def main():
# prepare input data
trf = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
inputs = trf(Image.open('./data/dog.jpg')).unsqueeze(0)
# load existing model from pytorchcv lib (pretrain=False)
net = ptcv_get_model('mobilenetv2_w1', pretrained=False)
net.eval()
output = net(inputs)
# print("original model:",net)
# fuse the Batch Normalization
net1 = fuse_bn_recursively(copy.deepcopy(net))
net1.eval()
output1 = net1(inputs)
# print("BN fusion model:",net1)
# compare the output
print("output of original model:",output.size())
print("=> min : %.4f, mean : %.4f max : %.4f"%(output.min().detach().numpy(),output.mean().detach().numpy(),output.max().detach().numpy()))
print("output of BNfusion model:",output1.size())
print("=> min : %.4f, mean : %.4f max : %.4f"%(output1.min().detach().numpy(),output1.mean().detach().numpy(),output1.max().detach().numpy()))
# transform to ONNX format for visualization
dummy_input = Variable(torch.randn([1, 3, 224, 224]))
torch.onnx.export(net, dummy_input, "./data/mobilenet_v2.onnx")
torch.onnx.export(net1, dummy_input, "./data/mobilenet_v2_nobn.onnx")
def infer(test_image_data_path, test_meta_data_path):
# DONOTCHANGE This Line
test_meta_data = pd.read_csv(test_meta_data_path, delimiter=',', header=0)
input_size=224 # you can change this according to your model.
batch_size=_BATCH_SIZE # you can change this. But when you use 'nsml submit --test' for test infer, there are only 200 number of data.
# test time gpu memory error, so i reduce batch size when test time
device = 0
dataloader = DataLoader(
AIRushDataset(test_image_data_path, test_meta_data, label=None,
transform=transforms.Compose([ObjROI()
,transforms.Resize((input_size, input_size))
, transforms.ToTensor()
#,transforms.Normalize(mean=mean_v, std=std_v)
])),
batch_size=batch_size,
shuffle=False,
num_workers=3,
pin_memory=True)
model_nsml.to(device)
model_nsml = fuse_bn_recursively(model_nsml)
model_nsml.eval()
predict_list = []
for batch_idx, image in enumerate(dataloader):
image = image.to(device)
output = model_nsml(image).double()
output_prob = F.softmax(output, dim=1)
predict = np.argmax(to_np(output_prob), axis=1)
predict_list.append(predict)
predict_vector = np.concatenate(predict_list, axis=0)
return predict_vector # this return type should be a numpy array which has shape of (138343, 1)
net = convert_resnet_family(net, se)
# Benchmarking
# First, we run the network the way it is
net.eval()
with torch.no_grad():
F.softmax(net(img), 1)
# Measuring non-optimized model performance
times = []
for i in range(50):
start = time.time()
with torch.no_grad():
res_0 = F.softmax(net(img), 1)
times.append(time.time() - start)
print('Non fused takes', np.mean(times), 'seconds')
net = fuse_bn_recursively(net)
net.eval()
times = []
for i in range(50):
start = time.time()
with torch.no_grad():
res_1 = F.softmax(net(img), 1)
times.append(time.time() - start)
print('Fused takes', np.mean(times), 'seconds')
diff = res_0 - res_1
print('L2 Norm of the element-wise difference:', diff.norm().item())
modelA = EfficientNet.from_name('efficientnet-b4', override_params={'num_classes': args.output_size})
modelB = make_model('se_resnext50_32x4d', num_classes=args.output_size, pretrained=False, pool=nn.AdaptiveAvgPool2d(1))
modelC = make_model('nasnetamobile', num_classes=args.output_size, pretrained=False, pool=nn.AdaptiveAvgPool2d(1))
# DONOTCHANGE: They are reserved for nsml
bind_model(modelA)
re_train_info = use_ensemble_model_session[0]#'efficientnet-b4'
nsml.load(checkpoint=re_train_info['checkpoint'], session=re_train_info['session'])
bind_model(modelB)
re_train_info = use_ensemble_model_session[1]#'se_resnext50_32x4d'
nsml.load(checkpoint=re_train_info['checkpoint'], session=re_train_info['session'])
bind_model(modelC)
re_train_info = use_ensemble_model_session[2]#'nasnetamobile'
nsml.load(checkpoint=re_train_info['checkpoint'], session=re_train_info['session'])
model = MyEnsembleTTA(modelA,modelB,modelC)
model = fuse_bn_recursively(model)
model = model.to(device) #use gpu
#summary(model, (3,args.input_size,args.input_size))
# DONOTCHANGE: They are reserved for nsml
bind_model(model)
#nsml.load(checkpoint='3', session='team_27/airush1/392 ')
#nsml.save('dontgiveup')
if args.pause:
nsml.paused(scope=locals())
if args.mode == "train":
# Warning: Do not load data before this line
dataloader, valid_dataloader = train_dataloader(args.input_size, args.batch_size*10, args.num_workers, test_bs = False
, br_multi_oh=True#)
,print_nor_info = False,use_last_fine_tune=True)