model.seen += imgs.size(0)
if (epoch + 1) % opt.evaluation_interval == 0 and epoch >= 19:
print("\n---- Evaluating Model ----")
print(opt.weights_folder)
# Evaluate the model on the validation set
if loss.item() < 4:
conf_thres = 0.3
else:
conf_thres = 0.1
loss, precision, recall, AP, f1, ap_class = evaluate(
model,
path=valid_path,
iou_thres=0.5,
conf_thres=conf_thres, # !!! origianl 0.5 current 0.3
nms_thres=0.05, #!!! original 0.5
img_size=640,
batch_size=opt.batch_size,
)
evaluation_metrics = [
("val_precision", precision.mean()),
("val_recall", recall.mean()),
("val_mAP", AP.mean()),
("val_f1", f1.mean()),
]
logger.list_of_scalars_summary(evaluation_metrics, epoch)
# Print class APs and mAP
ap_table = [["Index", "Class name", "AP"]]
for i, c in enumerate(ap_class):
_, output = classifier(features)
loss = criterion(output, clss)
print('Back propagation')
optimizer.zero_grad()
loss.backward()
optimizer.step()
writer.add_scalar('loss', loss.data.cpu().numpy(), iters)
train_info += ' loss: {:.4f}'.format(loss.data.cpu().numpy())
print(train_info)
if epoch % args.val_epoch == 0:
''' evaluate the model '''
acc = evaluate(feature_stractor, classifier, val_loader)
writer.add_scalar('val_acc', acc, iters)
print('Epoch: [{}] ACC:{}'.format(epoch, acc))
''' save best model '''
if acc > best_acc:
save_model(
feature_stractor,
os.path.join(feaStr_path, 'model_best_feaStr.pth.tar'))
save_model(
classifier,
os.path.join(class_path, 'model_best_class.pth.tar'))
best_acc = acc
''' save model '''
save_model(
feature_stractor,
os.path.join(feaStr_path, 'model_{}_feaStr.pth.tar'.format(epoch)))
_, output = classifier(features)
loss = criterion(output, clss)
optimizer.zero_grad()
loss.backward()
optimizer.step()
writer.add_scalar('loss', loss.data.cpu().numpy(), iters)
losses.append(loss)
train_info += ' loss: {:.4f}'.format(loss.data.cpu().numpy())
#print(train_info)
if epoch % args.val_epoch == 0:
''' evaluate the model '''
acc = evaluate(classifier, val_loader)
writer.add_scalar('val_acc', acc, iters)
print('Epoch: [{}] ACC:{}'.format(epoch, acc))
''' save best model '''
if acc > best_acc:
save_model(
classifier,
os.path.join(args.save_dir, 'model_best_class.pth.tar'))
best_acc = acc
''' save model '''
save_model(
classifier,
os.path.join(args.save_dir,
'model_{}_class.pth.tar'.format(epoch)))
plt.plot(range(len(losses)), losses)
model.load_state_dict(torch.load(opt.weights_path))
# Get dataloader
print("\n---- Evaluating Model ----")
opt.iou_thres = 0.5
ap = 0
APs = []
# for opt.iou_thres in np.arange(0.5,0.96,0.05):
for opt.iou_thres in np.arange(0.5, 0.54, 0.05):
loss, precision, recall, AP, f1, ap_class = evaluate(
model,
path=opt.valid_path,
iou_thres=opt.iou_thres, # !!! origianl 0.50.5,
conf_thres=opt.conf_thres, # !!! origianl 0.5 current 0.3
nms_thres=opt.nms_thres, #!!! original 0.5 0.001
img_size=opt.img_size,
batch_size=opt.batch_size,
)
#print(AP.mean())
ap += AP.mean()
APs.append([opt.iou_thres, AP.mean()])
evaluation_metrics = [
("val_precision", round(precision.mean(), 4)),
("val_recall", round(recall.mean(), 4)),
("val_mAP", round(AP.mean(), 4)),
("val_f1", round(f1.mean(), 4)),
]
#cls = torch.from_numpy(np.expand_dims(np.asarray(cls), axis=0))
#cls = cls.cuda()
loss = criterion(pred, batch_cls)
#print('Back propagation')
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_info += ' loss: {:.4f}'.format(loss.data.cpu().numpy())
print(train_info) #, end = '\r')
if epoch % args.val_epoch == 0:
''' evaluate the model '''
acc = evaluate(feature_stractor, rnn, val_loader, args.train_batch)
writer.add_scalar('val_acc', acc, iters)
print('Epoch: [{}] ACC:{}'.format(epoch, acc))
''' save best model '''
if acc > best_acc:
print('saved model', epoch)
save_model(
rnn, os.path.join(args.save_dir, 'model_best_rnn.pth.tar'))
save_model(
feature_stractor,
os.path.join(args.save_dir, 'model_best_fea.pth.tar'))
best_acc = acc
''' save model '''
save_model(
rnn,
os.path.join(args.save_dir, 'model_{}_rnn.pth.tar'.format(epoch)))