class PneumoniaTrainer:
def __init__(
self, model_name: str, epochs: int = 100, config: Optional[dict] = None
):
self.output_model = model_name + ".pth"
self.train_set = PneumoniaDataset(SetType.train)
self.train_loader = DataLoader(
PneumoniaDataset(SetType.train), batch_size=16, shuffle=True, num_workers=8
)
self.val_loader = DataLoader(
PneumoniaDataset(SetType.val, shuffle=False),
batch_size=16,
shuffle=False,
num_workers=8,
)
self.test_loader = DataLoader(
PneumoniaDataset(SetType.test, shuffle=False),
batch_size=16,
shuffle=False,
num_workers=8,
)
self.config = {
"pos_weight_bias": 0.5,
"starting_lr": 1e-2,
"momentum": 0.9,
"decay": 5e-4,
"lr_adjustment_factor": 0.3,
"scheduler_patience": 15,
"print_cadence": 100,
"comment": "Added large dense layer.",
"pos_weight": 1341 / 3875, # Number of negatives / positives.
}
self.epochs = epochs
self.device = torch.device("cuda:0")
self.writer = SummaryWriter(comment=self.config["comment"])
self.net = SimpleNet(1).to(self.device)
self.criterion = nn.BCEWithLogitsLoss(
pos_weight=torch.tensor(self.config["pos_weight"])
)
self.optimizer = optim.SGD(
self.net.parameters(),
lr=self.config["starting_lr"], # type: ignore
momentum=self.config["momentum"], # type: ignore
weight_decay=self.config["decay"], # type: ignore
)
self.scheduler = ReduceLROnPlateau(
self.optimizer,
factor=self.config["lr_adjustment_factor"], # type: ignore
mode="max",
verbose=True,
patience=self.config["scheduler_patience"], # type: ignore
)
print("Trainer Initialized.")
for dataset in [self.train_loader, self.test_loader, self.val_loader]:
print(f"Size of set: {len(dataset)}")
def train(self):
training_pass = 0
for epoch in range(self.epochs):
running_loss = 0.0
for i, (inputs, labels, metadata) in enumerate(self.train_loader):
self.net.train()
self.optimizer.zero_grad()
outputs = self.net(inputs.float().to(self.device))
loss = self.criterion(
outputs, labels.unsqueeze(1).float().to(self.device)
)
loss.backward()
self.optimizer.step()
running_loss += loss.item()
if i > 0 and i % self.config["print_cadence"] == 0:
mean_loss = running_loss / self.config["print_cadence"]
print(
f'Epoch: {epoch}\tBatch: {i}\tLoss: {mean_loss}'
)
self.writer.add_scalar(
"Train/RunningLoss",
mean_loss,
training_pass,
)
running_loss = 0.0
training_pass += 1
train_accuracy = self.log_training_metrics(epoch)
self.log_validation_metrics(epoch)
self.scheduler.step(train_accuracy)
accuracy, metrics = self.calculate_accuracy(self.test_loader)
self.writer.add_text("Test/Accuracy", f"{accuracy}")
for key, val in metrics.items():
self.writer.add_text(f"Test/{key}", f"{val}")
self.save_model()
def log_training_metrics(self, epoch: int):
accuracy, metrics = self.calculate_accuracy(self.train_loader)
self.writer.add_scalar(f"Train/Accuracy", accuracy, epoch)
for key, val in metrics.items():
self.writer.add_scalar(f"Train/{key}", val, epoch)
return accuracy
def log_validation_metrics(self, epoch: int):
accuracy, metrics = self.calculate_accuracy(self.val_loader)
self.writer.add_scalar("Validation/Accuracy", accuracy, epoch)
for key, val in metrics.items():
self.writer.add_scalar(f"Validation/{key}", val, epoch)
return accuracy
def calculate_accuracy(self, loader: DataLoader):
truth_list: list = []
pred_list: list = []
with torch.no_grad():
self.net.eval()
correct = 0.0
total = 0.0
for inputs, labels, metadata in loader:
outputs = self.net(inputs.float().to(self.device))
sigmoid = torch.nn.Sigmoid()
preds = sigmoid(outputs)
preds = np.round(preds.detach().cpu().squeeze(1))
pred_list.extend(preds) # type: ignore
truth_list.extend(labels)
total += labels.size(0)
correct += preds.eq(labels.float()).sum().item()
print(f"Correct:\t{correct}, Incorrect:\t{total-correct}")
tn, fp, fn, tp = confusion_matrix(truth_list, pred_list).ravel()
metrics = {
"Recall": tp / (tp + fn),
"Precision": tp / (tp + fp),
"FalseNegativeRate": fn / (tn + fn),
"FalsePositiveRate": fp / (tp + fp),
}
return correct / total, metrics
def save_model(self):
print("saving...")
torch.save(self.net.state_dict(), self.output_model)
def main():
es_staged_data_index = "cifar-metadata-1"
es_logging_index = "custom-net-cifar-12"
output_model = es_logging_index + ".pth"
es = Elasticsearch("localhost:9200")
data = [
doc["_source"] for doc in list(scan(es, index=es_staged_data_index))
]
np.random.seed(42)
np.random.shuffle(data)
training_data = [x for x in data if "train" in x["set_type"]]
testing_data = [x for x in data if "test" in x["set_type"]]
print(f"Size of training set: {len(training_data)}")
print(f"Size of testing set: {len(testing_data)}")
# didnt use this time around.
train_dataset_loader = _get_dataset_loader(training_data,
transform=transform_train,
shuffle=True)
test_dataset_loader = _get_dataset_loader(testing_data)
net = SimpleNet(10).cuda()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(),
lr=1e-3,
momentum=0.9,
weight_decay=5e-4)
# Train
print("training...")
for epoch in range(300):
running_loss = 0.0
for i, (inputs, labels) in enumerate(train_dataset_loader):
optimizer.zero_grad()
outputs = net(inputs.float().cuda())
loss = criterion(outputs, labels.long().cuda())
loss.backward()
optimizer.step()
# print stats
running_loss += loss.item()
print_on = 100
if (i + 1) % print_on == 0:
record = {
"timestamp": datetime.utcnow().isoformat(),
"cross-entropy-loss": running_loss / print_on,
"model-name": "train-simplenet-8"
}
es.index(index=es_logging_index, body=record)
print('[%d, %5d] loss %.3f' % (epoch + 1, i + 1, running_loss /
(print_on + 1)))
running_loss = 0.0
# Test
if epoch + 1 % 10:
print("testing...")
with torch.no_grad():
correct = 0.0
total = 0.0
i = 0.0
for inputs, labels in test_dataset_loader:
outputs = net(inputs.float().cuda())
#_, predicted = torch.max(outputs.data, 1)
_, predicted = outputs.max(1)
total += labels.size(0)
correct += predicted.eq(labels.cuda()).sum().item()
#correct += (predicted == labels.long().cuda()).sum().item()
i += 1
test_accuracy = correct / total
print(f"Test Accuracy: {test_accuracy}")
print(f"Correct: {correct}, Incorrect: {total-correct}")
record = {
"accuracy": test_accuracy,
"correct": correct,
"incorrect": total - correct,
"timestamp": datetime.utcnow().isoformat()
}
es.index(index=es_logging_index, body=record)
# Save
print("saving...")
torch.save(net.state_dict(), output_model)