def Modellist(model_name, num_classes=1000, use_attention=None):
if model_name == 'vgg11':
return vgg.VGG11(num_classes, use_attention=use_attention)
elif model_name == 'vgg13':
return vgg.VGG13(num_classes, use_attention=use_attention)
elif model_name == 'vgg16':
return vgg.VGG16(num_classes, use_attention=use_attention)
elif model_name == 'vgg19':
return vgg.VGG19(num_classes, use_attention=use_attention)
elif model_name == 'resnet18':
return resnet.ResNet18(num_classes, use_attention=use_attention)
elif model_name == 'resnet34':
return resnet.ResNet34(num_classes, use_attention=use_attention)
elif model_name == 'resnet50':
return resnet.ResNet50(num_classes, use_attention=use_attention)
elif model_name == 'resnet101':
return resnet.ResNet101(num_classes, use_attention=use_attention)
elif model_name == 'resnet152':
return resnet.ResNet152(num_classes, use_attention=use_attention)
elif model_name == 'densenet121':
return densenet.DenseNet121(num_classes, use_attention=use_attention)
elif model_name == 'densenet169':
return densenet.DenseNet169(num_classes, use_attention=use_attention)
elif model_name == 'densenet201':
return densenet.DenseNet201(num_classes, use_attention=use_attention)
elif model_name == 'densenet161':
return densenet.DenseNet161(num_classes, use_attention=use_attention)
elif model_name == 'mobilenetv3_small':
return mobilenetv3.mobilenetv3_small(num_classes)
elif model_name == 'mobilenetv3_large':
return mobilenetv3.mobilenetv3_large(num_classes)
elif model_name == 'efficientnet_b0':
return efficientnet.efficientnet_b0(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b1':
return efficientnet.efficientnet_b1(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b2':
return efficientnet.efficientnet_b2(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b3':
return efficientnet.efficientnet_b3(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b4':
return efficientnet.efficientnet_b4(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b5':
return efficientnet.efficientnet_b5(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b6':
return efficientnet.efficientnet_b6(num_classes,
use_attention=use_attention)
else:
raise ValueError("The model_name does not exist.")
def __init__(self, model, num_workers, lr, job_name):
self.lock = threading.Lock()
self.logger = Logger(job_name=job_name, file_dir=f"./measurement/logs/{job_name}_ps.log").logger
self.cm_t1_start = np.zeros(num_workers)
self.future_model = torch.futures.Future()
self.batch_update_size = num_workers
self.curr_update_size = 0
self.stop_flag = False
if model == 'resnet20':
self.model = resnet3.resnet20()
elif model == 'resnet56':
self.model = resnet3.resnet56()
elif model == 'resnet18':
self.model = resnet.ResNet18()
elif model == 'resnet50':
self.model = resnet.ResNet50()
elif model == 'vgg13':
self.model = vgg.VGG13()
elif model == 'vgg16':
self.model = vgg.VGG16()
elif model == 'densenet121':
self.model = densenet.DenseNet121()
elif model == 'alexnet':
self.model = alexnet.AlexNet()
elif model == 'googlenet':
self.model = googlenet.GoogLeNet()
elif model == 'mobilenet':
self.model = mobilenetv2.MobileNetV2()
self.lr = lr
for p in self.model.parameters():
p.grad = torch.zeros_like(p)
self.optimizer = optim.SGD(self.model.parameters(), lr=lr, momentum=0.9)
self.info_socketm = znet.SocketMsger.tcp_connect(DORKER0_IP, INFO_PORT)
self.info_socketm.send("PS")
self.info_socketm.send(f"1.0\n/home/ubuntu/measurement/logs/{job_name}_info0.log\n{job_name}")
self.ps_launched_lock = threading.Lock()
self.ps_launched = False
import model.mobilenetv3 as mbnet
import model.xception as xception
import model.resnet as resnet
import cv2
from solver import *
label_table = {0: "自然", 1: "高兴", 2: "悲伤", 3: "惊讶", 4: "恐惧", 5: "恶心", 6: "愤怒"}
color_table = [(100, 100, 0)] * 2 + [(0, 0, 155)] * 5
detector = CV2FaceDetector('ckpt/haarcascade_frontalface_default.xml')
# model = get_trained_model(mbnet.MobileNetV3_Small(), "ckpt/checkpoint_best.pth.tar")
# model = get_trained_model(xception.Xception(), "ckpt/checkpoint_best.pth.tar")
model = get_trained_model(resnet.ResNet50(),
"ckpt/checkpoint_best_resnet50.pth.tar")
smoother = LinearExponentialSmoothing(1)
classifier = ExpressionClassifier(model, label_table, smoother)
action_classifier = ActionClassifier()
cam_solver = CameraSolver(detector, classifier, action_classifier)
vizor = CV2Visualizer(label_table=classifier.express_table,
act_label_table=action_classifier.table,
color_table=color_table)
# cam_id = 0 # 本机摄像头
cam_id = 'rtsp://*****:*****@192.168.24.64:554/Streaming/Channels/101' # RTSP视频流
cam_solver.start(cam_id)
while True:
rt = cam_solver.get_solved_frame()
if rt:
vizor.update(*rt)
vizor.show()
import model.mobilenetv3 as mbnet
import model.xception as xception
import model.resnet as resnet
import cv2
from solver import *
label_table = {0:"自然", 1:"高兴", 2:"悲伤", 3:"惊讶", 4:"恐惧", 5:"恶心", 6:"愤怒"}
color_table = [(100, 100, 0)]* 2 + [(0,0,155)]*5
detector = CV2FaceDetector('ckpt/haarcascade_frontalface_default.xml')
# model = get_trained_model(mbnet.MobileNetV3_Small(), "ckpt/checkpoint_best.pth.tar")
# model = get_trained_model(xception.Xception(), "ckpt/checkpoint_best.pth.tar")
model = get_trained_model(resnet.ResNet50(), "ckpt/checkpoint_best_resnet50.pth.tar")
smoother = LinearExponentialSmoothing(1)
classifier = ExpressionClassifier(model, label_table, smoother)
# action_classifier = ActionClassifier()
cam_solver = CameraSolver(detector, classifier, None)
vizor = CV2Visualizer(label_table=classifier.express_table, act_label_table=None, color_table=color_table)
cam_id = 0 # 本机摄像头
# cam_id = 'rtsp://*****:*****@192.168.24.64:554/Streaming/Channels/101' # RTSP视频流
cam_solver.start(cam_id)
while True:
rt = cam_solver.get_solved_frame()
if rt:
vizor.update(*rt)
vizor.show()
cv2.waitKey(10)
cam_solver.close()
def main(train_path,
num_rois,
hf=True,
vf=False,
rot_90=False,
num_epoches=2000,
output_weight_path='weights/{}faster_rcnn.hdf5'.format(
time.localtime(time.time()))):
C = config.Config()
C.num_rois = int(num_rois)
C.use_horizontal_flips = hf
C.use_vertical_flips = vf
C.rot_90 = rot_90
C.model_path = output_weight_path
all_imgs, class_count, class_mapping = get_data(train_path)
if 'bg' not in class_count:
class_count['bg'] = 0
class_mapping['bg'] = len(class_mapping)
C.class_mapping = class_mapping
inv_map = {v: k for k, v in class_mapping.items()}
print('classes_count:', class_count)
print('class_mapping', class_mapping)
train_imgs = [s for s in all_imgs if s['type'] == 'train']
val_imgs = [s for s in all_imgs if s['type'] == 'val']
test_imgs = [s for s in all_imgs if s['type'] == 'test']
print("the number of train", len(train_imgs))
print('the number of val', len(val_imgs))
print('the number of test', len(test_imgs))
data_gen_train = data_generators.get_anchor_gt(train_imgs,
class_count,
C,
K.image_dim_ordering(),
mode='train')
data_gen_val = data_generators.get_anchor_gt(val_imgs,
class_count,
C,
K.image_dim_ordering(),
mode='val')
input_shape_img = (None, None, 3)
img_input = Input(shape=input_shape_img)
roi_input = Input(shape=(C.num_rois, 4))
shared_layers = resnet.ResNet50(img_input, trainable=True)
#RPN built on the base layers
num_anchors = len(C.anchor_box_scales) * len(C.anchor_box_ratios)
rpn = resnet.rpn(shared_layers, num_anchors)
classifier = resnet.classifier(shared_layers,
roi_input,
C.num_rois,
nb_classes=len(class_count),
trainable=True)
model_rpn = Model(img_input, rpn[:2])
model_classifier = Model([img_input, roi_input], classifier)
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
try:
print('loading weights')
model_rpn.load_weights(C.base_net_weight)
model_classifier.load_weights(C.base_net_weight)
except Exception as e:
print(e)
optimizer = Adam(lr=1e-3)
optimizer_classifier = Adam(lr=1e-3)
model_rpn.compile(optimizer=optimizer,
loss=[
model.losses.rpn_loss_cls(num_anchors),
model.losses.rpn_loss_regr(num_anchors)
])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
model.losses.class_loss_cls,
model.losses.class_loss_regr(len(class_count) - 1)
],
metrics={'dense_class_{}'.format(len(class_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 1000
iter_num = 0
losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()
best_loss = np.Inf
class_mapping_inv = inv_map
print('start training')
for epoch_num in range(num_epoches):
progbar = generic_utils.Progbar(epoch_length)
print('Epoch {}/{}'.format(epoch_num + 1, num_epoches))
while True:
try:
if len(rpn_accuracy_rpn_monitor) == epoch_length and C.verbose:
mean_overlapping_bboxes = float(
sum(rpn_accuracy_rpn_monitor)) / len(
rpn_accuracy_rpn_monitor)
rpn_accuracy_rpn_monitor = []
print(
'Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'
.format(mean_overlapping_bboxes, epoch_length))
if mean_overlapping_bboxes == 0:
print(
'RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.'
)
X, Y, img_data = data_gen_train.next()
loss_rpn = model_rpn.train_on_batch(X, Y)
P_rpn = model_rpn.predict_on_batch(X)
R = model.roi_helpers.rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
K.image_dim_ordering(),
use_regr=True,
overlap_thresh=0.7,
max_boxes=300)
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
X2, Y1, Y2 = model.roi_helpers.calc_iou(
R, img_data, C, class_mapping)
if X2 is None:
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
neg_samples = np.where(Y1[0, :, -1] == 1)
pos_samples = np.where(Y1[0, :, -1] == 0)
if len(neg_samples) > 0:
neg_samples = neg_samples[0]
else:
neg_samples = []
if len(pos_samples) > 0:
pos_samples = pos_samples[0]
else:
pos_samples = []
rpn_accuracy_rpn_monitor.append(len(pos_samples))
rpn_accuracy_for_epoch.append((len(pos_samples)))
if C.num_rois > 1:
if len(pos_samples) < C.num_rois / 2:
selected_pos_samples = pos_samples.tolist()
else:
selected_pos_samples = np.random.choice(
pos_samples, C.num_rois / 2,
replace=False).tolist()
try:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=False).tolist()
except:
selected_neg_samples = np.random.choice(
neg_samples,
C.num_rois - len(selected_pos_samples),
replace=True).tolist()
sel_samples = selected_pos_samples + selected_neg_samples
else:
# in the extreme case where num_rois = 1, we pick a random pos or neg sample
selected_pos_samples = pos_samples.tolist()
selected_neg_samples = neg_samples.tolist()
if np.random.randint(0, 2):
sel_samples = random.choice(neg_samples)
else:
sel_samples = random.choice(pos_samples)
loss_class = model_classifier.train_on_batch(
[X, X2[:, sel_samples, :]],
[Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
losses[iter_num, 0] = loss_rpn[1]
losses[iter_num, 1] = loss_rpn[2]
losses[iter_num, 2] = loss_class[1]
losses[iter_num, 3] = loss_class[2]
losses[iter_num, 4] = loss_class[3]
iter_num += 1
progbar.update(
iter_num,
[('rpn_cls', np.mean(losses[:iter_num, 0])),
('rpn_regr', np.mean(losses[:iter_num, 1])),
('detector_cls', np.mean(losses[:iter_num, 2])),
('detector_regr', np.mean(losses[:iter_num, 0]))])
if iter_num == epoch_length:
loss_rpn_cls = np.mean(losses[:, 0])
loss_rpn_regr = np.mean(losses[:, 1])
loss_class_cls = np.mean(losses[:, 2])
loss_class_regr = np.mean(losses[:, 3])
class_acc = np.mean(losses[:, 4])
mean_overlapping_bboxes = float(sum(
rpn_accuracy_for_epoch)) / len(rpn_accuracy_for_epoch)
rpn_accuracy_for_epoch = []
if C.verbose:
print(
'Mean number of bounding boxes from RPN overlapping ground truth boxes: {}'
.format(mean_overlapping_bboxes))
print(
'Classifier accuracy for bounding boxes from RPN: {}'
.format(class_acc))
print('Loss RPN classifier: {}'.format(loss_rpn_cls))
print('Loss RPN regression: {}'.format(loss_rpn_regr))
print('Loss Detector classifier: {}'.format(
loss_class_cls))
print('Loss Detector regression: {}'.format(
loss_class_regr))
print('Elapsed time: {}'.format(time.time() -
start_time))
curr_loss = loss_rpn_cls + loss_rpn_regr + loss_class_cls + loss_class_regr
iter_num = 0
start_time = time.time()
if curr_loss < best_loss:
if C.verbose:
print(
'Total loss decreased from {} to {}, saving weights'
.format(best_loss, curr_loss))
best_loss = curr_loss
model_all.save_weights(C.model_path)
break
except Exception as e:
print('Exception: {}'.format(e))
continue
print('Training complete, exiting.')
trainloader, testset, testloader, trainset_genuine = utils.load_data(
config.train_batch_size,
config.test_batch_size,
config.num_workers,
dataset='CIFAR10',
attack_set_size=config.attack_set_size,
binary=config.binary_dataset)
# classes = ('plane', 'car', 'bird', 'cat', 'deer',
# 'dog', 'frog', 'horse', 'ship', 'truck')
# Model
print('==> Building model..')
# net = VGG('VGG19')
if config.binary_dataset:
net = resnet.ResNet50(num_classes=1)
#net = keskar_models.C1(num_classes=1)
else:
net = resnet.ResNet50()
#net = keskar_models.C1(num_classes=10)
# net = resnet.ResNet50()
# net = PreActResNet18()
# net = GoogLeNet()
# net = DenseNet121()
# net = ResNeXt29_2x64d()
# net = MobileNet()
# net = MobileNetV2()
# net = DPN92()
# net = ShuffleNetG2()
# net = SENet18()
# net = ShuffleNetV2(1)
def main():
# Load the parameters from json file
args = parser.parse_args()
json_path = os.path.join(args.model_dir, 'params.json')
assert os.path.isfile(
json_path), "No json configuration file found at {}".format(json_path)
params = utils.Params(json_path)
# Set the random seed for reproducible experiments
random.seed(230)
torch.manual_seed(230)
np.random.seed(230)
torch.cuda.manual_seed(230)
warnings.filterwarnings("ignore")
# Set the logger
utils.set_logger(os.path.join(args.model_dir, 'train.log'))
# Create the input data pipeline
logging.info("Loading the datasets...")
# fetch dataloaders, considering full-set vs. sub-set scenarios
if params.subset_percent < 1.0:
train_dl = data_loader.fetch_subset_dataloader('train', params)
else:
train_dl = data_loader.fetch_dataloader('train', params)
dev_dl = data_loader.fetch_dataloader('dev', params)
logging.info("- done.")
"""
Load student and teacher model
"""
if "distill" in params.model_version:
# Specify the student models
if params.model_version == "cnn_distill": # 5-layers Plain CNN
print("Student model: {}".format(params.model_version))
model = net.Net(params).cuda()
elif params.model_version == "shufflenet_v2_distill":
print("Student model: {}".format(params.model_version))
model = shufflenet.shufflenetv2(class_num=args.num_class).cuda()
elif params.model_version == "mobilenet_v2_distill":
print("Student model: {}".format(params.model_version))
model = mobilenet.mobilenetv2(class_num=args.num_class).cuda()
elif params.model_version == 'resnet18_distill':
print("Student model: {}".format(params.model_version))
model = resnet.ResNet18(num_classes=args.num_class).cuda()
elif params.model_version == 'resnet50_distill':
print("Student model: {}".format(params.model_version))
model = resnet.ResNet50(num_classes=args.num_class).cuda()
elif params.model_version == "alexnet_distill":
print("Student model: {}".format(params.model_version))
model = alexnet.alexnet(num_classes=args.num_class).cuda()
elif params.model_version == "vgg19_distill":
print("Student model: {}".format(params.model_version))
model = models.vgg19_bn(num_classes=args.num_class).cuda()
elif params.model_version == "googlenet_distill":
print("Student model: {}".format(params.model_version))
model = googlenet.GoogleNet(num_class=args.num_class).cuda()
elif params.model_version == "resnext29_distill":
print("Student model: {}".format(params.model_version))
model = resnext.CifarResNeXt(cardinality=8,
depth=29,
num_classes=args.num_class).cuda()
elif params.model_version == "densenet121_distill":
print("Student model: {}".format(params.model_version))
model = densenet.densenet121(num_class=args.num_class).cuda()
# optimizer
if params.model_version == "cnn_distill":
optimizer = optim.Adam(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128))
else:
optimizer = optim.SGD(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128),
momentum=0.9,
weight_decay=5e-4)
iter_per_epoch = len(train_dl)
warmup_scheduler = utils.WarmUpLR(
optimizer, iter_per_epoch *
args.warm) # warmup the learning rate in the first epoch
# specify loss function
if args.self_training:
print(
'>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>self training>>>>>>>>>>>>>>>>>>>>>>>>>>>>>'
)
loss_fn_kd = loss_kd_self
else:
loss_fn_kd = loss_kd
"""
Specify the pre-trained teacher models for knowledge distillation
Checkpoints can be obtained by regular training or downloading our pretrained models
For model which is pretrained in multi-GPU, use "nn.DaraParallel" to correctly load the model weights.
"""
if params.teacher == "resnet18":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet18(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet18/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet18/0.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "alexnet":
print("Teacher model: {}".format(params.teacher))
teacher_model = alexnet.alexnet(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_alexnet/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "googlenet":
print("Teacher model: {}".format(params.teacher))
teacher_model = googlenet.GoogleNet(num_class=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_googlenet/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "vgg19":
print("Teacher model: {}".format(params.teacher))
teacher_model = models.vgg19_bn(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_vgg19/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "resnet50":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet50(num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet50/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet50/50.pth.tar'
elif params.teacher == "resnet101":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet101(num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet101/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "densenet121":
print("Teacher model: {}".format(params.teacher))
teacher_model = densenet.densenet121(
num_class=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_densenet121/best.pth.tar'
# teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "resnext29":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnext.CifarResNeXt(
cardinality=8, depth=29, num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnext29/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnext29/50.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "mobilenet_v2":
print("Teacher model: {}".format(params.teacher))
teacher_model = mobilenet.mobilenetv2(
class_num=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_mobilenet_v2/best.pth.tar'
elif params.teacher == "shufflenet_v2":
print("Teacher model: {}".format(params.teacher))
teacher_model = shufflenet.shufflenetv2(
class_num=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_shufflenet_v2/best.pth.tar'
utils.load_checkpoint(teacher_checkpoint, teacher_model)
# Train the model with KD
logging.info("Starting training for {} epoch(s)".format(
params.num_epochs))
train_and_evaluate_kd(model, teacher_model, train_dl, dev_dl,
optimizer, loss_fn_kd, warmup_scheduler, params,
args, args.restore_file)
# non-KD mode: regular training to obtain a baseline model
else:
print("Train base model")
if params.model_version == "cnn":
model = net.Net(params).cuda()
elif params.model_version == "mobilenet_v2":
print("model: {}".format(params.model_version))
model = mobilenet.mobilenetv2(class_num=args.num_class).cuda()
elif params.model_version == "shufflenet_v2":
print("model: {}".format(params.model_version))
model = shufflenet.shufflenetv2(class_num=args.num_class).cuda()
elif params.model_version == "alexnet":
print("model: {}".format(params.model_version))
model = alexnet.alexnet(num_classes=args.num_class).cuda()
elif params.model_version == "vgg19":
print("model: {}".format(params.model_version))
model = models.vgg19_bn(num_classes=args.num_class).cuda()
elif params.model_version == "googlenet":
print("model: {}".format(params.model_version))
model = googlenet.GoogleNet(num_class=args.num_class).cuda()
elif params.model_version == "densenet121":
print("model: {}".format(params.model_version))
model = densenet.densenet121(num_class=args.num_class).cuda()
elif params.model_version == "resnet18":
model = resnet.ResNet18(num_classes=args.num_class).cuda()
elif params.model_version == "resnet50":
model = resnet.ResNet50(num_classes=args.num_class).cuda()
elif params.model_version == "resnet101":
model = resnet.ResNet101(num_classes=args.num_class).cuda()
elif params.model_version == "resnet152":
model = resnet.ResNet152(num_classes=args.num_class).cuda()
elif params.model_version == "resnext29":
model = resnext.CifarResNeXt(cardinality=8,
depth=29,
num_classes=args.num_class).cuda()
# model = nn.DataParallel(model).cuda()
if args.regularization:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Loss of Regularization>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = loss_kd_regularization
elif args.label_smoothing:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Label Smoothing>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = loss_label_smoothing
else:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Normal Training>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = nn.CrossEntropyLoss()
if args.double_training: # double training, compare to self-KD
print(
">>>>>>>>>>>>>>>>>>>>>>>>Double Training>>>>>>>>>>>>>>>>>>>>>>>>"
)
checkpoint = 'experiments/pretrained_teacher_models/base_' + str(
params.model_version) + '/best.pth.tar'
utils.load_checkpoint(checkpoint, model)
if params.model_version == "cnn":
optimizer = optim.Adam(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128))
else:
optimizer = optim.SGD(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128),
momentum=0.9,
weight_decay=5e-4)
iter_per_epoch = len(train_dl)
warmup_scheduler = utils.WarmUpLR(optimizer,
iter_per_epoch * args.warm)
# Train the model
logging.info("Starting training for {} epoch(s)".format(
params.num_epochs))
train_and_evaluate(model, train_dl, dev_dl, optimizer, loss_fn, params,
args.model_dir, warmup_scheduler, args,
args.restore_file)
from model import resnet
import utils
import cv2
import numpy as np
import argparse
import torch
parser = argparse.ArgumentParser(description='GradCAM')
parser.add_argument('path', type=str, help='image path')
#parser.add_argument('state_dict_path',type=str, help='state_dict_path')
args = parser.parse_args()
model = resnet.ResNet50()
#model.load_state_dict(torch.load(args.state_dict_path))
image_path = args.path
img = cv2.imread(image_path, 1)
img = np.float32(cv2.resize(img, (224, 224))) / 255
input = utils.preprocess_image(img)
use_cuda = torch.cuda.is_available()
target_index = None
grad_cam = utils.GradCam(model=model, feature_module=model.layer4, \
target_layer_names=["2"], use_cuda=use_cuda)
mask = grad_cam(input, target_index)
utils.show_cam_on_image(img, mask)
gb_model = utils.GuidedBackpropReLUModel(model=model, use_cuda=use_cuda)
gb = gb_model(input, index=target_index)
gb = gb.transpose((1, 2, 0))
cam_mask = cv2.merge([mask, mask, mask])