|
| 1 | +""" |
| 2 | +CSV logger |
| 3 | +---------- |
| 4 | +
|
| 5 | +CSV logger for basic experiment logging that does not require opening ports |
| 6 | +
|
| 7 | +""" |
| 8 | +import io |
| 9 | +import os |
| 10 | +import csv |
| 11 | +import torch |
| 12 | +from argparse import Namespace |
| 13 | +from typing import Optional, Dict, Any, Union |
| 14 | + |
| 15 | +from pytorch_lightning import _logger as log |
| 16 | +from pytorch_lightning.core.saving import save_hparams_to_yaml |
| 17 | +from pytorch_lightning.loggers.base import LightningLoggerBase |
| 18 | +from pytorch_lightning.utilities.distributed import rank_zero_warn, rank_zero_only |
| 19 | + |
| 20 | + |
| 21 | +class ExperimentWriter(object): |
| 22 | + r""" |
| 23 | + Experiment writer for CSVLogger. |
| 24 | +
|
| 25 | + Currently supports to log hyperparameters and metrics in YAML and CSV |
| 26 | + format, respectively. |
| 27 | +
|
| 28 | + Args: |
| 29 | + log_dir: Directory for the experiment logs |
| 30 | + """ |
| 31 | + |
| 32 | + NAME_HPARAMS_FILE = 'hparams.yaml' |
| 33 | + NAME_METRICS_FILE = 'metrics.csv' |
| 34 | + |
| 35 | + def __init__(self, log_dir: str) -> None: |
| 36 | + self.hparams = {} |
| 37 | + self.metrics = [] |
| 38 | + |
| 39 | + self.log_dir = log_dir |
| 40 | + if os.path.exists(self.log_dir): |
| 41 | + rank_zero_warn( |
| 42 | + f"Experiment logs directory {self.log_dir} exists and is not empty." |
| 43 | + " Previous log files in this directory will be deleted when the new ones are saved!" |
| 44 | + ) |
| 45 | + os.makedirs(self.log_dir, exist_ok=True) |
| 46 | + |
| 47 | + self.metrics_file_path = os.path.join(self.log_dir, self.NAME_METRICS_FILE) |
| 48 | + |
| 49 | + def log_hparams(self, params: Dict[str, Any]) -> None: |
| 50 | + """Record hparams""" |
| 51 | + self.hparams.update(params) |
| 52 | + |
| 53 | + def log_metrics(self, metrics_dict: Dict[str, float], step: Optional[int] = None) -> None: |
| 54 | + """Record metrics""" |
| 55 | + def _handle_value(value): |
| 56 | + if isinstance(value, torch.Tensor): |
| 57 | + return value.item() |
| 58 | + return value |
| 59 | + |
| 60 | + if step is None: |
| 61 | + step = len(self.metrics) |
| 62 | + |
| 63 | + metrics = {k: _handle_value(v) for k, v in metrics_dict.items()} |
| 64 | + metrics['step'] = step |
| 65 | + self.metrics.append(metrics) |
| 66 | + |
| 67 | + def save(self) -> None: |
| 68 | + """Save recorded hparams and metrics into files""" |
| 69 | + hparams_file = os.path.join(self.log_dir, self.NAME_HPARAMS_FILE) |
| 70 | + save_hparams_to_yaml(hparams_file, self.hparams) |
| 71 | + |
| 72 | + if not self.metrics: |
| 73 | + return |
| 74 | + |
| 75 | + last_m = {} |
| 76 | + for m in self.metrics: |
| 77 | + last_m.update(m) |
| 78 | + metrics_keys = list(last_m.keys()) |
| 79 | + |
| 80 | + with io.open(self.metrics_file_path, 'w', newline='') as f: |
| 81 | + self.writer = csv.DictWriter(f, fieldnames=metrics_keys) |
| 82 | + self.writer.writeheader() |
| 83 | + self.writer.writerows(self.metrics) |
| 84 | + |
| 85 | + |
| 86 | +class CSVLogger(LightningLoggerBase): |
| 87 | + r""" |
| 88 | + Log to local file system in yaml and CSV format. Logs are saved to |
| 89 | + ``os.path.join(save_dir, name, version)``. |
| 90 | +
|
| 91 | + Example: |
| 92 | + >>> from pytorch_lightning import Trainer |
| 93 | + >>> from pytorch_lightning.loggers import CSVLogger |
| 94 | + >>> logger = CSVLogger("logs", name="my_exp_name") |
| 95 | + >>> trainer = Trainer(logger=logger) |
| 96 | +
|
| 97 | + Args: |
| 98 | + save_dir: Save directory |
| 99 | + name: Experiment name. Defaults to ``'default'``. |
| 100 | + version: Experiment version. If version is not specified the logger inspects the save |
| 101 | + directory for existing versions, then automatically assigns the next available version. |
| 102 | + """ |
| 103 | + |
| 104 | + def __init__(self, |
| 105 | + save_dir: str, |
| 106 | + name: Optional[str] = "default", |
| 107 | + version: Optional[Union[int, str]] = None): |
| 108 | + |
| 109 | + super().__init__() |
| 110 | + self._save_dir = save_dir |
| 111 | + self._name = name or '' |
| 112 | + self._version = version |
| 113 | + self._experiment = None |
| 114 | + |
| 115 | + @property |
| 116 | + def root_dir(self) -> str: |
| 117 | + """ |
| 118 | + Parent directory for all checkpoint subdirectories. |
| 119 | + If the experiment name parameter is ``None`` or the empty string, no experiment subdirectory is used |
| 120 | + and the checkpoint will be saved in "save_dir/version_dir" |
| 121 | + """ |
| 122 | + if not self.name: |
| 123 | + return self.save_dir |
| 124 | + return os.path.join(self.save_dir, self.name) |
| 125 | + |
| 126 | + @property |
| 127 | + def log_dir(self) -> str: |
| 128 | + """ |
| 129 | + The log directory for this run. By default, it is named |
| 130 | + ``'version_${self.version}'`` but it can be overridden by passing a string value |
| 131 | + for the constructor's version parameter instead of ``None`` or an int. |
| 132 | + """ |
| 133 | + # create a pseudo standard path ala test-tube |
| 134 | + version = self.version if isinstance(self.version, str) else f"version_{self.version}" |
| 135 | + log_dir = os.path.join(self.root_dir, version) |
| 136 | + return log_dir |
| 137 | + |
| 138 | + @property |
| 139 | + def save_dir(self) -> Optional[str]: |
| 140 | + return self._save_dir |
| 141 | + |
| 142 | + @property |
| 143 | + def experiment(self) -> ExperimentWriter: |
| 144 | + r""" |
| 145 | +
|
| 146 | + Actual ExperimentWriter object. To use ExperimentWriter features in your |
| 147 | + :class:`~pytorch_lightning.core.lightning.LightningModule` do the following. |
| 148 | +
|
| 149 | + Example:: |
| 150 | +
|
| 151 | + self.logger.experiment.some_experiment_writer_function() |
| 152 | +
|
| 153 | + """ |
| 154 | + if self._experiment: |
| 155 | + return self._experiment |
| 156 | + |
| 157 | + os.makedirs(self.root_dir, exist_ok=True) |
| 158 | + self._experiment = ExperimentWriter(log_dir=self.log_dir) |
| 159 | + return self._experiment |
| 160 | + |
| 161 | + @rank_zero_only |
| 162 | + def log_hyperparams(self, params: Union[Dict[str, Any], Namespace]) -> None: |
| 163 | + params = self._convert_params(params) |
| 164 | + self.experiment.log_hparams(params) |
| 165 | + |
| 166 | + @rank_zero_only |
| 167 | + def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None) -> None: |
| 168 | + self.experiment.log_metrics(metrics, step) |
| 169 | + |
| 170 | + @rank_zero_only |
| 171 | + def save(self) -> None: |
| 172 | + super().save() |
| 173 | + self.experiment.save() |
| 174 | + |
| 175 | + @rank_zero_only |
| 176 | + def finalize(self, status: str) -> None: |
| 177 | + self.save() |
| 178 | + |
| 179 | + @property |
| 180 | + def name(self) -> str: |
| 181 | + return self._name |
| 182 | + |
| 183 | + @property |
| 184 | + def version(self) -> int: |
| 185 | + if self._version is None: |
| 186 | + self._version = self._get_next_version() |
| 187 | + return self._version |
| 188 | + |
| 189 | + def _get_next_version(self): |
| 190 | + root_dir = os.path.join(self._save_dir, self.name) |
| 191 | + |
| 192 | + if not os.path.isdir(root_dir): |
| 193 | + log.warning('Missing logger folder: %s', root_dir) |
| 194 | + return 0 |
| 195 | + |
| 196 | + existing_versions = [] |
| 197 | + for d in os.listdir(root_dir): |
| 198 | + if os.path.isdir(os.path.join(root_dir, d)) and d.startswith("version_"): |
| 199 | + existing_versions.append(int(d.split("_")[1])) |
| 200 | + |
| 201 | + if len(existing_versions) == 0: |
| 202 | + return 0 |
| 203 | + |
| 204 | + return max(existing_versions) + 1 |
0 commit comments