def GetEnsembleResult():
# Priors
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
# Dataset
Dataset = BDXJTU2019_test(root='data', TEST_IMAGE_DIR='test')
Dataloader = data.DataLoader(Dataset,
1,
num_workers=1,
shuffle=False,
pin_memory=True)
# Network
cudnn.benchmark = True
net1 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net1.load_state_dict(
torch.load('./weights/best_models/se_resnext50_32x4d_SGD_1_20.pth'))
net1.eval()
net2 = MultiModalNet('multiscale_se_resnext_HR', 'DPN26', 0.5)
net2.load_state_dict(
torch.load(
'./weights/best_models/multiscale_se_resnext_HR_SGD_16.pth'))
net2.eval()
# filename = './submission/MM_epoch26_25_all_pretrained_2HR_616v2.txt'
submit_file = './submission/submission.txt'
f = open(submit_file, 'w+')
for (image_tensor, visit_tensor, anos) in tqdm(Dataloader):
Tensor_1 = net1.forward(image_tensor.cuda(), visit_tensor.cuda())
Tensor_HR = net2.forward(image_tensor.cuda(), visit_tensor.cuda())
preds = torch.nn.functional.normalize(Tensor_1) \
+ torch.nn.functional.normalize(Tensor_HR)
_, pred = preds.data.topk(1, 1, True, True)
#f.write(anos[0] + ',' + CLASSES[4] + '\r\n')
# print('{}\t{}'.format(anos[0][:-4],CLASSES[pred[0][0]]))
f.writelines('{}\t{}\n'.format(anos[0][:-4], CLASSES[pred[0][0]]))
f.close()
def GeResult():
# Priors
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
# Dataset
Dataset = BDXJTU2019_test(root='data')
Dataloader = data.DataLoader(Dataset,
1,
num_workers=1,
shuffle=False,
pin_memory=True)
# Network
cudnn.benchmark = True
#Network = pnasnet5large(6, None)
#Network = ResNeXt101_64x4d(6)
net = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_100/BDXJTU2019_SGD_20.pth'
))
net.eval()
filename = 'MM_epoch20_R_TTA.txt'
f = open(filename, 'w')
for (Input_O, Input_H, visit_tensor, anos) in Dataloader:
ConfTensor_O = net.forward(Input_O.cuda(), visit_tensor.cuda())
ConfTensor_H = net.forward(Input_H.cuda(), visit_tensor.cuda())
#ConfTensor_V = net.forward(Input_V.cuda())
preds = torch.nn.functional.normalize(
ConfTensor_O) + torch.nn.functional.normalize(
ConfTensor_H) #+torch.nn.functional.normalize(ConfTensor_V)
_, pred = preds.data.topk(1, 1, True, True)
#f.write(anos[0] + ',' + CLASSES[4] + '\r\n')
print(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.writelines(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.close()
def GeResult():
# Dataset
Dataset_val = MM_BDXJTU2019(root='/home/dell/Desktop/2019BaiduXJTU/data',
mode='1_val')
Dataloader_val = data.DataLoader(Dataset_val,
batch_size=1,
num_workers=4,
shuffle=True,
pin_memory=True)
class_names = [
'001', '002', '003', '004', '005', '006', '007', '008', '009'
]
# construct network
epoch = 12
net = MultiModalNet('se_resnet152', 'DPN26', 0.5)
# if torch.cuda.device_count() > 1:
# print("Let's use", torch.cuda.device_count(), "GPUs!")
# dim = 0 [30, xxx] -> [10, ...], [10, ...], [10, ...] on 3 GPUs
# net = nn.DataParallel(net)
net.to(device)
# net.load_state_dict(torch.load('/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet152_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_' + str(epoch) + '.pth'))
net.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet152_se_resnext50_32x4d_resample_pretrained_80w_1/inception_005.pth'
))
print('load ' + str(epoch) + ' epoch model')
net.eval()
results = []
results_anno = []
for i, (Input_img, Input_vis, Anno) in enumerate(Dataloader_val):
Input_img = Input_img.to(device)
Input_vis = Input_vis.to(device)
ConfTensor = net.forward(Input_img, Input_vis)
_, pred = ConfTensor.data.topk(1, 1, True, False)
results.append(pred.item())
results_anno.append(Anno) #append annotation results
if ((i + 1) % 1000 == 0):
print(i + 1)
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
# print accuracy of different input
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
cnf_matrix = confusion_matrix(results_anno, results)
cnf_tr = np.trace(cnf_matrix)
cnf_tr = cnf_tr.astype('float')
print(cnf_tr / len(Dataset_val))
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
title='Confusion matrix, without normalization')
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
normalize=True,
title='Normalized confusion matrix')
plt.show()
def GeResult():
# Dataset
Dataset = BDXJTU2019_test(root='/home/dell/Desktop/2019BaiduXJTU/data')
Dataloader = data.DataLoader(Dataset,
1,
num_workers=1,
shuffle=False,
pin_memory=True)
net1 = MultiModalNet1('se_resnet50', 'DPN26', 0.5)
net1.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet50_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_9.pth'
))
net1.to(device)
net1.eval()
net2 = MultiModalNet('se_resnet152', 'DPN26', 0.5)
net2.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet152_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_6.pth'
))
net2.to(device)
net2.eval()
net3 = MultiModalNet2('densenet201', 'DPN26', 0.5)
net3.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/densenet201_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_3.pth'
))
net3.to(device)
net3.eval()
net4 = MultiModalNet2('densenet201', 'DPN26', 0.5)
net4.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/densenet201_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_10.pth'
))
net4.to(device)
net4.eval()
net5 = MultiModalNet1('multiscale_se_resnext', 'DPN26', 0.5)
net5.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/multiscale_se_resnext_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_11.pth'
))
net5.to(device)
net5.eval()
net6 = MultiModalNet1('multiscale_resnet', 'DPN26', 0.5)
net6.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/multiscale_resnet_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_10.pth'
))
net6.to(device)
net6.eval()
net7 = MultiModalNet2('densenet201', 'DPN26', 0.5)
net7.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/densenet201_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_4.pth'
))
net7.to(device)
net7.eval()
#Network = pnasnet5large(6, None)
#Network = ResNeXt101_64x4d(6)
# net1 =MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net1.load_state_dict(torch.load('/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained/BDXJTU2019_SGD_16.pth'))
# net1.eval()
# net2 = MultiModalNet('multiscale_se_resnext_HR', 'DPN26', 0.5)
# net2.load_state_dict(torch.load('/home/zxw/2019BaiduXJTU/weights/MultiModal_50_MS_resample_pretrained_HR/BDXJTU2019_SGD_26.pth'))
# net2.eval()
# net3 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net3.load_state_dict(torch.load('/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained_w/BDXJTU2019_SGD_50.pth'))
# net3.eval()
# net4 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net4.load_state_dict(torch.load('/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained_1/BDXJTU2019_SGD_80.pth'))
# net4.eval()
filename = 'MM_ensemble4_TTA.txt'
f = open(filename, 'w')
for (Input_O, Input_H, visit_tensor, anos) in Dataloader:
ConfTensor_O = net1.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_H = net2.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_V = net3.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_V0 = net3.forward(Input_H.to(device),
visit_tensor.to(device))
ConfTensor_1 = net4.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_10 = net4.forward(Input_H.to(device),
visit_tensor.to(device))
ConfTensor_2 = net5.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_20 = net5.forward(Input_H.to(device),
visit_tensor.to(device))
ConfTensor_3 = net6.forward(Input_O.to(device),
visit_tensor.to(device))
ConfTensor_4 = net7.forward(Input_O.to(device),
visit_tensor.to(device))
preds = torch.nn.functional.normalize(
ConfTensor_O) + torch.nn.functional.normalize(
ConfTensor_H) + 2 * torch.nn.functional.normalize(
ConfTensor_V) + torch.nn.functional.normalize(
ConfTensor_V0) + torch.nn.functional.normalize(
ConfTensor_1
) + torch.nn.functional.normalize(
ConfTensor_10
) + 2 * torch.nn.functional.normalize(
ConfTensor_2) + torch.nn.functional.normalize(
ConfTensor_20
) + torch.nn.functional.normalize(
ConfTensor_3
) + 2 * torch.nn.functional.normalize(ConfTensor_4)
_, pred = preds.data.topk(1, 1, True, True)
#f.write(anos[0] + ',' + CLASSES[4] + '\r\n')
print(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.writelines(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.close()
def GeResult():
# Priors
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
# Dataset
Dataset = BDXJTU2019_test(root='data')
Dataloader = data.DataLoader(Dataset,
1,
num_workers=1,
shuffle=False,
pin_memory=True)
# Network
cudnn.benchmark = True
#Network = pnasnet5large(6, None)
#Network = ResNeXt101_64x4d(6)
net1 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net1.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained/BDXJTU2019_SGD_16.pth'
))
net1.eval()
net2 = MultiModalNet('multiscale_se_resnext_HR', 'DPN26', 0.5)
net2.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_50_MS_resample_pretrained_HR/BDXJTU2019_SGD_26.pth'
))
net2.eval()
net3 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net3.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained_w/BDXJTU2019_SGD_50.pth'
))
net3.eval()
net4 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net4.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained_1/BDXJTU2019_SGD_80.pth'
))
net4.eval()
filename = 'MM_epoch26_25_all_pretrained_2HR_616v2.txt'
f = open(filename, 'w')
for (Input_O, Input_H, visit_tensor, anos) in Dataloader:
ConfTensor_O = net1.forward(Input_O.cuda(), visit_tensor.cuda())
ConfTensor_H = net2.forward(Input_O.cuda(), visit_tensor.cuda())
ConfTensor_V = net3.forward(Input_O.cuda(), visit_tensor.cuda())
ConfTensor_1 = net4.forward(Input_O.cuda(), visit_tensor.cuda())
preds = torch.nn.functional.normalize(
ConfTensor_O) + torch.nn.functional.normalize(
ConfTensor_H) + torch.nn.functional.normalize(
ConfTensor_V) + torch.nn.functional.normalize(ConfTensor_1)
_, pred = preds.data.topk(1, 1, True, True)
#f.write(anos[0] + ',' + CLASSES[4] + '\r\n')
print(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.writelines(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.close()
def GeResult():
# Priors
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
# Dataset
Dataset = BDXJTU2019_test(root='../final')
batch_size = 32
print(batch_size)
Dataloader = data.DataLoader(Dataset,
batch_size,
num_workers=1,
shuffle=False,
pin_memory=True)
# Network
cudnn.benchmark = True
# Network = pnasnet5large(6, None)
# Network = ResNeXt101_64x4d(6)
net = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
print("weights model2")
net.load_state_dict(torch.load('model2/BDXJTU2019_SGD_18_28000.pth'))
net.eval()
filename = '[2_18].txt'
import numpy as np
f = open(filename, 'w')
csv6_O = open("csvO[2_18].csv", "w")
csv6_H = open("csvH[2_18].csv", "w")
for (Input_O, Input_H, visit_tensor, anoss) in Dataloader:
ConfTensor_Os = net.forward(Input_O.cuda(), visit_tensor.cuda())
ConfTensor_Hs = net.forward(Input_H.cuda(), visit_tensor.cuda())
# ConfTensor_V = net.forward(Input_V.cuda())
for id in range(ConfTensor_Os.shape[0]):
ConfTensor_O = ConfTensor_Os[id].reshape([1, 9])
ConfTensor_H = ConfTensor_Hs[id].reshape([1, 9])
anos = [anoss[id]]
preds_temp = torch.nn.functional.normalize(ConfTensor_O)
string = ""
for _ in range(9):
string = string + str(float(preds_temp[0][_])) + ","
string = string + "\n"
csv6_O.write(string)
####################
####################
preds_temp = torch.nn.functional.normalize(ConfTensor_H)
string = ""
for _ in range(9):
string = string + str(float(preds_temp[0][_])) + ","
string = string + "\n"
csv6_H.write(string)
####################
preds = torch.nn.functional.normalize(
ConfTensor_O) + torch.nn.functional.normalize(
ConfTensor_H
) # +torch.nn.functional.normalize(ConfTensor_V)
_, pred = preds.data.topk(1, 1, True, True)
# f.write(anos[0] + ',' + CLASSES[4] + '\r\n')
#print(preds[0],"a12343",preds[0].max())
# cls = pred[0][0]
# if preds[0].max() < 0.9:
# rnd = np.random.randint(1, 100)
# if rnd < 50:
# cls = 0
# elif rnd < 80:
# cls = 1
# elif rnd < 100:
# cls = 5
# print(rnd, "True", anos[0][:-4])
# print(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
f.writelines(anos[0][:-4] + '\t' + CLASSES[pred[0][0]] + '\n')
if int(anos[0][:-4]) % 500 == 0:
print(anos[0][:-4])
csv6_O.close()
csv6_H.close()
f.close()
def GeResult():
# Dataset
Dataset_val = MM_BDXJTU2019(root='/home/dell/Desktop/2019BaiduXJTU/data',
mode='val')
Dataloader_val = data.DataLoader(Dataset_val,
batch_size=1,
num_workers=2,
shuffle=True,
pin_memory=True)
class_names = [
'001', '002', '003', '004', '005', '006', '007', '008', '009'
]
net1 = MultiModalNet1('se_resnet50', 'DPN26', 0.5)
net1.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet50_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_9.pth'
))
net1.to(device)
net1.eval()
net2 = MultiModalNet('se_resnet152', 'DPN26', 0.5)
net2.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/se_resnet152_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_4.pth'
))
net2.to(device)
net2.eval()
net3 = MultiModalNet2('densenet201', 'DPN26', 0.5)
net3.load_state_dict(
torch.load(
'/home/dell/Desktop/2019BaiduXJTU/weights/densenet201_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_3.pth'
))
net3.to(device)
net3.eval()
# construct network
# net1 =MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net1.load_state_dict(torch.load('/home/dell/Desktop/2019BaiduXJTU/models/BDXJTU2019_SGD_16.pth'))
# net1.eval()
# net2 = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net2.load_state_dict(torch.load('/home/dell/Desktop/2019BaiduXJTU/models/BDXJTU2019_SGD_26.pth'))
# net2.eval()
# net3 =MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
# net3.load_state_dict(torch.load('/home/dell/Desktop/2019BaiduXJTU/models/BDXJTU2019_SGD_50.pth'))
# net3.eval()
results = []
results_anno = []
for i, (Input_img, Input_vis, Anno) in enumerate(Dataloader_val):
Input_img = Input_img.to(device)
Input_vis = Input_vis.to(device)
ConfTensor1 = net1.forward(Input_img, Input_vis)
ConfTensor2 = net2.forward(Input_img, Input_vis)
ConfTensor3 = net3.forward(Input_img, Input_vis)
ConfTensor = (torch.nn.functional.normalize(ConfTensor1) +
torch.nn.functional.normalize(ConfTensor2) +
torch.nn.functional.normalize(ConfTensor3)) / 3
score, pred = ConfTensor.data.topk(1, 1, True, False)
#print(score.item())
if (score.item() > 0.85):
results.append(pred.item())
results_anno.append(Anno) #append annotation results
if ((i + 1) % 2000 == 0):
print(i + 1)
print(len(results))
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
# print accuracy of different input
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
cnf_matrix = confusion_matrix(results_anno, results)
cnf_tr = np.trace(cnf_matrix)
cnf_tr = cnf_tr.astype('float')
print(cnf_tr / len(Dataset_val))
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
title='Confusion matrix, without normalization')
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
normalize=True,
title='Normalized confusion matrix')
plt.show()
def GeResult():
# Dataset
Dataset_val = MM_BDXJTU2019(root='data', mode='1_val')
Dataloader_val = data.DataLoader(Dataset_val,
batch_size=1,
num_workers=2,
shuffle=True,
pin_memory=True)
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
class_names = [
'001', '002', '003', '004', '005', '006', '007', '008', '009'
]
# construct network
epoch = 80
net = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_se_resnext50_32x4d_resample_pretrained_80w_1/BDXJTU2019_SGD_'
+ str(epoch) + '.pth'))
print('load ' + str(epoch) + ' epoch model')
net.eval()
results = []
results_anno = []
for i, (Input_img, Input_vis, Anno) in enumerate(Dataloader_val):
Input_img = Input_img.cuda()
Input_vis = Input_vis.cuda()
ConfTensor = net.forward(Input_img, Input_vis)
_, pred = ConfTensor.data.topk(1, 1, True, False)
results.append(pred.item())
results_anno.append(Anno) #append annotation results
if ((i + 1) % 1000 == 0):
print(i + 1)
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
# print accuracy of different input
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
cnf_matrix = confusion_matrix(results_anno, results)
cnf_tr = np.trace(cnf_matrix)
cnf_tr = cnf_tr.astype('float')
print(cnf_tr / len(Dataset_val))
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
title='Confusion matrix, without normalization')
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
normalize=True,
title='Normalized confusion matrix')
plt.show()
def GeResult():
# Dataset
Dataset_val = MM_BDXJTU2019_TTA(root='data', mode='val')
Dataloader_val = data.DataLoader(Dataset_val,
batch_size=1,
num_workers=4,
shuffle=True,
pin_memory=True)
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
class_names = [
'001', '002', '003', '004', '005', '006', '007', '008', '009'
]
# construct network
net = MultiModalNet('se_resnext50_32x4d', 'DPN26', 0.5)
net.load_state_dict(
torch.load(
'/home/zxw/2019BaiduXJTU/weights/MultiModal_100/BDXJTU2019_SGD_20.pth'
))
net.eval()
results = []
results_anno = []
for i, (Input_O, Input_H, Input_V, Input_visit,
Anno) in enumerate(Dataloader_val):
ConfTensor_O = net.forward(Input_O.cuda(), Input_visit.cuda())
ConfTensor_H = net.forward(Input_H.cuda(), Input_visit.cuda())
ConfTensor_V = net.forward(Input_V.cuda(), Input_visit.cuda())
ConfTensor = torch.nn.functional.normalize(
ConfTensor_O) + torch.nn.functional.normalize(
ConfTensor_H) + torch.nn.functional.normalize(ConfTensor_V)
_, pred = ConfTensor.data.topk(1, 1, True, False)
results.append(pred.item())
results_anno.append(Anno) #append annotation results
if (i % 1000 == 0):
print(i)
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
# print accuracy of different input
print('Accuracy of Orignal Input: %0.6f' %
(accuracy_score(results, results_anno, normalize=True)))
cnf_matrix = confusion_matrix(results_anno, results)
cnf_tr = np.trace(cnf_matrix)
cnf_tr = cnf_tr.astype('float')
print(cnf_tr / len(Dataset_val))
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
title='Confusion matrix, without normalization')
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=class_names,
normalize=True,
title='Normalized confusion matrix')
plt.show()
def GeResult():
# Priors
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.set_device(0)
model_names=["","se_resnext101_32x4d","se_resnext50_32x4d","se_resnet50","densenet169","densenet121"]
weight_names=["","model1","model2","model3","model_densenet169","model_densenet121"]
model_id=5
batch_size=4
print(model_id)
model_name=model_names[model_id]
weight_name=weight_names[model_id]
weightfiles=os.listdir(weight_name)
MAX=0
for file in weightfiles:
if file.count("_")<3:
continue
s=file.split("BDXJTU2019_SGD_")[1]
s=s.split("_")
s1=int(s[0])
s2=int(s[1].split(".")[0])
if MAX<s1*1000000+s2:
MAX=s1*1000000+s2
MAXfile=file
MAXs1=s1
print(MAXfile)
# Dataset
Dataset = MM_BDXJTU2019("data_txt", mode = 'val')
if model_name.count("dense")>0:
Dataset = MM_BDXJTU2019_for_dense("data_txt", mode='val')
elif model_name.count("nasnet")>0:
Dataset = MM_BDXJTU2019_for_nasnet("data_txt", mode='val')
print(batch_size)
Dataloader = data.DataLoader(Dataset,batch_size,
num_workers=1,
shuffle=False, pin_memory=True)
# Network
cudnn.benchmark = True
# Network = pnasnet5large(6, None)
# Network = ResNeXt101_64x4d(6)
net = MultiModalNet(model_name, 'DPN26', 0.5)
print("weights model"+str(model_id))
net.load_state_dict(torch.load(weight_name+'/'+MAXfile))
net.eval()
filename = 'val_result/['+str(model_id)+'_'+str(MAXs1)+'].txt'
import numpy as np
csvO=open('val_result/csvO['+str(model_id)+'_'+str(MAXs1)+'].csv',"w")
cnt=0
name_id=np.zeros([50000])
for i,(Input_O,visit_tensor, anoss,ids) in enumerate(Dataloader):
ConfTensor_Os = net.forward(Input_O.cuda(), visit_tensor.cuda())
for id in range(ConfTensor_Os.shape[0]):
ConfTensor_O=ConfTensor_Os[id].reshape([1,9])
anos=[anoss[id]]
name_id[cnt]=int(ids[id])
preds_temp = torch.nn.functional.normalize(ConfTensor_O)
string = ""
for _ in range(9):
string = string + str(float(preds_temp[0][_])) + ","
string = string + "\n"
csvO.write(string)
####################
preds = torch.nn.functional.normalize(ConfTensor_O)
_, pred = preds.data.topk(1, 1, True, True)
if cnt%100==0:
print(cnt)
cnt+=1
print(name_id[:5])
np.save("name_id"+str(model_id)+".npy",name_id)
csvO.close()
