summaryrefslogtreecommitdiff
path: root/src/text_recognizer/models/base.py
blob: 84a86cac4fdf6d68d9b053eefea56b860966f481 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
"""Abstract Model class for PyTorch neural networks."""

from abc import ABC, abstractmethod
from glob import glob
from pathlib import Path
import re
import shutil
from typing import Callable, Dict, Optional, Tuple, Type

from loguru import logger
import torch
from torch import nn
from torchsummary import summary


WEIGHT_DIRNAME = Path(__file__).parents[1].resolve() / "weights"


class Model(ABC):
    """Abstract Model class with composition of different parts defining a PyTorch neural network."""

    def __init__(
        self,
        network_fn: Type[nn.Module],
        network_args: Optional[Dict] = None,
        data_loader: Optional[Callable] = None,
        data_loader_args: Optional[Dict] = None,
        metrics: Optional[Dict] = None,
        criterion: Optional[Callable] = None,
        criterion_args: Optional[Dict] = None,
        optimizer: Optional[Callable] = None,
        optimizer_args: Optional[Dict] = None,
        lr_scheduler: Optional[Callable] = None,
        lr_scheduler_args: Optional[Dict] = None,
        device: Optional[str] = None,
    ) -> None:
        """Base class, to be inherited by model for specific type of data.

        Args:
            network_fn (Type[nn.Module]): The PyTorch network.
            network_args (Optional[Dict]): Arguments for the network. Defaults to None.
            data_loader (Optional[Callable]): A function that fetches train and val DataLoader.
            data_loader_args (Optional[Dict]):  Arguments for the DataLoader.
            metrics (Optional[Dict]): Metrics to evaluate the performance with. Defaults to None.
            criterion (Optional[Callable]): The criterion to evaulate the preformance of the network.
                Defaults to None.
            criterion_args (Optional[Dict]): Dict of arguments for criterion. Defaults to None.
            optimizer (Optional[Callable]): The optimizer for updating the weights. Defaults to None.
            optimizer_args (Optional[Dict]): Dict of arguments for optimizer. Defaults to None.
            lr_scheduler (Optional[Callable]): A PyTorch learning rate scheduler. Defaults to None.
            lr_scheduler_args (Optional[Dict]): Dict of arguments for learning rate scheduler. Defaults to
                None.
            device (Optional[str]): Name of the device to train on. Defaults to None.

        """

        # Fetch data loaders.
        if data_loader_args is not None:
            self._data_loaders = data_loader(**data_loader_args)
            dataset_name = self._data_loaders.__name__
            self._mapping = self._data_loaders.mapping
        else:
            self._mapping = None
            dataset_name = "*"
            self._data_loaders = None

        self._name = f"{self.__class__.__name__}_{dataset_name}_{network_fn.__name__}"

        if metrics is not None:
            self._metrics = metrics

        # Set the device.
        if device is None:
            self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        else:
            self._device = device

        # Load network.
        self._network = None
        self._network_args = network_args
        # If no network arguemnts are given, load pretrained weights if they exist.
        if self._network_args is None:
            self.load_weights(network_fn)
        else:
            self._network = network_fn(**self._network_args)

        # To device.
        self._network.to(self._device)

        # Set criterion.
        self._criterion = None
        if criterion is not None:
            self._criterion = criterion(**criterion_args)

        # Set optimizer.
        self._optimizer = None
        if optimizer is not None:
            self._optimizer = optimizer(self._network.parameters(), **optimizer_args)

        # Set learning rate scheduler.
        self._lr_scheduler = None
        if lr_scheduler is not None:
            # OneCycleLR needs the number of steps in an epoch as an input argument.
            if "OneCycleLR" in str(lr_scheduler):
                lr_scheduler_args["steps_per_epoch"] = len(self._data_loaders("train"))
            self._lr_scheduler = lr_scheduler(self._optimizer, **lr_scheduler_args)

        # Extract the input shape for the torchsummary.
        if isinstance(self._network_args["input_size"], int):
            self._input_shape = (1,) + tuple([self._network_args["input_size"]])
        else:
            self._input_shape = (1,) + tuple(self._network_args["input_size"])

        # Experiment directory.
        self.model_dir = None

        # Flag for stopping training.
        self.stop_training = False

    @property
    def __name__(self) -> str:
        """Returns the name of the model."""
        return self._name

    @property
    def input_shape(self) -> Tuple[int, ...]:
        """The input shape."""
        return self._input_shape

    @property
    def mapping(self) -> Dict:
        """Returns the class mapping."""
        return self._mapping

    def eval(self) -> None:
        """Sets the network to evaluation mode."""
        self._network.eval()

    def train(self) -> None:
        """Sets the network to train mode."""
        self._network.train()

    @property
    def device(self) -> str:
        """Device where the weights are stored, i.e. cpu or cuda."""
        return self._device

    @property
    def metrics(self) -> Optional[Dict]:
        """Metrics."""
        return self._metrics

    @property
    def criterion(self) -> Optional[Callable]:
        """Criterion."""
        return self._criterion

    @property
    def optimizer(self) -> Optional[Callable]:
        """Optimizer."""
        return self._optimizer

    @property
    def lr_scheduler(self) -> Optional[Callable]:
        """Learning rate scheduler."""
        return self._lr_scheduler

    @property
    def data_loaders(self) -> Optional[Dict]:
        """Dataloaders."""
        return self._data_loaders

    @property
    def network(self) -> nn.Module:
        """Neural network."""
        return self._network

    @property
    def weights_filename(self) -> str:
        """Filepath to the network weights."""
        WEIGHT_DIRNAME.mkdir(parents=True, exist_ok=True)
        return str(WEIGHT_DIRNAME / f"{self._name}_weights.pt")

    def summary(self) -> None:
        """Prints a summary of the network architecture."""
        device = re.sub("[^A-Za-z]+", "", self.device)
        summary(self._network, self._input_shape, device=device)

    def _get_state_dict(self) -> Dict:
        """Get the state dict of the model."""
        state = {"model_state": self._network.state_dict()}

        if self._optimizer is not None:
            state["optimizer_state"] = self._optimizer.state_dict()

        if self._lr_scheduler is not None:
            state["scheduler_state"] = self._lr_scheduler.state_dict()

        return state

    def load_checkpoint(self, path: Path) -> int:
        """Load a previously saved checkpoint.

        Args:
            path (Path): Path to the experiment with the checkpoint.

        Returns:
            epoch (int): The last epoch when the checkpoint was created.

        """
        logger.debug("Loading checkpoint...")
        if not path.exists():
            logger.debug("File does not exist {str(path)}")

        checkpoint = torch.load(str(path))
        self._network.load_state_dict(checkpoint["model_state"])

        if self._optimizer is not None:
            self._optimizer.load_state_dict(checkpoint["optimizer_state"])

        if self._lr_scheduler is not None:
            self._lr_scheduler.load_state_dict(checkpoint["scheduler_state"])

        epoch = checkpoint["epoch"]

        return epoch

    def save_checkpoint(self, is_best: bool, epoch: int, val_metric: str) -> None:
        """Saves a checkpoint of the model.

        Args:
            is_best (bool): If it is the currently best model.
            epoch (int): The epoch of the checkpoint.
            val_metric (str): Validation metric.

        Raises:
            ValueError: If the self.model_dir is not set.

        """
        state = self._get_state_dict()
        state["is_best"] = is_best
        state["epoch"] = epoch
        state["network_args"] = self._network_args

        if self.model_dir is None:
            raise ValueError("Experiment directory is not set.")

        self.model_dir.mkdir(parents=True, exist_ok=True)

        logger.debug("Saving checkpoint...")
        filepath = str(self.model_dir / "last.pt")
        torch.save(state, filepath)

        if is_best:
            logger.debug(
                f"Found a new best {val_metric}. Saving best checkpoint and weights."
            )
            shutil.copyfile(filepath, str(self.model_dir / "best.pt"))

    def load_weights(self, network_fn: Type[nn.Module]) -> None:
        """Load the network weights."""
        logger.debug("Loading network with pretrained weights.")
        filename = glob(self.weights_filename)[0]
        if not filename:
            raise FileNotFoundError(
                f"Could not find any pretrained weights at {self.weights_filename}"
            )
        # Loading state directory.
        state_dict = torch.load(filename, map_location=torch.device(self._device))
        self._network_args = state_dict["network_args"]
        weights = state_dict["model_state"]

        # Initializes the network with trained weights.
        self._network = network_fn(**self._network_args)
        self._network.load_state_dict(weights)

    def save_weights(self, path: Path) -> None:
        """Save the network weights."""
        logger.debug("Saving the best network weights.")
        shutil.copyfile(str(path / "best.pt"), self.weights_filename)