Spaces:
Running
Running
File size: 6,495 Bytes
17ee76b |
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 |
"""This module contains callbacks to be used along with `TorchModel`."""
import datetime
import logging
import os
import time
from abc import ABC, abstractmethod
import matplotlib.pyplot as plt
class Callback(ABC):
@abstractmethod
def on_training_start(self, epochs) -> None:
pass
@abstractmethod
def on_training_end(self, model) -> None:
pass
@abstractmethod
def on_epoch_start(self, epoch_num, epoch_iterations) -> None:
pass
@abstractmethod
def on_epoch_step(self, global_iteration, epoch_iteration, loss) -> None:
pass
@abstractmethod
def on_epoch_end(self, loss) -> None:
pass
@abstractmethod
def on_evaluation_start(self, val_iterations) -> None:
pass
@abstractmethod
def on_evaluation_step(self, iteration, model_outputs, targets, loss) -> None:
pass
@abstractmethod
def on_evaluation_end(self) -> None:
pass
@abstractmethod
def on_training_iteration_end(self, train_loss, val_loss) -> None:
pass
class DefaultModelCallback(Callback):
"""A callback that simply logs the loss for epochs during training and
evaluation."""
def __init__(self, log_every=10, visualization_dir=None) -> None:
"""
Args:
log_every (iterations): logging intervals
"""
super().__init__()
self.visualization_dir = visualization_dir
self._log_every = log_every
self._epochs = 0
self._epoch = 0
self._epoch_iterations = 0
self._val_iterations = 0
self._start_time = 0.0
self._train_losses = []
self._val_loss = []
def on_training_start(self, epochs) -> None:
logging.info(f"Training for {epochs} epochs")
self._epochs = epochs
self._train_losses = []
self._val_loss = []
def on_training_end(self, model) -> None:
if self.visualization_dir is not None:
plt.figure()
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.plot(
range(1, self._epochs + 1), self._train_losses, label="Training loss"
)
if self._val_loss:
plt.plot(
range(1, self._epochs + 1), self._val_loss, label="Validation loss"
)
plt.savefig(os.path.join(self.visualization_dir, "loss.png"))
plt.close()
def on_epoch_start(self, epoch_num: int, epoch_iterations: int) -> None:
self._epoch = epoch_num
self._epoch_iterations = epoch_iterations
self._start_time = time.time()
def on_epoch_step(
self, global_iteration: int, epoch_iteration: int, loss: float
) -> None:
if epoch_iteration % self._log_every == 0:
average_time = round(
(time.time() - self._start_time) / (epoch_iteration + 1), 3
)
loss_string = f"loss: {loss}"
# pylint: disable=line-too-long
logging.info(
f"Epoch {self._epoch}/{self._epochs} Iteration {epoch_iteration}/{self._epoch_iterations} {loss_string} Time: {average_time} seconds/iteration"
)
def on_epoch_end(self, loss) -> None:
self._train_losses.append(loss)
def on_evaluation_start(self, val_iterations) -> None:
self._val_iterations = val_iterations
def on_evaluation_step(self, iteration, model_outputs, targets, loss) -> None:
if iteration % self._log_every == 0:
logging.info(f"Iteration {iteration}/{self._val_iterations}")
def on_evaluation_end(self) -> None:
pass
def on_training_iteration_end(self, train_loss, val_loss) -> None:
# pylint: disable=line-too-long
train_loss_string = f"Train loss: {train_loss}"
if val_loss:
val_loss_string = f"Validation loss: {val_loss}"
logging.info(
f"""
============================================================================================================================
Epoch {self._epoch}/{self._epochs} {train_loss_string} {val_loss_string} time: {datetime.timedelta(seconds=time.time() - self._start_time)}
============================================================================================================================
"""
)
else:
logging.info(
f"""
============================================================================================================================
Epoch {self._epoch}/{self._epochs} {train_loss_string} time: {datetime.timedelta(seconds=time.time() - self._start_time)}
============================================================================================================================
"""
)
class TensorBoardCallback(Callback):
"""A callback that simply logs the loss for epochs during training and
evaluation."""
def __init__(self, tb_writer) -> None:
"""
Args:
tb_writer: tensorboard logger instance
"""
super().__init__()
self.tb_writer = tb_writer
self.epoch = 0
def on_training_start(self, epochs) -> None:
pass
def on_training_end(self, model) -> None:
pass
def on_epoch_start(self, epoch_num, epoch_iterations) -> None:
self.epoch = epoch_num
def on_epoch_step(self, global_iteration, epoch_iteration, loss) -> None:
self.tb_writer.add_scalars(
"Train loss (iterations)", {"Loss": loss}, global_iteration
)
def on_epoch_end(self, loss) -> None:
pass
def on_evaluation_start(self, val_iterations) -> None:
pass
def on_evaluation_step(self, iteration, model_outputs, targets, loss) -> None:
pass
def on_evaluation_end(self) -> None:
pass
def on_training_iteration_end(self, train_loss, val_loss) -> None:
if train_loss is not None:
self.tb_writer.add_scalars(
"Epoch loss", {"Loss (train)": train_loss}, self.epoch
)
if val_loss is not None:
self.tb_writer.add_scalars(
"Epoch loss", {"Loss (validation)": val_loss}, self.epoch
)
|