Spaces:
Runtime error
Runtime error
File size: 5,595 Bytes
2c0f55c | 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 | """A generic training wrapper."""
from copy import deepcopy
import logging
from typing import Callable, List, Optional
import torch
from torch.utils.data import DataLoader
LOGGER = logging.getLogger(__name__)
class Trainer:
def __init__(
self,
epochs: int = 20,
batch_size: int = 32,
device: str = "cpu",
optimizer_fn: Callable = torch.optim.Adam,
optimizer_kwargs: dict = {"lr": 1e-3},
use_scheduler: bool = False,
) -> None:
self.epochs = epochs
self.batch_size = batch_size
self.device = device
self.optimizer_fn = optimizer_fn
self.optimizer_kwargs = optimizer_kwargs
self.epoch_test_losses: List[float] = []
self.use_scheduler = use_scheduler
def forward_and_loss(model, criterion, batch_x, batch_y, **kwargs):
batch_out = model(batch_x)
batch_loss = criterion(batch_out, batch_y)
return batch_out, batch_loss
class GDTrainer(Trainer):
def train(
self,
dataset: torch.utils.data.Dataset,
model: torch.nn.Module,
test_len: Optional[float] = None,
test_dataset: Optional[torch.utils.data.Dataset] = None,
):
if test_dataset is not None:
train = dataset
test = test_dataset
else:
test_len = int(len(dataset) * test_len)
train_len = len(dataset) - test_len
lengths = [train_len, test_len]
train, test = torch.utils.data.random_split(dataset, lengths)
train_loader = DataLoader(
train,
batch_size=self.batch_size,
shuffle=True,
drop_last=True,
num_workers=6,
)
test_loader = DataLoader(
test,
batch_size=self.batch_size,
shuffle=True,
drop_last=True,
num_workers=6,
)
criterion = torch.nn.BCEWithLogitsLoss()
optim = self.optimizer_fn(model.parameters(), **self.optimizer_kwargs)
best_model = None
best_acc = 0
LOGGER.info(f"Starting training for {self.epochs} epochs!")
forward_and_loss_fn = forward_and_loss
if self.use_scheduler:
batches_per_epoch = len(train_loader) * 2 # every 2nd epoch
scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer=optim,
T_0=batches_per_epoch,
T_mult=1,
eta_min=5e-6,
# verbose=True,
)
use_cuda = self.device != "cpu"
for epoch in range(self.epochs):
LOGGER.info(f"Epoch num: {epoch}")
running_loss = 0
num_correct = 0.0
num_total = 0.0
model.train()
for i, (batch_x, _, batch_y) in enumerate(train_loader):
batch_size = batch_x.size(0)
num_total += batch_size
batch_x = batch_x.to(self.device)
batch_y = batch_y.unsqueeze(1).type(torch.float32).to(self.device)
batch_out, batch_loss = forward_and_loss_fn(
model, criterion, batch_x, batch_y, use_cuda=use_cuda
)
batch_pred = (torch.sigmoid(batch_out) + 0.5).int()
num_correct += (batch_pred == batch_y.int()).sum(dim=0).item()
running_loss += batch_loss.item() * batch_size
if i % 100 == 0:
LOGGER.info(
f"[{epoch:04d}][{i:05d}]: {running_loss / num_total} {num_correct/num_total*100}"
)
optim.zero_grad()
batch_loss.backward()
optim.step()
if self.use_scheduler:
scheduler.step()
running_loss /= num_total
train_accuracy = (num_correct / num_total) * 100
LOGGER.info(
f"Epoch [{epoch+1}/{self.epochs}]: train/loss: {running_loss}, train/accuracy: {train_accuracy}"
)
test_running_loss = 0.0
num_correct = 0.0
num_total = 0.0
model.eval()
eer_val = 0
for batch_x, _, batch_y in test_loader:
batch_size = batch_x.size(0)
num_total += batch_size
batch_x = batch_x.to(self.device)
with torch.no_grad():
batch_pred = model(batch_x)
batch_y = batch_y.unsqueeze(1).type(torch.float32).to(self.device)
batch_loss = criterion(batch_pred, batch_y)
test_running_loss += batch_loss.item() * batch_size
batch_pred = torch.sigmoid(batch_pred)
batch_pred_label = (batch_pred + 0.5).int()
num_correct += (batch_pred_label == batch_y.int()).sum(dim=0).item()
if num_total == 0:
num_total = 1
test_running_loss /= num_total
test_acc = 100 * (num_correct / num_total)
LOGGER.info(
f"Epoch [{epoch+1}/{self.epochs}]: test/loss: {test_running_loss}, test/accuracy: {test_acc}, test/eer: {eer_val}"
)
if best_model is None or test_acc > best_acc:
best_acc = test_acc
best_model = deepcopy(model.state_dict())
LOGGER.info(
f"[{epoch:04d}]: {running_loss} - train acc: {train_accuracy} - test_acc: {test_acc}"
)
model.load_state_dict(best_model)
return model
|