def forward(self, batch):
data_time = time.time() - self._last_time
output = self.model(batch[0])
# print(batch[1])
# loss = self.criterion(output, batch[1])
# import ipdb;ipdb.set_trace()
loss_cat = torch.nn.BCEWithLogitsLoss()(output[0][0], batch[1][0][:26])
loss_dim = torch.nn.MSELoss()(output[1][0], batch[1][0][26:29])
loss = loss_cat + loss_dim / 10
if self.multi_class:
mAP = self.calculate_mAP(output, batch[1])
self._preverse_for_show = [loss.detach(), data_time, mAP]
else:
prec1, prec5 = accuracy(output,
batch[1],
topk=(1,
min(5,
self.config.dataset.num_class)))
self._preverse_for_show = [
loss.detach(), data_time,
prec1.detach(),
prec5.detach()
]
return loss
def forward(self, batch):
data_time = time.time() - self._last_time
output = self.model(batch[0])
loss = self.criterion(output, batch[1])
if self.multi_class:
mAP = self.calculate_mAP(output, batch[1])
self._preverse_for_show = [loss.detach(), data_time, mAP]
else:
prec1, prec5 = accuracy(output, batch[1], topk=(1, min(5, self.config.dataset.num_class)))
self._preverse_for_show = [loss.detach(), data_time, prec1.detach(), prec5.detach()]
return loss
def evaluate(self):
batch_time = AverageMeter(0)
losses = AverageMeter(0)
if self.multi_class:
mAPs = AverageMeter(0)
else:
top1 = AverageMeter(0)
top5 = AverageMeter(0)
if self.inference_only:
spatial_crops = self.config.get('evaluate', {}).get('spatial_crops', 1)
temporal_samples = self.config.get('evaluate', {}).get('temporal_samples', 1)
else:
spatial_crops = 1
temporal_samples = 1
dup_samples = spatial_crops * temporal_samples
self.model.cuda().eval()
test_loader = self.data_loaders['test']
test_len = len(test_loader)
end = time.time()
for iter_idx in range(test_len):
inputs = self.get_batch('test')
isizes = inputs[0].shape
if self.config.net.model_type == '2D':
inputs[0] = inputs[0].view(
isizes[0] * dup_samples, -1, isizes[2], isizes[3])
else:
inputs[0] = inputs[0].view(
isizes[0], isizes[1], dup_samples, -1, isizes[3], isizes[4]
)
inputs[0] = inputs[0].permute(0, 2, 1, 3, 4, 5).contiguous()
inputs[0] = inputs[0].view(isizes[0] * dup_samples, isizes[1], -1, isizes[3], isizes[4])
output = self.model(inputs[0])
osizes = output.shape
output = output.view((osizes[0] // dup_samples, -1, osizes[1]))
output = torch.mean(output, 1)
loss = self.criterion(output, inputs[1])
num = inputs[0].size(0)
losses.update(loss.item(), num)
if self.multi_class:
mAP = self.calculate_mAP(output, inputs[1])
mAPs.update(mAP, num)
else:
prec1, prec5 = accuracy(output, inputs[1], topk=(1, min(5, self.config.dataset.num_class)))
top1.update(prec1.item(), num)
top5.update(prec5.item(), num)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if iter_idx % self.config.trainer.print_freq == 0:
self.logger.info('Test: [{0}/{1}]\tTime {batch_time.val:.3f} ({batch_time.avg:.3f})'.format(
iter_idx, test_len, batch_time=batch_time))
total_num = torch.Tensor([losses.count]).cuda()
loss_sum = torch.Tensor([losses.avg*losses.count]).cuda()
if self.config.gpus > 1:
all_reduce(total_num, False)
all_reduce(loss_sum, False)
final_loss = loss_sum.item()/total_num.item()
if self.multi_class:
mAP_sum = torch.Tensor([mAPs.avg*mAPs.count]).cuda()
if self.config.gpus > 1:
all_reduce(mAP_sum)
final_mAP = mAP_sum.item()/total_num.item()
self.logger.info(' * mAP {:.3f}\tLoss {:.3f}\ttotal_num={}'.format(final_mAP, final_loss,
total_num.item()))
metric = Top1Metric(final_mAP, 0, final_loss)
else:
top1_sum = torch.Tensor([top1.avg*top1.count]).cuda()
top5_sum = torch.Tensor([top5.avg*top5.count]).cuda()
if self.config.gpus > 1:
all_reduce(top1_sum, False)
all_reduce(top5_sum, False)
final_top1 = top1_sum.item()/total_num.item()
final_top5 = top5_sum.item()/total_num.item()
self.logger.info(' * Prec@1 {:.3f}\tPrec@5 {:.3f}\tLoss {:.3f}\ttotal_num={}'.format(final_top1,
final_top5, final_loss, total_num.item()))
metric = Top1Metric(final_top1, final_top5, final_loss)
self.model.cuda().train()
return metric
def evaluate(self):
global predict_mode
output_data = []
batch_time = AverageMeter(0)
losses = AverageMeter(0)
if self.multi_class:
mAPs = AverageMeter(0)
else:
top1 = AverageMeter(0)
top5 = AverageMeter(0)
if self.inference_only:
spatial_crops = self.config.get('evaluate',
{}).get('spatial_crops', 1)
temporal_samples = self.config.get('evaluate',
{}).get('temporal_samples', 1)
else:
spatial_crops = 1
temporal_samples = 1
dup_samples = spatial_crops * temporal_samples
self.model.cuda().eval()
test_loader = self.data_loaders['test']
test_len = len(test_loader)
end = time.time()
for iter_idx in range(test_len):
inputs = self.get_batch('test')
isizes = inputs[0].shape
if self.config.net.model_type == '2D':
inputs[0] = inputs[0].view(isizes[0] * dup_samples, -1,
isizes[2], isizes[3])
else:
inputs[0] = inputs[0].view(isizes[0], isizes[1], dup_samples,
-1, isizes[3], isizes[4])
inputs[0] = inputs[0].permute(0, 2, 1, 3, 4, 5).contiguous()
inputs[0] = inputs[0].view(isizes[0] * dup_samples, isizes[1],
-1, isizes[3], isizes[4])
output = self.model(inputs[0])
osizes = output[0].shape
output[0] = output[0].view(
(osizes[0] // dup_samples, -1, osizes[1]))
output[0] = torch.mean(output[0], 1)
osizes = output[1].shape
output[1] = output[1].view(
(osizes[0] // dup_samples, -1, osizes[1]))
output[1] = torch.mean(output[1], 1)
# print(output)
# import ipdb;ipdb.set_trace()
# for __, _ in zip(output[0].tolist(), output[1].tolist()):
# out_data.append(','.join([str(_) for _ in __+___]))
out_data = [
','.join([str(_) for _ in __ + ___])
for __, ___ in zip(output[0].tolist(), output[1].tolist())
]
# print(out_data)
for i in range(len(out_data)):
output_data.append(out_data[i])
# loss = self.criterion(output, inputs[1])
loss_cat = torch.nn.BCEWithLogitsLoss()(output[0][0],
inputs[1][0][:26])
loss_dim = torch.nn.MSELoss()(output[1][0], inputs[1][0][26:29])
loss = loss_cat + loss_dim / 10
num = inputs[0].size(0)
losses.update(loss.item(), num)
if self.multi_class:
mAP = self.calculate_mAP(output, inputs[1])
mAPs.update(mAP, num)
else:
prec1, prec5 = accuracy(
output,
inputs[1],
topk=(1, min(5, self.config.dataset.num_class)))
top1.update(prec1.item(), num)
top5.update(prec5.item(), num)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if iter_idx % self.config.trainer.print_freq == 0:
self.logger.info(
'Test: [{0}/{1}]\tTime {batch_time.val:.3f} ({batch_time.avg:.3f})'
.format(iter_idx, test_len, batch_time=batch_time))
total_num = torch.Tensor([losses.count]).cuda()
loss_sum = torch.Tensor([losses.avg * losses.count]).cuda()
if self.config.gpus > 1:
all_reduce(total_num, False)
all_reduce(loss_sum, False)
final_loss = loss_sum.item() / total_num.item()
if self.multi_class:
mAP_sum = torch.Tensor([mAPs.avg * mAPs.count]).cuda()
if self.config.gpus > 1:
all_reduce(mAP_sum)
final_mAP = mAP_sum.item() / total_num.item()
self.logger.info(' * mAP {:.3f}\tLoss {:.3f}\ttotal_num={}'.format(
final_mAP, final_loss, total_num.item()))
metric = Top1Metric(final_mAP, 0, final_loss)
else:
top1_sum = torch.Tensor([top1.avg * top1.count]).cuda()
top5_sum = torch.Tensor([top5.avg * top5.count]).cuda()
if self.config.gpus > 1:
all_reduce(top1_sum, False)
all_reduce(top5_sum, False)
final_top1 = top1_sum.item() / total_num.item()
final_top5 = top5_sum.item() / total_num.item()
self.logger.info(
' * Prec@1 {:.3f}\tPrec@5 {:.3f}\tLoss {:.3f}\ttotal_num={}'.
format(final_top1, final_top5, final_loss, total_num.item()))
metric = Top1Metric(final_top1, final_top5, final_loss)
self.model.cuda().train()
if not predict_mode:
return metric
else:
return metric, output_data