def test(model, device, test_loader, loss_function):
'''
Evaluates DTP
args:
model: DTP as defined in the DynamicTrajectoryPredictor class
device: GPU or CPU
test_loader: Dataloader to produce stacks of optical flow images
loss_function: eg. MSE
returns:
MSE and FDE at intervals of 5,10,15 frames into the future
Outputs and targets 15 frames into the future
'''
model.eval()
test_loss = 0
all_outputs_5 = np.array([])
all_targets_5 = np.array([])
all_outputs_10 = np.array([])
all_targets_10 = np.array([])
all_outputs_15 = np.array([])
all_targets_15 = np.array([])
with torch.no_grad():
for batch_idx, data in enumerate(test_loader):
flow, targets = data['flow_stack'].to(device), data['labels'].to(
device)
flow = flow.float()
targets = targets.float()
output = model(flow)
test_loss += loss_function(output,
targets).item() # sum up batch loss
output_5 = torch.cat((output[:, 0:5], output[:, 15:20]), dim=1)
output_10 = torch.cat((output[:, 0:10], output[:, 15:25]), dim=1)
output_15 = output
targets_5 = torch.cat((targets[:, 0:5], targets[:, 15:20]), dim=1)
targets_10 = torch.cat((targets[:, 0:10], targets[:, 15:25]),
dim=1)
targets_15 = targets
all_outputs_5 = np.append(all_outputs_5,
output_5.detach().cpu().numpy())
all_targets_5 = np.append(all_targets_5,
targets_5.detach().cpu().numpy())
all_outputs_10 = np.append(all_outputs_10,
output_10.detach().cpu().numpy())
all_targets_10 = np.append(all_targets_10,
targets_10.detach().cpu().numpy())
all_outputs_15 = np.append(all_outputs_15,
output_15.detach().cpu().numpy())
all_targets_15 = np.append(all_targets_15,
targets_15.detach().cpu().numpy())
MSE_5 = utils.calc_mse(all_outputs_5, all_targets_5)
FDE_5 = utils.calc_fde(all_outputs_5, all_targets_5, n=5)
MSE_10 = utils.calc_mse(all_outputs_10, all_targets_10)
FDE_10 = utils.calc_fde(all_outputs_10, all_targets_10, n=10)
MSE_15 = utils.calc_mse(all_outputs_15, all_targets_15)
FDE_15 = utils.calc_fde(all_outputs_15, all_targets_15, n=15)
print(
'Validation: \t\t\t\tMSE@5: {:.0f} \tFDE@5: {:.0f} \tMSE@10: {:.0f} \tFDE@10: {:.0f} \tMSE@15: {:.0f} \tFDE@15: {:.0f}'
.format(MSE_5, FDE_5, MSE_10, FDE_10, MSE_15, FDE_15))
return MSE_5, FDE_5, MSE_10, FDE_10, MSE_15, FDE_15, all_outputs_15, all_targets_15
def train(model, device, train_loader, optimizer, epoch, loss_function):
'''
Trains DTP
args:
model: DTP as defined in the DynamicTrajectoryPredictor class
device: GPU or CPU
train_loader: Dataloader to produce stacks of optical flow images
optimizer: eg. ADAM
epoch: Current epoch (for printing progress)
loss_function: eg. MSE
'''
model.train()
all_outputs_5 = np.array([])
all_targets_5 = np.array([])
all_outputs_10 = np.array([])
all_targets_10 = np.array([])
all_outputs_15 = np.array([])
all_targets_15 = np.array([])
for batch_idx, data in enumerate(train_loader):
if batch_idx % 50 == 0:
print('Batch ', batch_idx, ' of ', len(train_loader))
flow, targets = data['flow_stack'].to(device), data['labels'].to(
device)
targets = targets.float()
optimizer.zero_grad()
flow = flow.float()
output = model(flow)
loss = loss_function(output, targets)
loss.backward()
optimizer.step()
output_5 = torch.cat((output[:, 0:5], output[:, 15:20]), dim=1)
output_10 = torch.cat((output[:, 0:10], output[:, 15:25]), dim=1)
output_15 = output
targets_5 = torch.cat((targets[:, 0:5], targets[:, 15:20]), dim=1)
targets_10 = torch.cat((targets[:, 0:10], targets[:, 15:25]), dim=1)
targets_15 = targets
all_outputs_5 = np.append(all_outputs_5,
output_5.detach().cpu().numpy())
all_targets_5 = np.append(all_targets_5,
targets_5.detach().cpu().numpy())
all_outputs_10 = np.append(all_outputs_10,
output_10.detach().cpu().numpy())
all_targets_10 = np.append(all_targets_10,
targets_10.detach().cpu().numpy())
all_outputs_15 = np.append(all_outputs_15,
output_15.detach().cpu().numpy())
all_targets_15 = np.append(all_targets_15,
targets_15.detach().cpu().numpy())
MSE_5 = utils.calc_mse(all_outputs_5, all_targets_5)
FDE_5 = utils.calc_fde(all_outputs_5, all_targets_5, n=5)
MSE_10 = utils.calc_mse(all_outputs_10, all_targets_10)
FDE_10 = utils.calc_fde(all_outputs_10, all_targets_10, n=10)
MSE_15 = utils.calc_mse(all_outputs_15, all_targets_15)
FDE_15 = utils.calc_fde(all_outputs_15, all_targets_15, n=15)
print(
'Train Epoch: {} [{}/{} ({:.0f}%)] \tMSE@5: {:.0f} \tFDE@5: {:.0f} \tMSE@10: {:.0f} \tFDE@10: {:.0f} \tMSE@15: {:.0f} \tFDE@15: {:.0f}'
.format(epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), MSE_5, FDE_5, MSE_10,
FDE_10, MSE_15, FDE_15))