File size: 3,215 Bytes
f3270e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (C) 2021-2025, Mindee.

# This program is licensed under the Apache License 2.0.
# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.

import math

import matplotlib.pyplot as plt
import numpy as np


def plot_samples(images, targets):
    # Unnormalize image
    num_samples = min(len(images), 12)
    num_cols = min(len(images), 4)
    num_rows = int(math.ceil(num_samples / num_cols))
    _, axes = plt.subplots(num_rows, num_cols, figsize=(20, 5))
    for idx in range(num_samples):
        img = (255 * images[idx].numpy()).round().clip(0, 255).astype(np.uint8)
        if img.shape[0] == 3 and img.shape[2] != 3:
            img = img.transpose(1, 2, 0)

        row_idx = idx // num_cols
        col_idx = idx % num_cols
        ax = axes[row_idx] if num_rows > 1 else axes
        ax = ax[col_idx] if num_cols > 1 else ax

        ax.imshow(img)
        ax.set_title(targets[idx])
    # Disable axis
    for ax in axes.ravel():
        ax.axis("off")

    plt.show()


def plot_recorder(lr_recorder, loss_recorder, beta: float = 0.95, **kwargs) -> None:
    """Display the results of the LR grid search.
    Adapted from https://github.com/frgfm/Holocron/blob/master/holocron/trainer/core.py.

    Args:
        lr_recorder: list of LR values
        loss_recorder: list of loss values
        beta (float, optional): smoothing factor
        **kwargs: keyword arguments from `matplotlib.pyplot.show`.
    """
    if len(lr_recorder) != len(loss_recorder) or len(lr_recorder) == 0:
        raise AssertionError("Both `lr_recorder` and `loss_recorder` should have the same length")

    # Exp moving average of loss
    smoothed_losses = []
    avg_loss = 0.0
    for idx, loss in enumerate(loss_recorder):
        avg_loss = beta * avg_loss + (1 - beta) * loss
        smoothed_losses.append(avg_loss / (1 - beta ** (idx + 1)))

    # Properly rescale Y-axis
    data_slice = slice(
        min(len(loss_recorder) // 10, 10),
        -min(len(loss_recorder) // 20, 5) if len(loss_recorder) >= 20 else len(loss_recorder),
    )
    vals = np.array(smoothed_losses[data_slice])
    min_idx = vals.argmin()
    max_val = vals.max() if min_idx is None else vals[: min_idx + 1].max()  # type: ignore[misc]
    delta = max_val - vals[min_idx]

    plt.plot(lr_recorder[data_slice], smoothed_losses[data_slice])
    plt.xscale("log")
    plt.xlabel("Learning Rate")
    plt.ylabel("Training loss")
    plt.ylim(vals[min_idx] - 0.1 * delta, max_val + 0.2 * delta)
    plt.grid(True, linestyle="--", axis="x")
    plt.show(**kwargs)


class EarlyStopper:
    def __init__(self, patience: int = 5, min_delta: float = 0.01):
        self.patience = patience
        self.min_delta = min_delta
        self.counter = 0
        self.min_validation_loss = float("inf")

    def early_stop(self, validation_loss: float) -> bool:
        if validation_loss < self.min_validation_loss:
            self.min_validation_loss = validation_loss
            self.counter = 0
        elif validation_loss > (self.min_validation_loss + self.min_delta):
            self.counter += 1
            if self.counter >= self.patience:
                return True
        return False