import torch.nn as nn import torchvision.models as models import flexflow.torch.fx as fx # alexnet = models.alexnet() # fx.torch_to_flexflow(alexnet, "alexnet.ff") # # vgg16 = models.vgg16() # fx.torch_to_flexflow(vgg16, "vgg16.ff") # # squeezenet = models.squeezenet1_0() # fx.torch_to_flexflow(squeezenet, "squeezenet.ff") # densenet = models.densenet161() # fx.torch_to_flexflow(densenet, "densenet.ff") # inception = models.inception_v3() # fx.torch_to_flexflow(inception, "inception.ff") googlenet = models.googlenet() fx.torch_to_flexflow(googlenet, "googlenet.ff") # shufflenet = models.shufflenet_v2_x1_0() # fx.torch_to_flexflow(shufflenet, "shufflenet.ff") # mobilenet = models.mobilenet_v2() # fx.torch_to_flexflow(mobilenet, "mobilenet.ff")
def classify_images():
net = models.googlenet(pretrained=True)
net.eval()
transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
paths = {
'../data/ILSVRC2012_img_val/raw/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify.csv',
'../data/ILSVRC2012_img_val/originalImgModification/denoised/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_denoised.csv',
'../data/ILSVRC2012_img_val/originalImgModification/sharpen/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_sharpened.csv',
'../data/ILSVRC2012_img_val/originalImgModification/bilateralfilter/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_bilateralfilter.csv',
'../data/ILSVRC2012_img_val/originalImgModification/gaussianblur/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_gaussianblur.csv',
'../data/ILSVRC2012_img_val/originalImgModification/medianblur/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_medianblur.csv',
'../data/ILSVRC2012_img_val/perturbed/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed.csv', #you may have to change this filepath
'../data/ILSVRC2012_img_val/perturbedModification/denoised/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed_denoised.csv',
'../data/ILSVRC2012_img_val/perturbedModification/sharpen/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed_sharpened.csv',
'../data/ILSVRC2012_img_val/perturbedModification/bilateralfilter/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed_bilateralfilter.csv',
'../data/ILSVRC2012_img_val/perturbedModification/gaussianblur/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed_gaussianblur.csv',
'../data/ILSVRC2012_img_val/perturbedModification/medianblur/':
'../data/ILSVRC2012_img_val/classification/imagenet_classify_perturbed_medianblur.csv'
}
for input, output in paths.items():
files = [f for f in glob.glob(input + "**/*.JPEG", recursive=True)]
bar = progressbar.ProgressBar(maxval=len(files)).start()
classify = {}
i = 0
for path in files:
img = Image.open(path)
if (img.mode != "RGB"):
img = img.convert(mode="RGB")
image = transform(img)
f_image = net.forward(
Variable(image[None, :, :, :],
requires_grad=True)).data.cpu().numpy().flatten()
I = (np.array(f_image)).flatten().argsort()[::-1]
label = I[0]
classify[os.path.basename(path)] = label
i = i + 1
bar.update(i)
with open(output, 'w+') as f:
writer = csv.writer(f, lineterminator='\n')
for row in classify.items():
writer.writerow(row)
def get_model(name, device):
if name == "normal_cnn":
model = Net().to(device)
return model
elif name == "alexnet":
model = models.alexnet().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg11":
model = models.vgg11().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg13":
model = models.vgg13().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg16":
model = models.vgg16().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg19":
model = models.vgg19().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg11_bn":
model = models.vgg11_bn().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg13_bn":
model = models.vgg13_bn().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg16_bn":
model = models.vgg16_bn().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "vgg19_bn":
model = models.vgg19_bn().to(device)
num_features = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_features, 1).to(device)
return model
elif name == "densenet121":
model = models.densenet121().to(device)
num_features = model.classifier.in_features
model.classifier = nn.Linear(num_features, 1).to(device)
return model
elif name == "densenet161":
model = models.densenet161().to(device)
num_features = model.classifier.in_features
model.classifier = nn.Linear(num_features, 1).to(device)
return model
elif name == "densenet169":
model = models.densenet169().to(device)
num_features = model.classifier.in_features
model.classifier = nn.Linear(num_features, 1).to(device)
return model
elif name == "densenet201":
model = models.densenet201().to(device)
num_features = model.classifier.in_features
model.classifier = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnet18":
model = models.resnet18().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnet34":
model = models.resnet34().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnet50":
model = models.resnet50().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnet101":
model = models.resnet101().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnet152":
model = models.resnet152().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "googlenet":
model = models.googlenet(aux_logits=False).to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "inception_v3":
model = models.inception_v3(aux_logits=False).to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "shufflenet_v2_x0_5":
model = models.shufflenet_v2_x0_5().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "shufflenet_v2_x1_0":
model = models.shufflenet_v2_x1_0().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "shufflenet_v2_x1_5":
model = models.shufflenet_v2_x1_5().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "shufflenet_v2_x2_0":
model = models.shufflenet_v2_x2_0().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "mobilenet_v2":
model = models.mobilenet_v2().to(device)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnext50_32x4d":
model = models.resnext50_32x4d().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "resnext101_32x8d":
model = models.resnext101_32x8d().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "wide_resnet50_2":
model = models.wide_resnet50_2().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "wide_resnet101_2":
model = models.wide_resnet101_2().to(device)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 1).to(device)
return model
elif name == "mnasnet0_5":
model = models.mnasnet0_5().to(device)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 1).to(device)
return model
elif name == "mnasnet0_75":
model = models.mnasnet0_75().to(device)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 1).to(device)
return model
elif name == "mnasnet1_0":
model = models.mnasnet1_0().to(device)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 1).to(device)
return model
elif name == "mnasnet1_3":
model = models.mnasnet1_3().to(device)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 1).to(device)
return model
def test_googlenet_eval(self):
m = torch.jit.script(models.googlenet(pretrained=True).eval())
self.checkModule(m, "googlenet", torch.rand(1, 3, 224, 224))
def create_model(model_type, num_classes, feature_extract, pretrained):
"""
Creates a model.
:param model_type: Model type.
:param num_classes: Number of classes.
:param feature_extract: A boolean indicating if we are extracting features.
:param pretrained: A boolean indicating if pretrained weights should be used.
:return: Model.
"""
device = get_device()
if 'resnet18' == model_type:
model = models.resnet18(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'resnet34' == model_type:
model = models.resnet34(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'resnet50' == model_type:
model = models.resnet50(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'resnet101' == model_type:
model = models.resnet101(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'resnet152' == model_type:
model = models.resnet152(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'alexnet' == model_type:
model = models.alexnet(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(4096, num_classes)
elif 'vgg11' == model_type:
model = models.vgg11(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg11_bn' == model_type:
model = models.vgg11_bn(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg13' == model_type:
model = models.vgg13(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg13_bn' == model_type:
model = models.vgg13_bn(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg16' == model_type:
model = models.vgg16(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg16_bn' == model_type:
model = models.vgg16_bn(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg19' == model_type:
model = models.vgg19(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'vgg19_bn' == model_type:
model = models.vgg19_bn(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
num_classes)
elif 'squeezenet1_0' == model_type:
model = models.squeezenet1_0(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Conv2d(512,
num_classes,
kernel_size=(1, 1),
stride=(1, 1))
model.num_classes = num_classes
elif 'squeezenet1_1' == model_type:
model = models.squeezenet1_1(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Conv2d(512,
num_classes,
kernel_size=(1, 1),
stride=(1, 1))
model.num_classes = num_classes
elif 'densenet121' == model_type:
model = models.densenet121(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier = nn.Linear(model.classifier.in_features, num_classes)
elif 'densenet161' == model_type:
model = models.densenet161(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier = nn.Linear(model.classifier.in_features, num_classes)
elif 'densenet169' == model_type:
model = models.densenet169(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier = nn.Linear(model.classifier.in_features, num_classes)
elif 'densenet201' == model_type:
model = models.densenet201(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier = nn.Linear(model.classifier.in_features, num_classes)
elif 'googlenet' == model_type:
model = models.googlenet(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'shufflenet_v2_x0_5' == model_type:
model = models.shufflenet_v2_x0_5(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'shufflenet_v2_x1_0' == model_type:
model = models.shufflenet_v2_x1_0(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'mobilenet_v2' == model_type:
model = models.mobilenet_v2(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Linear(model.classifier[1].in_features,
num_classes)
elif 'resnext50_32x4d' == model_type:
model = models.resnext50_32x4d(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Linear(model.classifier[1].in_features,
num_classes)
elif 'resnext101_32x8d' == model_type:
model = models.resnext101_32x8d(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'wide_resnet50_2' == model_type:
model = models.wide_resnet50_2(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'wide_resnet101_2' == model_type:
model = models.wide_resnet101_2(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.fc = nn.Linear(model.fc.in_features, num_classes)
elif 'mnasnet0_5' == model_type:
model = models.mnasnet0_5(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Linear(model.classifier[1].in_features,
num_classes)
elif 'mnasnet1_0' == model_type:
model = models.mnasnet1_0(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.classifier[1] = nn.Linear(model.classifier[1].in_features,
num_classes)
else:
model = models.inception_v3(pretrained=pretrained)
set_parameter_requires_grad(model, feature_extract)
model.AuxLogits.fc = nn.Linear(model.AuxLogits.fc.in_features,
num_classes)
model.fc = nn.Linear(model.fc.in_features, num_classes)
return model.to(device)
def googlenet():
return models.googlenet(pretrained=True)
def initialize_model(model_name,
embedding_dim,
feature_extracting,
use_pretrained=True):
if model_name == "densenet161":
model_ft = models.densenet161(pretrained=use_pretrained,
memory_efficient=True)
set_parameter_requires_grad(model_ft, feature_extracting)
#print(model_ft)
num_features = model_ft.classifier.in_features
print(num_features)
#print(embedding_dim)
model_ft.classifier = nn.Linear(num_features, embedding_dim)
#print(model_ft)
if model_name == "densenet169":
model_ft = models.densenet161(pretrained=use_pretrained,
memory_efficient=True)
set_parameter_requires_grad(model_ft, feature_extracting)
#print(model_ft)
num_features = model_ft.classifier.in_features
print(num_features)
#print(embedding_dim)
model_ft.classifier = nn.Linear(num_features, embedding_dim)
#print(model_ft)
if model_name == "densenet121":
model_ft = models.densenet121(pretrained=use_pretrained,
memory_efficient=True)
set_parameter_requires_grad(model_ft, feature_extracting)
#print(model_ft)
num_features = model_ft.classifier.in_features
print(num_features)
#print(embedding_dim)
model_ft.classifier = nn.Linear(num_features, embedding_dim)
#print(model_ft)
if model_name == "densenet201":
model_ft = models.densenet201(pretrained=use_pretrained,
memory_efficient=True)
set_parameter_requires_grad(model_ft, feature_extracting)
#print(model_ft)
num_features = model_ft.classifier.in_features
print(num_features)
#print(embedding_dim)
model_ft.classifier = nn.Linear(num_features, embedding_dim)
#print(model_ft)
elif model_name == "resnet101":
model_ft = models.resnet101(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
print(model_ft)
elif model_name == "resnet34":
model_ft = models.resnet34(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
print(model_ft)
elif model_name == "adl_resnet50":
model_ft = adl_resnet50(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
print(model_ft)
elif model_name == "inceptionv3":
model_ft = models.inception_v3(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
elif model_name == "seresnext":
model_ft = se_resnext101_32x4d(num_classes=1000)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.last_linear.in_features
model_ft.last_linear = nn.Linear(num_features, embedding_dim)
elif model_name == "seresnext50":
model_ft = se_resnext50_32x4d(num_classes=1000)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.last_linear.in_features
model_ft.last_linear = nn.Linear(num_features, embedding_dim)
elif model_name == "googlenet":
model_ft = models.googlenet(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
#print(model_ft)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
elif model_name == "mobilenet2":
model_ft = models.MobileNetV2(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extracting)
num_features = model_ft.classifier[1].in_features
model_ft.classifier[1] = nn.Linear(num_features, embedding_dim)
print(model_ft)
elif model_name == "mnasnet":
model_ft = mnasnet1_0(pretrained=use_pretrained)
#model_ft = MnasNet()
set_parameter_requires_grad(model_ft, feature_extracting)
print(model_ft.classifier[0])
print(model_ft.classifier[1])
num_features = model_ft.classifier[1].in_features
print(num_features)
print(embedding_dim)
#model_ft.classifier[0] = nn.Dropout(p=0.2, inplace=False)
model_ft.classifier[1] = nn.Linear(num_features, embedding_dim)
print(model_ft)
elif model_name == "adl_googlenet":
model_ft = GoogLeNet()
set_parameter_requires_grad(model_ft, feature_extracting)
print(model_ft)
num_features = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_features, embedding_dim)
else:
raise ValueError
return model_ft
def main():
#=======================================
# 1. Load dataset
#=======================================
# CIFAR data
data_tf = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_set = datasets.CIFAR10('./data',
train=True,
transform=data_tf,
download=True)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=64,
shuffle=True)
test_set = datasets.CIFAR10('./data',
train=False,
transform=data_tf,
download=True)
valid_loader = torch.utils.data.DataLoader(test_set,
batch_size=128,
shuffle=False)
# custom datasets
# train_folders, valid_folders = get_folders(
# config.train_data, config.valid_data)
# train_datasets = MyDataSet(train_folders, transforms=None)
# valid_datasets = MyDataSet(valid_folders, transforms=None, train=False)
# train_loader = DataLoader(dataset=train_datasets,
# batch_size=config.batch_size, shuffle=True)
# valid_loader = DataLoader(dataset=valid_datasets,
# batch_size=config.batch_size, shuffle=True)
print("Train numbers:{:d}".format(len(train_loader)))
print("Test numbers:{:d}".format(len(valid_loader)))
#=======================================
# 2. Define network and Load model
#=======================================
if config.pretrained:
model = models.googlenet(num_classes=config.num_classes)
checkpoint = torch.load(config.model_path)
model.load_state_dict(checkpoint['model_state_dict'])
print("load model success")
else:
model = models.googlenet(pretrained=True)
# adjust last fc layer to class number
channel_in = model.fc.in_features
model.fc = nn.Linear(channel_in, config.num_classes)
model.to(device)
#=======================================
# 3. Define Loss function and optimizer
#=======================================
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.Adam(model.parameters(),
lr=config.lr,
amsgrad=True,
weight_decay=config.weight_decay)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1)
#=======================================
# 4. Train and Test the network
#=======================================
best_accuracy = 0.
epoch = 0
resume = False
# ====4.1 restart the training process====
if resume:
checkpoint = torch.load(config.model_path)
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint["epoch"]
loss = checkpoint["loss"]
best_accuracy = checkpoint["best_accuracy"]
# print checkpoint
print('epoch:', epoch)
print('best_accuracy:', best_accuracy)
print("model's state_dict:")
for param_tensor in model.state_dict():
print(param_tensor, '\t', model.state_dict()[param_tensor].size())
for epoch in range(1, config.epochs + 1):
# ====4.2 start training====
print("======start training=====")
model.train()
torch.cuda.empty_cache()
start = time.time()
index = 1
sum_loss = 0.
correct = 0.
total = 0.
for images, labels in train_loader:
# clear the cuda cache
torch.cuda.empty_cache()
images = images.to(device)
labels = labels.to(device, dtype=torch.long)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and Optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
# print training loss and accuracy
_, predicted = torch.max(outputs.data, 1)
sum_loss += loss.item()
total += labels.size(0)
correct += (predicted == labels).sum()
if index % 10 == 0:
print("Iter: %d ===> Loss: %.8f | Acc: %.3f%%" %
(index, sum_loss / index, 100. * correct / total))
index += 1
end = time.time()
print("Epoch [%d/%d], Loss: %.8f, Time: %.1fsec!" %
(epoch, config.epochs, loss.item(), (end - start)))
vis.line(X=[epoch],
Y=[loss.item()],
win='loss',
opts=dict(title='train loss'),
update='append')
# ====4.3 start evalidating====
model.eval()
correct_prediction = 0.
total = 0
with torch.no_grad():
for images, labels in valid_loader:
# clear the cuda cache
torch.cuda.empty_cache()
images = images.to(device)
labels = labels.to(device, dtype=torch.long)
# print prediction
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
# total labels number
total += labels.size(0)
# add correct
correct_prediction += (predicted == labels).sum().item()
# print("total correct number: ", correct_prediction)
accuracy = 100. * correct_prediction / total
print("Accuracy: %.4f%%" % accuracy)
scheduler.step()
if accuracy > best_accuracy:
if not os.path.exists(config.checkpoint):
os.mkdir(config.checkpoint)
# save networks
torch.save(
{
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
'best_accuracy': accuracy,
},
os.path.join("%s-%.3f.pth") % (config.best_models, accuracy))
print("save networks for epoch:", epoch)
best_accuracy = accuracy
def forward(self, x):
batch_size = x.size(0)
feat = self.FNet(x)
# incept4e and incept5b are middle layer of GNet
# then used to calculate OAM
incept4e = self.incept4e(feat)
incept5b = self.incept5b(incept4e)
oam = (self.OAM(incept4e) + self.OAM(incept5b)) / 2
# N x 480 x 14 x 14
caps = [
feat * self.CAMs[i](feat).view(batch_size, 1, 1, -1)
for i in range(self.cams)
]
# #cams of N x 480 x 14 x 14
cams = [
self.GNet(caps[i]).view(batch_size, -1) for i in range(self.cams)
] # each of size N x 1024
embeddings = [self.CA_learners[i](cams[i]) for i in range(self.cams)]
adv_reverse = [
self.CA_sub_adv(embeddings[i]) for i in range(self.cams)
]
adv = [self.CA_sub(embeddings[i]) for i in range(self.cams)]
return cams, embeddings, adv, adv_reverse
if __name__ == '__main__':
googlenet = models.googlenet(pretrained=False)
summary(googlenet, (3, 512, 512))
learning_rate = 10
n_iterations = 100
layers = [11] # For resnet50, layer can range from 0 to 9
# n_octaves = 7
# octave_scale = 1.4
# channel = 0
# ~~~~~~~~~~~~~~~~~~~
# set mean and std_deviation for imagenet
mean = [.485, .456, .406]
std = [.229, .224, .225]
# initialize net
net = models.googlenet(pretrained=True).to(device)
net.eval()
# preprocess image
img = Image.open("dog1.jpg")
img.show()
img = transform(img).unsqueeze(0).to(device)
children = list(net.children())
for i in range(len(layers)):
layers[i] = nn.Sequential(*children)[:layers[i] - len(children)]
for _ in range(n_iterations):
for layer in layers:
# apply jitter
y_jitter, x_jitter = np.random.randint(-5, 5, size=2)
def initialize_model(model_name='vgg16',
num_classes=1,
feature_extract=True,
use_pretrained=True):
# Initialize these variables which will be set in this if statement. Each of these
# variables is model specific.
model_ft = None
input_size = 0
if model_name == "googlenet":
""" googlenet
"""
model_ft = models.googlenet(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "resnet152":
""" Resnet152
"""
model_ft = models.resnet152(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "vgg16":
""" VGG16
"""
model_ft = models.vgg16(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier[6].in_features
model_ft.classifier[6] = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "resnet18":
""" Resnet18
"""
model_ft = models.resnet18(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "alexnet":
""" Alexnet
"""
model_ft = models.alexnet(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier[6].in_features
model_ft.classifier[6] = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "vgg11_bn":
""" VGG11_bn
"""
model_ft = models.vgg11_bn(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier[6].in_features
model_ft.classifier[6] = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "squeezenet":
""" Squeezenet
"""
model_ft = models.squeezenet1_0(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
model_ft.classifier[1] = nn.Conv2d(512,
num_classes,
kernel_size=(1, 1),
stride=(1, 1))
model_ft.num_classes = num_classes
input_size = 224
elif model_name == "densenet":
""" Densenet
"""
model_ft = models.densenet121(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.classifier.in_features
model_ft.classifier = nn.Linear(num_ftrs, num_classes)
input_size = 224
elif model_name == "inception":
""" Inception v3
Be careful, expects (299,299) sized images and has auxiliary output
"""
model_ft = models.inception_v3(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
# Handle the auxilary net
num_ftrs = model_ft.AuxLogits.fc.in_features
model_ft.AuxLogits.fc = nn.Linear(num_ftrs, num_classes)
# Handle the primary net
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 299
else:
print("Invalid model name, exiting...")
exit()
return model_ft, input_size
def get_model(model_id, use_pretrained):
model_ft = None
if model_id == PyTorchModelsEnum.ALEXNET:
model_ft = models.alexnet(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.DENSENET121:
model_ft = models.densenet121(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.DENSENET161:
model_ft = models.densenet161(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.DENSENET169:
model_ft = models.densenet169(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.DENSENET201:
model_ft = models.densenet201(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.GOOGLENET:
model_ft = models.googlenet(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.INCEPTION_V3:
model_ft = models.inception_v3(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.MOBILENET_V2:
model_ft = models.mobilenet_v2(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.MNASNET_0_5:
model_ft = models.mnasnet0_5(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.MNASNET_0_75: # no pretrained
model_ft = models.mnasnet0_75(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.MNASNET_1_0:
model_ft = models.mnasnet1_0(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.MNASNET_1_3:
model_ft = models.mnasnet1_3(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNET18:
model_ft = models.resnet18(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNET34:
model_ft = models.resnet34(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNET50:
model_ft = models.resnet50(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNET101:
model_ft = models.resnet101(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNET152:
model_ft = models.resnet152(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNEXT50:
model_ft = models.resnext50_32x4d(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.RESNEXT101:
model_ft = models.resnext101_32x8d(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SHUFFLENET_V2_0_5:
model_ft = models.shufflenet_v2_x0_5(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SHUFFLENET_V2_1_0:
model_ft = models.shufflenet_v2_x1_0(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SHUFFLENET_V2_1_5:
model_ft = models.shufflenet_v2_x1_5(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SHUFFLENET_V2_2_0:
model_ft = models.shufflenet_v2_x2_0(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SQUEEZENET1_0:
model_ft = models.squeezenet1_0(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.SQUEEZENET1_1:
model_ft = models.squeezenet1_1(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG11:
model_ft = models.vgg11(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG11_BN:
model_ft = models.vgg11_bn(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG13:
model_ft = models.vgg13(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG13_BN:
model_ft = models.vgg13_bn(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG16:
model_ft = models.vgg16(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG16_BN:
model_ft = models.vgg16_bn(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG19:
model_ft = models.vgg19(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.VGG19_BN:
model_ft = models.vgg19_bn(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.WIDE_RESNET50:
model_ft = models.wide_resnet50_2(pretrained=use_pretrained)
elif model_id == PyTorchModelsEnum.WIDE_RESNET101:
model_ft = models.wide_resnet101_2(pretrained=use_pretrained)
return model_ft
import cv2
# from model_utils import model, generate_mask, segmentation_model, layers
import runway
from runway.data_types import image, number, boolean
"""
works on torch_dreams v1.1.0
for tests use: $ pip install git+https://github.com/Mayukhdeb/torch-dreams
to run server on localhost: $ python runway_model.py
"""
"""
set up segmentation utils
"""
segmentation_model = models.segmentation.fcn_resnet101(pretrained=True).eval()
model = models.googlenet(pretrained=True).eval()
seg_transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
def generate_mask(model,
image,
transforms=seg_transforms,
invert=False,
factor=1.0):
inp = seg_transforms(image).unsqueeze(0)
def main():
print(torch.__version__)
# برای تقسیم داده ها به آموزش و تست به صورت خودکار
# این تابع درصدی از داده های آموزش را ضمن حفظ اسامی پوشه ها، به پوشه تست منتقل می کند
if(not os.path.exists(vars.val_dir)):
utils.create_validation_data(vars.train_dir, vars.val_dir, vars.val_split_ratio, 'jpg')
def handler(signum, frame):
print('Signal handler called with signal', signum)
print('Training will finish after this epoch')
vars.stop_training = True
#raise OSError("Couldn't open vars.device!")
signal.signal(signal.SIGINT, handler) # only in python version >= 3.2
print("Start Time: ", strftime("%Y-%m-%d %H:%M:%S", localtime()))
print("Active Mode: " + vars.mode)
plt.ion() # interactive mode
######################################################################
# Load Data
# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
'train': transforms.Compose([
transforms.Resize([vars.input_size, vars.input_size]),
#transforms.ColorJitter(0.1, 0.1, 0.1, 0.01),
#transforms.RandomRotation(5),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'val': transforms.Compose([
transforms.Resize([vars.input_size, vars.input_size]),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'test': transforms.Compose([
transforms.Resize([vars.input_size, vars.input_size]),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}
# image_dataset_train = {'train': datasets.ImageFolder(os.path.join(vars.data_dir, 'train'), data_transforms['train'])}
# image_dataset_test = {'val': datasets.ImageFolder(os.path.join(vars.data_dir, 'val'), data_transforms['val'])}
# image_dataset_train.update(image_dataset_test)
# خط پایین با سه خط بالا برابری می کند!
image_datasets = {x: datasets.ImageFolder(os.path.join(vars.data_dir, x), data_transforms[x])
for x in ['train', 'val']}
vars.dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=vars.batch_size, shuffle=True, num_workers=0)
for x in ['train', 'val']}
vars.dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
vars.class_names = image_datasets['train'].classes
# Get a batch of training data
inputs, classes = next(iter(vars.dataloaders['train']))
# Make a grid from batch
out = torchvision.utils.make_grid(inputs)
#utils.imshow(out, title=[vars.class_names[x] for x in classes])
######################################################################
# Finetuning the convnet
# Load a pretrained model and reset final fully connected layer.
##\\//\\//model, input_size = initialize_model(model_name, num_classes, feature_extract, use_pretrained=True)
if(vars.model_name.find('vgg') != -1):
model = models.vgg11_bn(pretrained=vars.pre_trained)
num_ftrs = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_ftrs, len(vars.class_names))
elif(vars.model_name == 'resnet152'):
model = models.resnet152(pretrained=vars.pre_trained)
elif(vars.model_name == 'resnet50'):
model = models.resnet50(pretrained=vars.pre_trained)
elif(vars.model_name == 'resnet18'):
model = models.resnet18(pretrained=vars.pre_trained)
elif(vars.model_name == 'googlenet'):
model = models.googlenet(pretrained=vars.pre_trained)
elif(vars.model_name == 'darknet53'):
model = darknet.darknet53(1000)
optimizer = optim.SGD(model.parameters(), lr = vars.learning_rate, momentum=0.9)
if(vars.pre_trained):
#model.load_state_dict( torch.load('D:\\Projects\\_Python\\car Detection\\model_best.pth.tar') )
checkpoint = torch.load('D:\\Projects\\_Python\\car Detection\\model_best.pth.tar')
start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
if vars.device.type == 'cuda':
# best_acc1 may be from a checkpoint from a different GPU
best_acc1 = best_acc1.to(vars.device)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
elif(vars.model_name == 'car3conv'):
model = mymodels.car3conv()
elif(vars.model_name == 'car5conv'):
model = mymodels.car5conv()
elif(vars.model_name == 'car2conv'):
model = mymodels.car2conv()
elif(vars.model_name == 'mymodel'):
model = mymodels.MyModel()
elif(vars.model_name == 'mymodel2'):
model = mymodels.MyModel2()
else:
return hogmain.main(None)
if(vars.model_name != 'vgg11'):
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, len(vars.class_names))
model.to(vars.device)
if(vars.mode == 'test'):#test
#model.load_state_dict(torch.load("D:\\Projects\\Python\\Zeitoon Detection\"))
model.load_state_dict(torch.load(vars.test_model))
# log_dir = '{}-{}-{}-batch-{}'.format(vars.model_name, 'SGD', 'cuda', vars.batch_size)
# if(vars.pre_trained):
# log_dir = log_dir + '-pretrained'
# if not os.path.exists(log_dir):
# os.mkdir(log_dir)
log_file = open(".\\Time-{}-{}.log".format(vars.model_name, vars.batch_size),"w")
for dev in ['cuda', 'cpu']:
vars.device = torch.device(dev)
model = model.to(vars.device)
#run model on one batch to allocate required memory on device (and have more exact results)
inputs = inputs.to(vars.device)
outputs = model(inputs)
s = test_model(model, vars.criterion, 'val', 100)
log_file.write(s)
#log_file.write('\n' + '-'*80)
log_file.write(summary(model, input_size=(3, vars.input_size, vars.input_size), batch_size=-1, device=vars.device.type))
log_file.close()
print(summary(model, input_size=(3, vars.input_size, vars.input_size), batch_size=-1, device=vars.device.type))
else:
print(summary(model, input_size=(3, vars.input_size, vars.input_size), batch_size=-1, device=vars.device.type))
#model.load_state_dict(torch.load("C:\\Projects\\Python\\Car Detection\\darknet53-SGD-cuda-batch-32\\ep7-acc97.83-loss0.0667.pth"))
optimizer = optim.SGD(model.parameters(), lr = vars.learning_rate, momentum=0.9)
#optimizer = optim.Adam(model.parameters(), lr=0.05)
# Decay LR by a factor of 0.6 every 6 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size = vars.scheduler_step_size, gamma = vars.scheduler_gamma)
model = model.to(vars.device)
model = train_model(model, vars.criterion, optimizer, exp_lr_scheduler, vars.num_epochs)
visualize_model(model)
plt.ioff()
plt.show()
def adam(model: nn.Module, lr=0.001):
return torch.optim.Adam(model.parameters(), lr=lr)
def sgd(model: nn.Module, lr=0.001):
return torch.optim.SGD(model.parameters(), lr=lr)
OPTIMIZERS = {
'adam': adam,
'sgd': sgd,
}
FEATURE_EXTRACTORS = {
'resnet18': lambda: models.resnet18(pretrained=False),
'googlenet': lambda: models.googlenet(pretrained=False),
'resnet18-np': lambda: ResNet18NoFlattening(pretrained=False),
'resnet12-np': lambda: ResNet12NoFlattening(),
'resnet12-np-o': lambda: ResNet12NoFlatteningOriginal(),
'conv256-np-o': lambda: ConvNet256Original(),
'conv64-np-o': lambda: ConvNet64Original(),
'conv64-p-o': lambda: ConvNet64PoolingOriginal(),
}
class FSLEpisodeSampler(Dataset):
def __init__(self, subdataset: LabeledSubdataset, n_way: int, n_shot: int, batch_size: int, balanced: bool,
device=torch.device("cuda:0" if torch.cuda.is_available() else "cpu")):
self.subdataset = subdataset
self.n_way = n_way
self.n_shot = n_shot
def init_from_googlenet(self):
googlenet = models.googlenet(pretrained=True)
self.load_state_dict(googlenet.state_dict(), strict=False)
self.transform_input = True
def main():
opts = get_train_args()
print("load data ...")
train_data = datasets.ImageFolder(
root="data/train",
transform=transforms.Compose([
transforms.Resize((256, 256)), # 한 축을 128로 조절하고
#transforms.CenterCrop(256), # square를 한 후,
transforms.ToTensor(), # Tensor로 바꾸고 (0~1로 자동으로 normalize)
transforms.Normalize(
(0.5, 0.5, 0.5), # -1 ~ 1 사이로 normalize
(0.5, 0.5, 0.5)), # (c - m)/s 니까...
]))
test_data = datasets.ImageFolder(
root="data/test",
transform=transforms.Compose([
transforms.Resize((256, 256)), # 한 축을 128로 조절하고
#transforms.CenterCrop(256), # square를 한 후,
transforms.ToTensor(), # Tensor로 바꾸고 (0~1로 자동으로 normalize)
transforms.Normalize(
(0.5, 0.5, 0.5), # -1 ~ 1 사이로 normalize
(0.5, 0.5, 0.5)), # (c - m)/s 니까...
]))
test_loader = DataLoader(test_data,
batch_size=opts.batch_size,
shuffle=False,
num_workers=opts.num_processes)
classes = train_data.classes
print(classes)
print("load model ...")
if opts.model == 'resnet':
model = models.resnet50(progress=True)
model.load_state_dict(torch.load('resnet_model.pt'))
elif opts.model == 'vggnet':
model = models.vgg13_bn(progress=True)
model.load_state_dict(torch.load('vggnet_model.pt'))
elif opts.model == 'googlenet':
model = models.googlenet(progress=True)
model.load_state_dict(torch.load('googlenet_model.pt'))
elif opts.model == 'densenet':
model = models.densenet121(progress=True)
model.load_state_dict(torch.load('densenet_model.pt'))
else:
model = models.resnext50_32x4d(progress=True)
model.load_state_dict(torch.load('resnext_model.pt'))
print(opts.model)
model.cuda()
print("start inference")
idx = 0
with torch.no_grad():
with open(opts.model + '_result.txt', 'a') as f:
for i, (inputs, labels) in enumerate(test_loader):
inputs = inputs.cuda()
outputs = model(inputs)
_, predicted = outputs.max(1)
for j, meta in enumerate(predicted):
predicted_class = classes[meta]
plant_class = predicted_class.split('_')[0]
disease_class = predicted_class.split('_')[1]
f.write(
str(test_loader.dataset.samples[idx][0].split('/')
[-1].split('.')[0]) + '\t' + plant_class + '\t' +
disease_class + '\n')
idx += 1
def main():
# Number of images to perturb, 0 means "all of them"
N = 0
# List to hold L2 norms of r for all perturbed images so rho can be caluclated at the end
r_arr = []
# List to hold original labels
orig_labels = []
# List to hold perturbed labels
pert_labels = []
# List to hold L2 norms
L2_norms = []
# Cumulative sum for rho
rho_sum = 0
j = 0
# Network you're using (can change to whatever)
net = models.googlenet(pretrained=True)
# Switch to evaluation mode
net.eval()
def clip_tensor(A, minv, maxv):
A = torch.max(A, minv * torch.ones(A.shape))
A = torch.min(A, maxv * torch.ones(A.shape))
return A
base = "../data/ILSVRC2012_img_val/"
# Get list of files in ImageNet directory (MAKE SURE `base` ABOVE IS CORRECT)
# First get the list of already perturbed files
pert_walk_gen = os.walk(base + "perturbed/", topdown=True)
_, _, pert_files_raw = next(pert_walk_gen)
pert_files = [p for p in pert_files_raw
if p != ".gitignore"] # Remove .gitignore
# Now process the files
raw_walk_gen = os.walk(base + "raw/", topdown=True)
_, _, files_raw = next(raw_walk_gen)
files_tmp = [f for f in files_raw
if f != ".gitignore"] # Remove .gitignore
files = [f for f in files_tmp if not (f in pert_files)
] # Don't process pictures that have already been perturbed
sorted_files = sorted(files, key=lambda item: int(item[18:23]))
end = N if N != 0 else len(sorted_files)
try:
# Now for the correct number of images
for name in sorted_files[0:end]:
orig_img = Image.open(base + "raw/" + name)
# Preprocessing only works for colour images
if (orig_img.mode != "RBG"):
orig_img = orig_img.convert(mode="RGB")
if (orig_img.mode == "RGB"): # Belt-and-suspenders check
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
# Get vector form of image so L2 norm of image x can be calculated (See denominator of eqn 15 in DeepFool paper)
img_arr = np.array(orig_img)
img_vect = img_arr.ravel()
L2_norms.append(np.linalg.norm(img_vect))
# Remove the mean
im = transforms.Compose([
transforms.Resize(256),
#transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])(orig_img)
r, loop_i, label_orig, label_pert, pert_image = deepfool(
im, net)
# Add L2 norm of perturbation to array (See numerator of eqn 15 in DeepFool paper)
r_norm = np.linalg.norm(r)
r_arr.append(r_norm)
# Add original labels and perturbed labels to array (just in case you need them later, not rlly using rn)
orig_labels.append(label_orig)
pert_labels.append(label_pert)
labels = open(os.path.join('synset_words.txt'),
'r').read().split('\n')
str_label_orig = labels[np.int(label_orig)].split(',')[0]
str_label_pert = labels[np.int(label_pert)].split(',')[0]
print("Original label = ", str_label_orig)
print("Perturbed label = ", str_label_pert)
print("Number of Iterations = ", loop_i)
clip = lambda x: clip_tensor(x, 0, 255)
tf = transforms.Compose([
transforms.Normalize(mean=[0, 0, 0],
std=[(1 / 0.229), (1 / 0.244),
(1 / 0.255)]),
transforms.Normalize(mean=[-0.485, -0.456, -0.406],
std=[1, 1, 1]),
transforms.Lambda(clip),
transforms.ToPILImage(),
#transforms.Resize(256),
#transforms.CenterCrop(224)
])
# Add to cumulative sum term to get rho (See eqn 15 in DeepFool paper)
rho_sum = rho_sum + r_norm / np.linalg.norm(img_vect)
# Write image file to directory to hold perturbed images
if (os.path.exists(base + 'perturbed/') != 1):
os.mkdir(base + 'perturbed/')
tf(pert_image.cpu()[0]).save(base + 'perturbed/' + name)
# Create .csv with original saved image name and predicted label
with open(base + 'orig_names_and_labels.csv', 'a') as orig_f:
writer = csv.writer(orig_f, lineterminator='\n')
writer.writerow([name, str(label_orig)])
# Create .csv with perturbed saved image name and predicted label
with open(base + 'pert_names_and_labels.csv', 'a') as pert_f:
writer = csv.writer(pert_f, lineterminator='\n')
writer.writerow([name, str(label_pert)])
j += 1
else:
raise TypeError(
f"expected Image Mode RBG, got {orig_img.mode}")
except KeyboardInterrupt:
pass
# Compute average robustness (rho) for the simulation (See eqn 15 in DeepFool paper)
rho = (1 / j) * rho_sum
print(f"Number of samples: {j}, Average robustness: {rho}")
with open(base + "rho.txt", 'a') as rho_f:
rho_f.write(f"Number of samples: {j}, Average robustness: {rho}\n")
def training(self):
self.tr_loss_track, self.val_loss_track, self.tr_acc_track, self.val_acc_track, self.val_acc_history = [], [], [], [], []
print('training', self.model_type, self.optim_type, self.learning_rate)
# Defining pretrained model
if self.model_type == "vgg19":
model = models.vgg19(pretrained=True)
model.classifier[6] = nn.Linear(4096, self.class_labels)
elif self.model_type == "resnet50":
model = models.resnet50(pretrained=True)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, self.class_labels)
elif self.model_type == "resnet101":
model = models.resnet101(pretrained=True)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, self.class_labels)
elif self.model_type == "resnet151":
model = models.resnet101(pretrained=True)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, self.class_labels)
elif self.model_type == "googlenet":
model = models.googlenet(pretrained=True)
model.classifier = nn.Linear(1000, self.class_labels)
elif self.model_type == "densenet121":
model = models.densenet121(pretrained=True)
num_features = model.classifier.in_features
model.classifier = nn.Linear(num_features, self.class_labels)
# #Without fine tuning
elif self.model_type == "un_vgg19":
# print("unpretrained1")
model = models.vgg19(pretrained=False)
model.classifier[6] = nn.Linear(4096, self.class_labels)
# set pretrained =False for models
elif self.model_type == "un_resnet50":
# print("unpretrained2")
model = models.resnet50(pretrained=False)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, self.class_labels)
model = model.to(self.device)
criterian = nn.CrossEntropyLoss()
if self.optim_type == "adam":
optimizer = torch.optim.Adam(model.parameters(),
lr=self.learning_rate)
elif self.optim_type == "sgd":
optimizer = torch.optim.SGD(model.parameters(),
lr=self.learning_rate,
momentum=0.9)
# training
start_time = time.time()
best_loss = float('inf')
best_model = None
best_acc = 0.0
# best_model_wts = copy.deepcopy(model.state_dict())
for epoch in range(1, self.num_epochs + 1):
print('epoch {}/{}'.format(epoch, self.num_epochs))
for t in ['train', 'val']:
num_correct = 0.0
num_samples = 0
r_loss = 0.0
running_corrects = 0
if t == 'train':
# training mode
model.train()
else:
# evaluate model
model.eval()
count = 0
for data in self.dataloaders[t]:
count += 1
# data has three types files, labels and filename
files, labels, filename = data
files = Variable(files.to(self.device)) # to gpu or cpu
labels = Variable(labels.to(self.device))
optimizer.zero_grad(
) # clearning old gradient from last step
with torch.set_grad_enabled(t == 'train'):
pred = model(files)
# loss computation
loss = criterian(pred, labels)
_, prediction = torch.max(pred, 1)
# backprop gradients at training time
if t == 'train':
loss.backward()
optimizer.step()
# print(t +' iteration {}:loss {} '.format(count,r_loss))
# statistics
r_loss += loss.item() * files.size(0)
print(t + ' iteration {}:loss {} '.format(count, r_loss))
running_corrects += torch.sum(prediction == labels.data)
epoch_loss = r_loss / len(self.dataloaders[t].dataset)
epoch_acc = running_corrects.double() / len(
self.dataloaders[t].dataset)
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
t, epoch_loss, epoch_acc))
# print(t +' epoch {}:loss {} '.format(epoch,r_loss))
# deep copy the model
# print('epoch_acc',epoch_acc,'best_acc',best_acc)
if t == 'val' and epoch_acc > best_acc:
print('inside check point if')
best_acc = epoch_acc
best_model_wts = copy.deepcopy(model.state_dict())
self.checkpoint(best_model_wts, best_loss, epoch,
self.learning_rate)
if t == 'val':
self.val_acc_history.append(epoch_acc.item())
self.val_loss_track.append(epoch_loss)
if t == 'train':
self.tr_loss_track.append(epoch_loss)
self.tr_acc_track.append(epoch_acc.item())
time_elapsed = time.time() - start_time
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
# load best model weights
# model.load_state_dict(best_model_wts)
# updating best model in checkpoint
self.plot_losses_both(self.tr_loss_track, self.val_loss_track)
self.plot_loss_Val(self.val_loss_track)
self.plot_loss_Accu(self.val_acc_history)
def test_googlenet(self):
process_model(models.googlenet(), self.image,
_C_tests.forward_googlenet, "GoogLeNet")
def main():
data_transforms = {
"train":
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]),
"val":
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]),
"test":
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]),
}
model_ft = models.googlenet(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 2.
# Alternatively, it can be generalized to nn.Linear(num_ftrs, len(class_names)).
model_ft.aux_logits = False
model_ft.fc = nn.Linear(num_ftrs, 2)
device = "cpu"
model_ft = model_ft.to(device)
# Load model
model_ft.load_state_dict(torch.load("weightsOld"))
print("loaded weights")
image = Image.open("data/train/notPotH/negative982.JPG")
target_class = 1
prep_image = data_transforms["val"](image)
print(prep_image.shape)
prep_image.requires_grad_()
scores = model_ft(prep_image.reshape(1, *prep_image.shape))
score_max_index = scores.argmax()
print("index: ", score_max_index)
score_max = scores[0, 0]
score_max.backward()
print(prep_image.grad.shape)
saliency = prep_image.grad.numpy()
saliency = np.abs(saliency)
saliency = saliency.reshape(224, 224, 3)
saliency = np.max(saliency, axis=2)
print(
f"max: {np.max(saliency):.3f} min: {np.min(saliency):.3f} mean: {np.mean(saliency):.3f}"
)
# saliency, _ = torch.max(prep_image.grad.data.abs(), dim=1)
# print(saliency.shape)
# saliency = exposure.rescale_intensity(saliency, (0.0, 1.0))
saliency = exposure.equalize_hist(saliency)
print(
f"max: {np.max(saliency):.3f} min: {np.min(saliency):.3f} mean: {np.mean(saliency):.3f}"
)
plt.imshow(saliency, cmap=plt.cm.hot)
plt.axis("off")
plt.show()
def __init__(self, model_type=3, cnn_type=19, cnn_layer_test=0):
super(Classic_feat, self).__init__()
self.model_type = model_type
self.cnn_type = cnn_type
self.bn = True # BenchNorm
self.use_layer = cnn_layer_test
# select cnn_model from AlexNet, GoogLeNet, VggNet
if self.model_type == ClassicCNN.AlexNet.value:
print('use AlexNet')
self.model = models.alexnet(pretrained=True)
self.sel_feat_layer = {0, 3, 6, 12} # conv1, conv2, conv3, conv5
elif self.model_type == ClassicCNN.GoogLeNet.value:
if self.cnn_type == 1: # GoogLeNet v1
print('use GoogLeNet v1')
self.model = models.googlenet(pretrained=True)
else:
print('use Inception v3')
self.model = models.inception_v3(pretrained=True)
elif self.model_type == ClassicCNN.VggNet.value:
if cnn_type == 16:
print('use Vgg16')
self.model = models.vgg16(pretrained=True) # 31
self.sel_feat_layer = {
2, 7, 12, 19, 30
} # conv1_1, conv2_2, conv3_2, conv4_2, conv5_4
else: # use Vgg19 with bn
self.cnn_type = 19
if self.bn:
print('use Vgg19_bn')
self.model = models.vgg19_bn(pretrained=True)
self.sel_feat_layer = {
3, 10, 17, 30, 52
} # conv1, conv2, conv3, conv4, conv5
else:
print('use Vgg19')
self.model = models.vgg19(pretrained=True) # 37
self.sel_feat_layer = {
2, 7, 12, 21, 36
} # conv1, conv2, conv3, conv4, conv5
# use grad-CAM
params = list(self.model.parameters())
self.weight_softmax = np.squeeze(params[-2].data.numpy())
self.CAMs = None
self.model.eval()
if use_gpu:
print('use cuda model')
self.model = self.model.cuda()
if self.model_type != ClassicCNN.GoogLeNet.value:
self.features = nn.ModuleList(list(self.model.features)).eval()
self.classify = nn.ModuleList(list(self.model.classifier)).eval()
# test feature
if self.use_layer == 0: # run combine cnn_features
self.use_layer = -4
print('test linear combined cnn_features')
elif self.use_layer > 4: # run combine features with grad-CAM
print('test combined cnn_features with grad-CAM')
else: # test each conv layers
print('test single conv layer', self.use_layer)
def Netowork_test(test):
transform = transforms.Compose([ # [1]
transforms.Resize(256), # [2]
transforms.CenterCrop(224), # [3]
transforms.ToTensor(), # [4]
transforms.Normalize( # [5]
mean=[0.485, 0.456, 0.406], # [6]
std=[0.229, 0.224, 0.225] # [7]
)
])
img = Image.open(test)
img_t = transform(img)
batch_t = torch.unsqueeze(img_t, 0)
Alexnet = models.alexnet(True)
Alexnet.eval()
out_alexnet = Alexnet(batch_t)
VGG_16 = models.vgg16_bn(True)
VGG_16.eval()
out_Vgg_16 = VGG_16(batch_t)
Googlenet = models.googlenet(True)
Googlenet.eval()
out_googlenet = Googlenet(batch_t)
# Inception = models.inception_v3(True)
# Inception.eval()
# out_inception = Inception(batch_t)
test_app = "\n" + test + "\n"
with open('imagenet_classes.txt') as f:
classes = [line.strip() for line in f.readlines()]
print("Alex-Net")
_, indices = torch.sort(out_alexnet, descending=True)
percentage = torch.nn.functional.softmax(out_alexnet, dim=1)[0] * 100
# [print (classes[idx], percentage[idx].item()) for idx in indices[0][:5]]
file1 = open("Alexnet_without_prob.txt", "a+")
file2 = open("Alexnet.txt", "a+")
file1.write(test_app)
file2.write(test_app)
for idx in indices[0][:5]:
x, y = classes[idx], percentage[idx].item()
file1.writelines("__" + str(x))
file2.writelines(str(x) + str(y) + "\n")
file1.close()
file2.close()
print("VGG-16")
_, indices = torch.sort(out_Vgg_16, descending=True)
percentage = torch.nn.functional.softmax(out_Vgg_16, dim=1)[0] * 100
# [print(classes[idx], percentage[idx].item()) for idx in indices[0][:5]]
file1 = open("VGG_16_without_prob.txt", "a+")
file2 = open("VGG_16.txt", "a+")
file1.write(test_app)
file2.write(test_app)
for idx in indices[0][:5]:
x, y = classes[idx], percentage[idx].item()
file1.writelines("__" + str(x))
file2.writelines(str(x) + str(y) + "\n")
file1.close()
file2.close()
print("Googlenet")
_, indices = torch.sort(out_googlenet, descending=True)
percentage = torch.nn.functional.softmax(out_googlenet, dim=1)[0] * 100
# [print(classes[idx], percentage[idx].item()) for idx in indices[0][:5]]
file1 = open("GoogleNet_without_prob.txt", "a+")
file2 = open("GoogleNet.txt", "a+")
file1.write(test_app)
file2.write(test_app)
for idx in indices[0][:5]:
x, y = classes[idx], percentage[idx].item()
file1.writelines("__" + str(x))
file2.writelines(str(x) + str(y) + "\n")
file1.close()
file2.close()
transforms.CenterCrop(227),
transforms.ToTensor()
]),
}
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
data_transforms[x])
for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=20,
shuffle=True, num_workers=4)
for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes
print("class_names", class_names)
model_ft = models.googlenet(pretrained=False,num_classes=2)
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
model_ft = nn.DataParallel(model_ft)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
# Observe that all parameters are being optimized
# optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
optimizer_ft = optim.Adam(model_ft.parameters())
# Decay LR by a factor of 0.1 every 7 epochs
BATCH_SIZE = 10
EPOCH = 5
LR = 1e-3
full_dataset = dataset("traindata.h5")
train_size = int(len(full_dataset)*0.8)
valid_size = len(full_dataset) - train_size
train_dataset, valid_dataset = torch.utils.data.random_split(full_dataset, [train_size, valid_size])
print("Train Data Size:", len(train_dataset))
print("Valid Data Size:", len(valid_dataset))
train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, num_workers=0, shuffle=True)
valid_loader = DataLoader(valid_dataset, batch_size=1, num_workers=0, shuffle=False)
vgg16 = models.vgg16(pretrained=True)
vgg16.eval()
googlenet = models.googlenet(pretrained=True)
googlenet.eval()
model = "vgg"
if model == "vgg":
net = MLFFNN([25088, 4096, 1024])
elif model == "googlenet":
net = MLFFNN([1024, 512, 128])
criterion = nn.CrossEntropyLoss()
mse = nn.MSELoss()
optimizer = optim.Adam(net.parameters(), lr=LR)
net.to(DEVICE)
googlenet.to(DEVICE)
def __init__(self, sample_nums=16):
super(SingleObjRegressor, self).__init__()
self.backbone = models.googlenet(pretrained=True)
self.sample_nums = sample_nums
def main():
m_Wts_preload = None #"recentPreWeights" # None/ filename of saved weights
b_size = 192 # batch size
tb_comment = "_1GN_E40_128_Ad_WM_IN_lr-15_b1-0.92_LWts(samSize)_60_epochs" # comment for tensorboard to save file
data_dir = "data" # data directory
num_epoch = 60 # num epochs
l_rate = 0.001 #learning rate
lr_step = 15 #decay of l_rate after these steps
w_decay = 0.001 #Adam
betas = (0.92, 0.999) #Adam
momentum = 0.9 # SGD momentum
weightz = [1558 / 9247,
7689 / 9247] #weights for loss function as per sample size
statistics_file = "meanAndSD.npy"
#data transformations and augmentation
data_transforms = {
'train':
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'val':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'test':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}
image_datasets = {
x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x])
for x in ["train", "val", "test"]
}
dataloaders = {
x: torch.utils.data.DataLoader(
image_datasets[x],
batch_size=b_size,
shuffle=True,
num_workers=4,
pin_memory=True,
)
for x in ["train", "val", "test"]
}
dataset_sizes = {
x: len(image_datasets[x])
for x in ["train", "val", "test"]
}
class_names = image_datasets["train"].classes
#choose GPU if available
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def imshow(inp, title=None):
"""Imshow for Tensor."""
inp = inp.numpy().transpose((1, 2, 0))
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
inp = std * inp + mean
inp = np.clip(inp, 0, 1)
plt.imshow(inp)
if title is not None:
plt.title(title)
plt.show()
# plt.pause(0.001) # pause a bit so that plots are updated
# Get a batch of training data
inputs, classes = next(iter(dataloaders["train"]))
# Make a grid from batch
out = torchvision.utils.make_grid(inputs)
imshow(out, title=[class_names[x] for x in classes])
# writer object for Tensorboard
writer = SummaryWriter(comment=tb_comment)
#function to plot confusion matrix from a confusion matrix
def plot_confusion_matrix(cm,
target_names,
title="Confusion matrix",
cmap=None,
normalize=True):
accuracy = np.trace(cm) / float(np.sum(cm))
misclass = 1 - accuracy
if cmap is None:
cmap = plt.get_cmap("Blues")
fig = plt.figure(figsize=(8, 6))
plt.imshow(cm, interpolation="nearest", cmap=cmap)
plt.title(title)
plt.colorbar()
if target_names is not None:
tick_marks = np.arange(len(target_names))
plt.xticks(tick_marks, target_names, rotation=45)
plt.yticks(tick_marks, target_names)
if normalize:
cm = cm.astype("float") / cm.sum(axis=1)[:, np.newaxis]
thresh = cm.max() / 1.5 if normalize else cm.max() / 2
for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
if normalize:
plt.text(
j,
i,
"{:0.4f}".format(cm[i, j]),
horizontalalignment="center",
color="white" if cm[i, j] > thresh else "black",
)
else:
plt.text(
j,
i,
"{:,}".format(cm[i, j]),
horizontalalignment="center",
color="white" if cm[i, j] > thresh else "black",
)
plt.tight_layout()
plt.ylabel("True label")
plt.xlabel(
"Predicted label\naccuracy={:0.4f}; misclass={:0.4f}".format(
accuracy, misclass))
writer.add_figure("Confusion_Matrix",
fig,
global_step=None,
close=True,
walltime=None)
plt.show()
def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
since = time.time()
#to load pre-trained weights
if m_Wts_preload is not None:
print("Loading locally pre-trained weights...")
model.load_state_dict(torch.load(m_Wts_preload))
bestPre_model_wts = copy.deepcopy(model.state_dict())
bestRec_model_wts = copy.deepcopy(model.state_dict())
best_precision = 0.0
best_recall = 0.0
for epoch in range(num_epochs):
print("Epoch {}/{}".format(epoch, num_epochs - 1))
print("-" * 10)
# Each epoch has a training and validation phase
for phase in ["train", "val", "test"]:
epoch_preds = None
epoch_labels = None
if phase == "train":
model.train() # Set model to training mode
else:
model.eval() # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0
# Iterate over data.
for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward
# track history if only in train
with torch.set_grad_enabled(phase == "train"):
outputs = model(inputs)
# print(outputs.shape)s
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
# backward + optimize only if in training phase
if phase == "train":
loss.backward()
optimizer.step()
# statistics
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
if epoch_labels is None:
epoch_labels = labels.cpu().numpy()
else:
epoch_labels = np.concatenate(
(epoch_labels, labels.cpu().numpy()))
# epoch_labels.append(label)
if epoch_preds is None:
epoch_preds = preds.cpu().numpy()
else:
epoch_preds = np.concatenate(
(epoch_preds, preds.cpu().numpy()))
# epoch_preds.append(preds)
if phase == "train":
scheduler.step()
# calculation of evaluation metrics
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = running_corrects.double() / dataset_sizes[phase]
epoch_prec = precision_score(epoch_labels,
epoch_preds,
average="weighted")
epoch_rec = recall_score(epoch_labels,
epoch_preds,
average="weighted")
epoch_f1 = f1_score(epoch_labels,
epoch_preds,
average="weighted")
#plot confusion matrix for validation and test sets only
if phase in ["val", "test"] and epoch == num_epochs - 1:
cm = confusion_matrix(epoch_labels,
epoch_preds,
labels=None,
normalize=None)
plot_confusion_matrix(
cm,
target_names=["notPotH", "potH"],
title="Confusion Matrix " + phase + " set",
cmap=None,
normalize=False,
)
plot_confusion_matrix(
cm,
target_names=["notPotH", "potH"],
title="Normalized Confusion Matrix " + phase + " set",
cmap=None,
normalize=True,
)
#write to tensorboard
writer.add_scalars("loss", {phase: epoch_loss}, epoch)
writer.add_scalars("precision", {phase: epoch_prec}, epoch)
writer.add_scalars("recall", {phase: epoch_rec}, epoch)
writer.add_scalars("f1_score", {phase: epoch_f1}, epoch)
writer.flush()
print(
"{} Loss: {:.4f} Acc: {:.4f} Ep_Prec: {:.4f} Ep_Rec: {:.4f} Ep_F1Score: {:.4f}"
.format(phase, epoch_loss, epoch_acc, epoch_prec,
epoch_rec, epoch_f1))
# deep copy the model
if phase == "val" and epoch_prec > best_precision:
best_precision = epoch_prec
bestPre_model_wts = copy.deepcopy(model.state_dict())
if phase == "val" and epoch_rec > best_recall:
best_recall = epoch_rec
bestRec_model_wts = copy.deepcopy(model.state_dict())
if (epoch % 3 == 0 or epoch
== num_epoch): # save model weights every 3rd epoch
torch.save(model.state_dict(), "recentWtz")
print("Time Elapsed:", (time.time() - since) / 60, " mins")
print()
time_elapsed = time.time() - since
print("Training complete in {:.0f}m {:.0f}s".format(
time_elapsed // 60, time_elapsed % 60))
print("Best val Prec: {:4f} Rec: {:4f}".format(best_precision,
best_recall))
torch.save(bestPre_model_wts, "recentPreWeights")
torch.save(bestRec_model_wts, "recentRecWeights")
return model
def visualize_model(model, num_images=6):
was_training = model.training
model.eval()
images_so_far = 0
fig = plt.figure()
with torch.no_grad():
for i, (inputs, labels) in enumerate(dataloaders["val"]):
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
for j in range(inputs.size()[0]):
images_so_far += 1
ax = plt.subplot(num_images // 2, 2, images_so_far)
ax.axis("off")
ax.set_title("predicted: {}".format(class_names[preds[j]]))
imshow(inputs.cpu().data[j])
if images_so_far == num_images:
model.train(mode=was_training)
return
model.train(mode=was_training)
model_ft = models.googlenet(pretrained=True)
#model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 2.
# Alternatively, it can be generalized to nn.Linear(num_ftrs, len(class_names)).
model_ft.aux_logits = False
model_ft.fc = nn.Linear(num_ftrs, 2)
model_ft = model_ft.to(device)
# print(torch.from_numpy(np.float32(weightz)))
criterion = nn.CrossEntropyLoss(
weight=torch.from_numpy(np.float32(weightz)).to(device))
# Observe that all parameters are being optimized
# optimizer_ft = optim.SGD(model_ft.parameters(), lr=l_rate, momentum=momentum)
optimizer_ft = optim.Adam(model_ft.parameters(),
lr=l_rate,
betas=betas,
weight_decay=w_decay)
# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft,
step_size=lr_step,
gamma=0.1)
model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
num_epoch)
torch.save(model_ft.state_dict(), "recentWtzz")
visualize_model(model_ft)
writer.close()
def __init__(self, *args, **kwargs):
super().__init__()
self.num_classes = kwargs['num_classes']
self.batch_size = kwargs['batch_size'] # 64
# Pretrained image model
self.inception = models.googlenet(pretrained=True, aux_logits=True)
# set_parameter_requires_grad(model=self.inception, feature_extracting=True)
# Handle the auxilary net
self.num_ftrs = self.inception.aux1.fc2.in_features
self.inception.aux1.fc2 = nn.Linear(self.num_ftrs, self.num_classes)
self.num_ftrs = self.inception.aux2.fc2.in_features
self.inception.aux2.fc2 = nn.Linear(self.num_ftrs, self.num_classes)
# Handle the primary net
self.num_ftrs = self.inception.fc.in_features
print('self.num_ftrs: {}'.format(self.num_ftrs))
self.inception.fc = nn.Linear(self.num_ftrs, self.num_classes)
# Return features before fc layer.
self.inception.fc = nn.Identity()
# print('self.inception:\n{}'.format(self.inception))
self.image_size = 224
# Text model
self.vocab_size = kwargs['vocab_size']
self.embedding_dim = kwargs['embedding_dim'] # 50
# self.glove_embedding = kwargs['glove_embedding']
# # self.word_embedding = nn.Embedding(num_embeddings=self.vocab_size,
# # embedding_dim=self.embedding_dim)
# # self.word_embedding.load_state_dict({'weight': self.glove_embedding})
# self.word_embedding = nn.Embedding.from_pretrained(
# embeddings=self.glove_embedding)
params_model = {
'bsize': self.batch_size,
'word_emb_dim': self.embedding_dim,
'enc_lstm_dim': 2048,
'pool_type': 'max',
'dpout_model': 0.0,
'version': 1
}
self.infersent = InferSent(params_model)
self.infersent.load_state_dict(
torch.load(
os.path.join(os.getcwd(), '../data/encoder/infersent1.pkl')))
self.infersent.set_w2v_path(w2v_path=os.path.join(
os.getcwd(), '../data/glove/glove.840B.300d.txt'))
self.infersent.build_vocab_k_words(K=self.vocab_size)
print('self.infersent:\n{}'.format(self.infersent))
# LSTM
self.hidden_size = 1024
# self.lstm = nn.LSTM(input_size=4096,
# hidden_size=self.hidden_size, num_layers=1)
# print('self.lstm:\n{}'.format(self.lstm))
# Fully connected layers
self.fc_size = 512
self.encode_dim = 4096
self.fc1 = nn.Linear(in_features=self.num_ftrs + self.encode_dim,
out_features=self.fc_size)
self.fc1_bn = nn.BatchNorm1d(num_features=self.fc_size)
self.fc2 = nn.Linear(in_features=self.fc_size,
out_features=self.num_classes)
print('self.fc1:\n{}'.format(self.fc1))
print('self.fc2:\n{}'.format(self.fc2))
def main(train_loader, train_datasets, valid_loader):
'''args: model_arch, train_loader, train_datasets, valid_loader, epochs, learning_rate, hidden_units, device, save_dir
saves checkpoint under save_dir
returns: checkpoint with {'epoch_tot',
'model',
'criterion',
'optimizer',
'optimizer.state_dict',
'model.state_dict',
'model.class_to_idx'}'''
arg = input_args()
model_arch = arg.arch
hidden_units = arg.hidden_units
learning_rate = arg.learning_rate
device = arg.device
epochs = arg.epochs
save_dir = arg.save_dir
if model_arch == 'alexnet':
gs_vgg = models.alexnet(pretrained=True)
elif model_arch == 'vgg11':
gs_vgg = models.vgg11(pretrained=True)
elif model_arch == 'vgg11_bn':
gs_vgg = models.vgg11_bn(pretrained=True)
elif model_arch == 'vgg13':
gs_vgg = models.vgg13(pretrained=True)
elif model_arch == 'vgg13_bn':
gs_vgg = models.vgg13_bn(pretrained=True)
elif model_arch == 'vgg16':
gs_vgg = models.vgg16(pretrained=True)
elif model_arch == 'vgg16_bn':
gs_vgg = models.vgg16_bn(pretrained=True)
elif model_arch == 'vgg19':
gs_vgg = models.vgg19(pretrained=True)
elif model_arch == 'vgg19_bn':
gs_vgg = models.vgg19_bn(pretrained=True)
elif model_arch == 'resnet18':
gs_vgg = models.resnet18(pretrained=True)
elif model_arch == 'resnet34':
gs_vgg = models.resnet34(pretrained=True)
elif model_arch == 'resnet50':
gs_vgg = models.resnet50(pretrained=True)
elif model_arch == 'resnet101':
gs_vgg = models.resnet101(pretrained=True)
elif model_arch == 'resnet152':
gs_vgg = models.resnet152(pretrained=True)
elif model_arch == 'squeezenet1_0':
gs_vgg = models.squeezenet1_0(pretrained=True)
elif model_arch == 'squeezenet1_1':
gs_vgg = models.squeezenet1_1(pretrained=True)
elif model_arch == 'densenet121':
gs_vgg = models.densenet121(pretrained=True)
elif model_arch == 'densenet169':
gs_vgg = models.densenet169(pretrained=True)
elif model_arch == 'densenet161':
gs_vgg = models.densenet161(pretrained=True)
elif model_arch == 'densenet201':
gs_vgg = models.densenet201(pretrained=True)
elif model_arch == 'inception_v3':
gs_vgg = models.inception_v3(pretrained=True)
elif model_arch == 'googlenet':
gs_vgg = models.googlenet(pretrained=True)
elif model_arch == 'shufflenet_v2_x0_5':
gs_vgg = models.shufflenet_v2_x0_5(pretrained=True)
elif model_arch == 'shufflenet_v2_x1_0':
gs_vgg = models.shufflenet_v2_x1_0(pretrained=True)
elif model_arch == 'shufflenet_v2_x1_5':
gs_vgg = models.shufflenet_v2_x1_5(pretrained=True)
elif model_arch == 'shufflenet_v2_x2_0':
gs_vgg = models.shufflenet_v2_x2_0(pretrained=True)
elif model_arch == 'mobilenet_v2':
gs_vgg = models.mobilenet_v2(pretrained=True)
elif model_arch == 'resnext50_32x4d':
gs_vgg = models.resnext50_32x4d(pretrained=True)
elif model_arch == 'resnext101_32x8d':
gs_vgg = models.resnext101_32x8d(pretrained=True)
elif model_arch == 'wide_resnet50_2':
gs_vgg = models.wide_resnet50_2(pretrained=True)
elif model_arch == 'wide_resnet101_2':
gs_vgg = models.wide_resnet101_2(pretrained=True)
elif model_arch == 'mnasnet0_5':
gs_vgg = models.mnasnet0_5(pretrained=True)
elif model_arch == 'mnasnet0_75':
gs_vgg = models.mnasnet0_75(pretrained=True)
elif model_arch == 'mnasnet1_0':
gs_vgg = models.mnasnet1_0(pretrained=True)
elif model_arch == 'mnasnet1_3':
gs_vgg = models.mnasnet1_3(pretrained=True)
epoch_tot = 0
for parameters in gs_vgg.parameters():
parameters.requires_grad = False
try:
input_layer = gs_vgg.classifier[0].in_features
except:
input_layer = gs_vgg.classifier.in_features
hidden_layers = [(int(hidden_units * 0.68)), (int(hidden_units * 0.32))]
output_layer = len(train_loader)
gs_vgg.classifier = nn.Sequential(
nn.Linear(input_layer, hidden_layers[0]),
nn.ReLU(),
nn.Dropout(p=0.3),
nn.Linear(hidden_layers[0], hidden_layers[1]),
nn.ReLU(),
#nn.Linear(hidden_layers[1], output_layer),
nn.Linear(hidden_layers[1], 102),
nn.LogSoftmax(dim=1))
criterion = nn.NLLLoss()
optimizer = optim.Adam(gs_vgg.classifier.parameters(), lr=learning_rate)
gs_vgg.to(device)
step_num = 0
epoch = epochs
running_loss = 0
print_every = 10
for e in range(epoch):
epoch_tot += 1
for images, labels in train_loader:
gs_vgg.train()
images, labels = images.to(device), labels.to(device)
optimizer.zero_grad()
output = gs_vgg.forward(images)
loss = criterion(output, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
step_num += 1
if step_num % print_every == 0:
test_loss = 0
accuracy = 0
gs_vgg.eval()
for images, labels in valid_loader:
images, labels = images.to(device), labels.to(device)
output = gs_vgg.forward(images)
loss = criterion(output, labels)
test_loss += loss.item()
prob = torch.exp(output)
top_p, top_class = prob.topk(1, dim=1)
equals = top_class == labels.view(*top_class.shape)
accuracy += torch.mean(equals.type(
torch.FloatTensor)).item()
print(
f"Total Epochs: {epoch_tot}.. "
f"Train loss: {running_loss/print_every:.3f}.. "
f"Test loss: {test_loss/len(valid_loader):.3f}.. "
f"Test accuracy: {(accuracy/len(valid_loader))*100:.1f}%")
running_loss = 0
gs_vgg.train()
gs_vgg.class_to_idx = train_datasets.class_to_idx
gs_checkpoint = {
'epoch_tot': epoch_tot,
'model': gs_vgg,
'criterion': criterion,
'optimizer': optimizer,
'optimizer.state_dict': optimizer.state_dict(),
'model.state_dict': gs_vgg.state_dict(),
'model.class_to_idx': gs_vgg.class_to_idx
}
torch.save(gs_checkpoint, save_dir)
return gs_checkpoint
def __init__(self,
data_dir='./data/',
model='resnet50',
save_features='False',
save_file='features',
max_pooling='False',
generate_vat='False',
resize=224):
self.data_dir = data_dir
self.model = model
self.save_features = save_features
self.save_file = save_file
self.max_pooling = max_pooling
self.generate_vat = generate_vat
self.resize = resize
self.num_images = len(os.listdir(data_dir))
# Let's get our class labels.
LABELS_URL = 'https://s3.amazonaws.com/outcome-blog/imagenet/labels.json'
response = requests.get(
LABELS_URL) # Make an HTTP GET request and store the response.
labels = {int(key): value for key, value in response.json().items()}
self.trans = transforms.Compose([
transforms.Resize((resize, resize)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if (self.model.lower() == 'alexnet'):
self.model1 = models.alexnet(pretrained=True)
elif (self.model.lower() == 'google'):
self.model1 = models.googlenet(pretrained=True)
elif (self.model.lower() == 'resnet18'):
self.model1 = models.resnet18(pretrained=True)
elif (self.model.lower() == 'resnet50'):
self.model1 = models.resnet50(pretrained=True)
elif (self.model.lower() == 'resnet101'):
self.model1 = models.resnet101(pretrained=True)
elif (self.model.lower() == 'vgg'):
self.model1 = models.vgg16(pretrained=True)
else:
print('\nInvalid model.... Exiting!\n')
exit()
self.new_model = torch.nn.Sequential(
*list(self.model1.children())[:-2])