def sdn_test_early_exits(model, loader, device='cpu'):
model.eval()
early_output_counts = [0] * model.num_output
non_conf_output_counts = [0] * model.num_output
top1 = data.AverageMeter()
top5 = data.AverageMeter()
total_time = 0
with torch.no_grad():
for batch in loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
start_time = time.time()
output, output_id, is_early = model(b_x)
end_time = time.time()
total_time+= (end_time - start_time)
if is_early:
early_output_counts[output_id] += 1
else:
non_conf_output_counts[output_id] += 1
prec1, prec5 = data.accuracy(output, b_y, topk=(1, 5))
top1.update(prec1[0], b_x.size(0))
top5.update(prec5[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top5_acc = top5.avg.data.cpu().numpy()[()]
return top1_acc, top5_acc, early_output_counts, non_conf_output_counts, total_time
def sdn_test(model, loader, device='cpu'):
model.eval()
top1 = []
top5 = []
for output_id in range(model.num_output):
t1 = data.AverageMeter()
t5 = data.AverageMeter()
top1.append(t1)
top5.append(t5)
with torch.no_grad():
for batch in loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
output = model(b_x)
for output_id in range(model.num_output):
cur_output = output[output_id]
prec1, prec5 = data.accuracy(cur_output, b_y, topk=(1, 5))
top1[output_id].update(prec1[0], b_x.size(0))
top5[output_id].update(prec5[0], b_x.size(0))
top1_accs = []
top5_accs = []
for output_id in range(model.num_output):
top1_accs.append(top1[output_id].avg.data.cpu().numpy()[()])
top5_accs.append(top5[output_id].avg.data.cpu().numpy()[()])
return top1_accs, top5_accs
def cnn_test(model, loader, device='cpu'):
model.eval()
top1 = data.AverageMeter()
top5 = data.AverageMeter()
with torch.no_grad():
for batch in loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
output = model(b_x)
prec1, prec5 = data.accuracy(output, b_y, topk=(1, 5))
top1.update(prec1[0], b_x.size(0))
top5.update(prec5[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top5_acc = top5.avg.data.cpu().numpy()[()]
return top1_acc, top5_acc
def cnn_test_time(model, loader, device='cpu'):
model.eval()
top1 = data.AverageMeter()
top3 = data.AverageMeter()
total_time = 0
with torch.no_grad():
for batch in loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
start_time = time.time()
output = model(b_x)
end_time = time.time()
total_time += (end_time - start_time)
prec1, prec3 = data.accuracy(output, b_y, topk=(1, 3))
top1.update(prec1[0], b_x.size(0))
top3.update(prec3[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top3_acc = top3.avg.data.cpu().numpy()[()]
return top1_acc, top3_acc, total_time
def cnn_test(model, loader, device='cpu'):
model.eval()
top1 = data.AverageMeter()
top3 = data.AverageMeter()
with torch.no_grad():
for batch in loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
output = model(b_x)
if isinstance(output, list):
output = output[0]
prec1, prec3 = data.accuracy(output, b_y, topk=(1, 3))
top1.update(prec1[0], b_x.size(0))
top3.update(prec3[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top3_acc = top3.avg.data.cpu().numpy()[()]
return top1_acc, top3_acc