms
+from text_recognizer.models.base import Model
+
+
+class WandbCallback(Callback):
+ """A custom W&B metric logger for the trainer."""
+
+ def __init__(self, log_batch_frequency: int = None) -> None:
+ """Short summary.
+
+ Args:
+ log_batch_frequency (int): If None, metrics will be logged every epoch.
+ If set to an integer, callback will log every metrics every log_batch_frequency.
+
+ """
+ super().__init__()
+ self.log_batch_frequency = log_batch_frequency
+
+ def _on_batch_end(self, batch: int, logs: Dict) -> None:
+ if self.log_batch_frequency and batch % self.log_batch_frequency == 0:
+ wandb.log(logs, commit=True)
+
+ def on_train_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Logs training metrics."""
+ if logs is not None:
+ logs["lr"] = self.model.optimizer.param_groups[0]["lr"]
+ self._on_batch_end(batch, logs)
+
+ def on_validation_batch_end(self, batch: int, logs: Optional[Dict] = None) -> None:
+ """Logs validation metrics."""
+ if logs is not None:
+ self._on_batch_end(batch, logs)
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Logs at epoch end."""
+ wandb.log(logs, commit=True)
+
+
+class WandbImageLogger(Callback):
+ """Custom W&B callback for image logging."""
+
+ def __init__(
+ self,
+ example_indices: Optional[List] = None,
+ num_examples: int = 4,
+ transform: Optional[bool] = None,
+ ) -> None:
+ """Initializes the WandbImageLogger with the model to train.
+
+ Args:
+ example_indices (Optional[List]): Indices for validation images. Defaults to None.
+ num_examples (int): Number of random samples to take if example_indices are not specified. Defaults to 4.
+ transform (Optional[Dict]): Use transform on image or not. Defaults to None.
+
+ """
+
+ super().__init__()
+ self.caption = None
+ self.example_indices = example_indices
+ self.test_sample_indices = None
+ self.num_examples = num_examples
+ self.transform = (
+ self._configure_transform(transform) if transform is not None else None
+ )
+
+ def _configure_transform(self, transform: Dict) -> Callable:
+ args = transform["args"] or {}
+ return getattr(transforms, transform["type"])(**args)
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Sets the model and extracts validation images from the dataset."""
+ self.model = model
+ self.caption = "Validation Examples"
+ if self.example_indices is None:
+ self.example_indices = np.random.randint(
+ 0, len(self.model.val_dataset), self.num_examples
+ )
+ self.images = self.model.val_dataset.dataset.data[self.example_indices]
+ self.targets = self.model.val_dataset.dataset.targets[self.example_indices]
+ self.targets = self.targets.tolist()
+
+ def on_test_begin(self) -> None:
+ """Get samples from test dataset."""
+ self.caption = "Test Examples"
+ if self.test_sample_indices is None:
+ self.test_sample_indices = np.random.randint(
+ 0, len(self.model.test_dataset), self.num_examples
+ )
+ self.images = self.model.test_dataset.data[self.test_sample_indices]
+ self.targets = self.model.test_dataset.targets[self.test_sample_indices]
+ self.targets = self.targets.tolist()
+
+ def on_test_end(self) -> None:
+ """Log test images."""
+ self.on_epoch_end(0, {})
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Get network predictions on validation images."""
+ images = []
+ for i, image in enumerate(self.images):
+ image = self.transform(image) if self.transform is not None else image
+ pred, conf = self.model.predict_on_image(image)
+ if isinstance(self.targets[i], list):
+ ground_truth = "".join(
+ [
+ self.model.mapper(int(target_index) - 26)
+ if target_index > 35
+ else self.model.mapper(int(target_index))
+ for target_index in self.targets[i]
+ ]
+ ).rstrip("_")
+ else:
+ ground_truth = self.model.mapper(int(self.targets[i]))
+ caption = f"Prediction: {pred} Confidence: {conf:.3f} Ground Truth: {ground_truth}"
+ images.append(wandb.Image(image, caption=caption))
+
+ wandb.log({f"{self.caption}": images}, commit=False)
+
+
+class WandbSegmentationLogger(Callback):
+ """Custom W&B callback for image logging."""
+
+ def __init__(
+ self,
+ class_labels: Dict,
+ example_indices: Optional[List] = None,
+ num_examples: int = 4,
+ ) -> None:
+ """Initializes the WandbImageLogger with the model to train.
+
+ Args:
+ class_labels (Dict): A dict with int as key and class string as value.
+ example_indices (Optional[List]): Indices for validation images. Defaults to None.
+ num_examples (int): Number of random samples to take if example_indices are not specified. Defaults to 4.
+
+ """
+
+ super().__init__()
+ self.caption = None
+ self.class_labels = {int(k): v for k, v in class_labels.items()}
+ self.example_indices = example_indices
+ self.test_sample_indices = None
+ self.num_examples = num_examples
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Sets the model and extracts validation images from the dataset."""
+ self.model = model
+ self.caption = "Validation Segmentation Examples"
+ if self.example_indices is None:
+ self.example_indices = np.random.randint(
+ 0, len(self.model.val_dataset), self.num_examples
+ )
+ self.images = self.model.val_dataset.dataset.data[self.example_indices]
+ self.targets = self.model.val_dataset.dataset.targets[self.example_indices]
+ self.targets = self.targets.tolist()
+
+ def on_test_begin(self) -> None:
+ """Get samples from test dataset."""
+ self.caption = "Test Segmentation Examples"
+ if self.test_sample_indices is None:
+ self.test_sample_indices = np.random.randint(
+ 0, len(self.model.test_dataset), self.num_examples
+ )
+ self.images = self.model.test_dataset.data[self.test_sample_indices]
+ self.targets = self.model.test_dataset.targets[self.test_sample_indices]
+ self.targets = self.targets.tolist()
+
+ def on_test_end(self) -> None:
+ """Log test images."""
+ self.on_epoch_end(0, {})
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Get network predictions on validation images."""
+ images = []
+ for i, image in enumerate(self.images):
+ pred_mask = (
+ self.model.predict_on_image(image).detach().squeeze(0).cpu().numpy()
+ )
+ gt_mask = np.array(self.targets[i])
+ images.append(
+ wandb.Image(
+ image,
+ masks={
+ "predictions": {
+ "mask_data": pred_mask,
+ "class_labels": self.class_labels,
+ },
+ "ground_truth": {
+ "mask_data": gt_mask,
+ "class_labels": self.class_labels,
+ },
+ },
+ )
+ )
+
+ wandb.log({f"{self.caption}": images}, commit=False)
+
+
+class WandbReconstructionLogger(Callback):
+ """Custom W&B callback for image reconstructions logging."""
+
+ def __init__(
+ self, example_indices: Optional[List] = None, num_examples: int = 4,
+ ) -> None:
+ """Initializes the WandbImageLogger with the model to train.
+
+ Args:
+ example_indices (Optional[List]): Indices for validation images. Defaults to None.
+ num_examples (int): Number of random samples to take if example_indices are not specified. Defaults to 4.
+
+ """
+
+ super().__init__()
+ self.caption = None
+ self.example_indices = example_indices
+ self.test_sample_indices = None
+ self.num_examples = num_examples
+
+ def set_model(self, model: Type[Model]) -> None:
+ """Sets the model and extracts validation images from the dataset."""
+ self.model = model
+ self.caption = "Validation Reconstructions Examples"
+ if self.example_indices is None:
+ self.example_indices = np.random.randint(
+ 0, len(self.model.val_dataset), self.num_examples
+ )
+ self.images = self.model.val_dataset.dataset.data[self.example_indices]
+
+ def on_test_begin(self) -> None:
+ """Get samples from test dataset."""
+ self.caption = "Test Reconstructions Examples"
+ if self.test_sample_indices is None:
+ self.test_sample_indices = np.random.randint(
+ 0, len(self.model.test_dataset), self.num_examples
+ )
+ self.images = self.model.test_dataset.data[self.test_sample_indices]
+
+ def on_test_end(self) -> None:
+ """Log test images."""
+ self.on_epoch_end(0, {})
+
+ def on_epoch_end(self, epoch: int, logs: Dict) -> None:
+ """Get network predictions on validation images."""
+ images = []
+ for image in self.images:
+ reconstructed_image = (
+ self.model.predict_on_image(image).detach().squeeze(0).cpu().numpy()
+ )
+ images.append(image)
+ images.append(reconstructed_image)
+
+ wandb.log(
+ {f"{self.caption}": [wandb.Image(image) for image in images]}, commit=False,
+ )
diff --git a/training/trainer/train.py b/training/trainer/train.py
new file mode 100644
index 0000000..b770c94
--- /dev/null
+++ b/training/trainer/train.py
@@ -0,0 +1,325 @@
+"""Training script for PyTorch models."""
+
+from pathlib import Path
+import time
+from typing import Dict, List, Optional, Tuple, Type
+import warnings
+
+from einops import rearrange
+from loguru import logger
+import numpy as np
+import torch
+from torch import Tensor
+from torch.optim.swa_utils import update_bn
+from training.trainer.callbacks import Callback, CallbackList, LRScheduler, SWA
+from training.trainer.util import log_val_metric
+import wandb
+
+from text_recognizer.models import Model
+
+
+torch.backends.cudnn.benchmark = True
+np.random.seed(4711)
+torch.manual_seed(4711)
+torch.cuda.manual_seed(4711)
+
+
+warnings.filterwarnings("ignore")
+
+
+class Trainer:
+ """Trainer for training PyTorch models."""
+
+ def __init__(
+ self,
+ max_epochs: int,
+ callbacks: List[Type[Callback]],
+ transformer_model: bool = False,
+ max_norm: float = 0.0,
+ freeze_backbone: Optional[int] = None,
+ ) -> None:
+ """Initialization of the Trainer.
+
+ Args:
+ max_epochs (int): The maximum number of epochs in the training loop.
+ callbacks (CallbackList): List of callbacks to be called.
+ transformer_model (bool): Transformer model flag, modifies the input to the model. Default is False.
+ max_norm (float): Max norm for gradient cl:ipping. Defaults to 0.0.
+ freeze_backbone (Optional[int]): How many epochs to freeze the backbone for. Used when training
+ Transformers. Default is None.
+
+ """
+ # Training arguments.
+ self.start_epoch = 1
+ self.max_epochs = max_epochs
+ self.callbacks = callbacks
+ self.freeze_backbone = freeze_backbone
+
+ # Flag for setting callbacks.
+ self.callbacks_configured = False
+
+ self.transformer_model = transformer_model
+
+ self.max_norm = max_norm
+
+ # Model placeholders
+ self.model = None
+
+ def _configure_callbacks(self) -> None:
+ """Instantiate the CallbackList."""
+ if not self.callbacks_configured:
+ # If learning rate schedulers are present, they need to be added to the callbacks.
+ if self.model.swa_scheduler is not None:
+ self.callbacks.append(SWA())
+ elif self.model.lr_scheduler is not None:
+ self.callbacks.append(LRScheduler())
+
+ self.callbacks = CallbackList(self.model, self.callbacks)
+
+ def compute_metrics(
+ self, output: Tensor, targets: Tensor, loss: Tensor, batch_size: int
+ ) -> Dict:
+ """Computes metrics for output and target pairs."""
+ # Compute metrics.
+ loss = loss.detach().float().item()
+ output = output.detach()
+ targets = targets.detach()
+ if self.model.metrics is not None:
+ metrics = {}
+ for metric in self.model.metrics:
+ if metric == "cer" or metric == "wer":
+ metrics[metric] = self.model.metrics[metric](
+ output,
+ targets,
+ batch_size,
+ self.model.mapper(self.model.pad_token),
+ )
+ else:
+ metrics[metric] = self.model.metrics[metric](output, targets)
+ else:
+ metrics = {}
+ metrics["loss"] = loss
+
+ return metrics
+
+ def training_step(self, batch: int, samples: Tuple[Tensor, Tensor],) -> Dict:
+ """Performs the training step."""
+ # Pass the tensor to the device for computation.
+ data, targets = samples
+ data, targets = (
+ data.to(self.model.device),
+ targets.to(self.model.device),
+ )
+
+ batch_size = data.shape[0]
+
+ # Placeholder for uxiliary loss.
+ aux_loss = None
+
+ # Forward pass.
+ # Get the network prediction.
+ if self.transformer_model:
+ if self.freeze_backbone is not None and batch < self.freeze_backbone:
+ with torch.no_grad():
+ image_features = self.model.network.extract_image_features(data)
+
+ if isinstance(image_features, Tuple):
+ image_features, _ = image_features
+
+ output = self.model.network.decode_image_features(
+ image_features, targets[:, :-1]
+ )
+ else:
+ output = self.model.network.forward(data, targets[:, :-1])
+ if isinstance(output, Tuple):
+ output, aux_loss = output
+ output = rearrange(output, "b t v -> (b t) v")
+ targets = rearrange(targets[:, 1:], "b t -> (b t)").long()
+ else:
+ output = self.model.forward(data)
+
+ if isinstance(output, Tuple):
+ output, aux_loss = output
+ targets = data
+
+ # Compute the loss.
+ loss = self.model.criterion(output, targets)
+
+ if aux_loss is not None:
+ loss += aux_loss
+
+ # Backward pass.
+ # Clear the previous gradients.
+ for p in self.model.network.parameters():
+ p.grad = None
+
+ # Compute the gradients.
+ loss.backward()
+
+ if self.max_norm > 0:
+ torch.nn.utils.clip_grad_norm_(
+ self.model.network.parameters(), self.max_norm
+ )
+
+ # Perform updates using calculated gradients.
+ self.model.optimizer.step()
+
+ metrics = self.compute_metrics(output, targets, loss, batch_size)
+
+ return metrics
+
+ def train(self) -> None:
+ """Runs the training loop for one epoch."""
+ # Set model to traning mode.
+ self.model.train()
+
+ for batch, samples in enumerate(self.model.train_dataloader()):
+ self.callbacks.on_train_batch_begin(batch)
+ metrics = self.training_step(batch, samples)
+ self.callbacks.on_train_batch_end(batch, logs=metrics)
+
+ @torch.no_grad()
+ def validation_step(self, batch: int, samples: Tuple[Tensor, Tensor],) -> Dict:
+ """Performs the validation step."""
+
+ # Pass the tensor to the device for computation.
+ data, targets = samples
+ data, targets = (
+ data.to(self.model.device),
+ targets.to(self.model.device),
+ )
+
+ batch_size = data.shape[0]
+
+ # Placeholder for uxiliary loss.
+ aux_loss = None
+
+ # Forward pass.
+ # Get the network prediction.
+ # Use SWA if available and using test dataset.
+ if self.transformer_model:
+ output = self.model.network.forward(data, targets[:, :-1])
+ if isinstance(output, Tuple):
+ output, aux_loss = output
+ output = rearrange(output, "b t v -> (b t) v")
+ targets = rearrange(targets[:, 1:], "b t -> (b t)").long()
+ else:
+ output = self.model.forward(data)
+
+ if isinstance(output, Tuple):
+ output, aux_loss = output
+ targets = data
+
+ # Compute the loss.
+ loss = self.model.criterion(output, targets)
+
+ if aux_loss is not None:
+ loss += aux_loss
+
+ # Compute metrics.
+ metrics = self.compute_metrics(output, targets, loss, batch_size)
+
+ return metrics
+
+ def validate(self) -> Dict:
+ """Runs the validation loop for one epoch."""
+ # Set model to eval mode.
+ self.model.eval()
+
+ # Summary for the current eval loop.
+ summary = []
+
+ for batch, samples in enumerate(self.model.val_dataloader()):
+ self.callbacks.on_validation_batch_begin(batch)
+ metrics = self.validation_step(batch, samples)
+ self.callbacks.on_validation_batch_end(batch, logs=metrics)
+ summary.append(metrics)
+
+ # Compute mean of all metrics.
+ metrics_mean = {
+ "val_" + metric: np.mean([x[metric] for x in summary])
+ for metric in summary[0]
+ }
+
+ return metrics_mean
+
+ def fit(self, model: Type[Model]) -> None:
+ """Runs the training and evaluation loop."""
+
+ # Sets model, loads the data, criterion, and optimizers.
+ self.model = model
+ self.model.prepare_data()
+ self.model.configure_model()
+
+ # Configure callbacks.
+ self._configure_callbacks()
+
+ # Set start time.
+ t_start = time.time()
+
+ self.callbacks.on_fit_begin()
+
+ # Run the training loop.
+ for epoch in range(self.start_epoch, self.max_epochs + 1):
+ self.callbacks.on_epoch_begin(epoch)
+
+ # Perform one training pass over the training set.
+ self.train()
+
+ # Evaluate the model on the validation set.
+ val_metrics = self.validate()
+ log_val_metric(val_metrics, epoch)
+
+ self.callbacks.on_epoch_end(epoch, logs=val_metrics)
+
+ if self.model.stop_training:
+ break
+
+ # Calculate the total training time.
+ t_end = time.time()
+ t_training = t_end - t_start
+
+ self.callbacks.on_fit_end()
+
+ logger.info(f"Training took {t_training:.2f} s.")
+
+ # "Teardown".
+ self.model = None
+
+ def test(self, model: Type[Model]) -> Dict:
+ """Run inference on test data."""
+
+ # Sets model, loads the data, criterion, and optimizers.
+ self.model = model
+ self.model.prepare_data()
+ self.model.configure_model()
+
+ # Configure callbacks.
+ self._configure_callbacks()
+
+ self.callbacks.on_test_begin()
+
+ self.model.eval()
+
+ # Check if SWA network is available.
+ self.model.use_swa_model()
+
+ # Summary for the current test loop.
+ summary = []
+
+ for batch, samples in enumerate(self.model.test_dataloader()):
+ metrics = self.validation_step(batch, samples)
+ summary.append(metrics)
+
+ self.callbacks.on_test_end()
+
+ # Compute mean of all test metrics.
+ metrics_mean = {
+ "test_" + metric: np.mean([x[metric] for x in summary])
+ for metric in summary[0]
+ }
+
+ # "Teardown".
+ self.model = None
+
+ return metrics_mean
diff --git a/training/trainer/util.py b/training/trainer/util.py
new file mode 100644
index 0000000..7cf1b45
--- /dev/null
+++ b/training/trainer/util.py
@@ -0,0 +1,28 @@
+"""Utility functions for training neural networks."""
+from typing import Dict, Optional
+
+from loguru import logger
+
+
+def log_val_metric(metrics_mean: Dict, epoch: Optional[int] = None) -> None:
+ """Logging of val metrics to file/terminal."""
+ log_str = "Validation metrics " + (f"at epoch {epoch} - " if epoch else " - ")
+ logger.debug(log_str + " - ".join(f"{k}: {v:.4f}" for k, v in metrics_mean.items()))
+
+
+class RunningAverage:
+ """Maintains a running average."""
+
+ def __init__(self) -> None:
+ """Initializes the parameters."""
+ self.steps = 0
+ self.total = 0
+
+ def update(self, val: float) -> None:
+ """Updates the parameters."""
+ self.total += val
+ self.steps += 1
+
+ def __call__(self) -> float:
+ """Computes the running average."""
+ return self.total / float(self.steps)
generated by cgit v1.2.3-70-g09d2 (git 2.46.0) at 2024-11-10 07:02:21 +0000