def main():
usage = "usage: %prog [option] img_path"
parser = OptionParser(usage)
options, args = parser.parse_args()
print('options', options)
print('args', args)
if len(args) != 1:
parser.error('incorrect number of arguments')
path = [args[0]]
model = get_pretrained_model('densenet161').to(device)
img_pil = list(map(pil_loader, path))
img_tensor = list(map(_preprocess, img_pil))
img_variable = torch.stack(img_tensor).to(device)
p, x = gcam(model, img_variable)
t = crop_heat(x, img_variable, threshold=0.63)
print('Saving images at current directory...')
print('orig.png ...')
img_pil[0].save('./orig.png')
print('boundbox.png ...')
Image.fromarray(
add_boundedbox(x[0], add_heatmap_ts(x[0], img_variable[0], need_transpose_color=False), threshold=0.63,
need_transpose_color=True)).save('./boundbox.png')
print('croped.png ...')
tsimg2img(t[0]).save('./croped.png')
def main():
dls = dataloader.get_fold_loaders(k, BATCH_SIZE)
for i, d in enumerate(dls):
torch.save(d, "loaders/" + str(i) + ".pth")
accs = []
norm_confmats = []
confmats = []
best_state_dict_init = torch.load("models/inits/best.pth")
for fold in range(k):
mod = model.get_pretrained_model(layer_names=setting["layers"],
type_init=setting["init"]).to(device)
mod.load_state_dict(best_state_dict_init)
optim = model.get_optimizer(mod,
feature_extract=True,
lr=setting["lr"],
mom=setting["mom"])
criterion = nn.CrossEntropyLoss()
for e in range(EPOCHS):
mod, valloss, _, confmat = traintest.trainepoch(
mod, dls[fold], criterion, optim, device)
valacc = get_acc_from_conf(confmat)
if e == EPOCHS - 1:
confmats.append(confmat)
norm_confmat = normalize(confmat)
norm_confmats.append(norm_confmat)
accs.append(valacc)
torch.save(mod.state_dict(), "models/folds/" + str(fold))
def main():
usage = "usage: %prog [option] img_path dir_path"
parser = OptionParser(usage)
parser.add_option('-d',
'--draw',
help='show the images',
action='store_true',
default=False,
dest='draw')
options, args = parser.parse_args()
print('options', options)
print('args', args)
if len(args) != 2:
parser.error('incorrect number of arguments')
dataloaders, dataset_sizes = get_dataloaders(study_name=None,
data_dir=args[1],
batch_size=30,
batch_eval_ten=15,
shuffle=False)
img_path = args[0]
model = get_pretrained_model('densenet161').to(device)
if options.draw:
findTOP5pic(model, dataloaders, img_path)
else:
findTOP5addr(model, dataloaders, img_path)
def main():
usage = 'usage: %prog [options] data_dir'
parser = OptionParser(usage)
parser.add_option(
'-p',
'--phase',
help=
'valid or train, assume the same directory structure with the training set, '
'where there should be a file named <PHASE>.csv in the data directory',
action='store',
type='string',
default='valid',
dest='phase')
parser.add_option(
'-m',
'--model',
help=
'the model to evaluate, one of [densenet161, densenet169, resnet50, vgg19, agnet]',
action='store',
type='string',
default='agnet',
dest='model')
parser.add_option('-s',
'--study',
help='for evaluating on a specific study',
action='store',
type='string',
default='all',
dest='study')
parser.add_option('-d',
'--draw',
help='draw roc curves',
action='store_true',
default=False,
dest='draw')
options, args = parser.parse_args()
print('options', options)
print('args', args)
if len(args) != 1:
parser.error('incorrect number of arguments')
data_dir = args[0]
if not options.model in common.model_names:
parser.error('Unknown model name')
if not options.study in common.study_names + ['all']:
parser.error('Unknown study name')
study_name = None
if options.study != 'all':
study_name = options.study
dataloader = any_dataloader(
options.phase,
study_name=study_name,
data_dir=args[0],
)
model = get_pretrained_model(options.model).to(device)
evaluate(model, dataloader, options.phase, options.draw)
def run(setting,
n,
save_dir,
folder,
early_stop=True,
split=0.75,
init_dict=None):
name = convert_to_name(setting)
model_save_dir = save_dir + '/'
history = {
"loss": [],
"acc": [],
"normacc": [],
"ipacc": [],
"npacc": [],
"confmat": [],
"best_avg": 0
}
mod = model.get_pretrained_model(layer_names=setting["layers"],
type_init=setting["init"]).to(device)
if init_dict is not None:
mod.load_state_dict(init_dict)
optim = model.get_optimizer(mod,
feature_extract=True,
lr=setting["lr"],
mom=setting["mom"])
criterion = nn.CrossEntropyLoss()
Path(model_save_dir + name + "/" + n).mkdir(parents=True, exist_ok=True)
torch.save(mod.state_dict(), model_save_dir + name + "/" + n + '/epoch_0')
stop = False
if early_stop:
dataloaders = dataloader.get_loaders(BATCH_SIZE, split)
while not stop:
print(stopcrit.checks)
mod, valloss, valacc, confmat = traintest.trainepoch(
mod, dataloaders, criterion, optim, device)
#normalacc, ipacc, npacc = accs_from_confmat(confmat)
history["loss"].append(valloss)
history["acc"].append(valacc)
#history["normacc"].append(normalacc)
#history["ipacc"].append(ipacc)
#history["npacc"].append(npacc)
history["confmat"].append(confmat)
stop = stopcrit.check(valacc, mod.state_dict())
else:
dataloaders = dataloader.get_loaders(BATCH_SIZE, split)
for epoch in range(EPOCHS):
if split == 1.0:
validate = False
else:
validate = True
mod, valloss, valacc, confmat = traintest.trainepoch(
mod, dataloaders, criterion, optim, device, validate)
if valloss is not None:
#normalacc, ipacc, npacc = accs_from_confmat(confmat)
history["loss"].append(valloss)
history["acc"].append(valacc)
#history["normacc"].append(normalacc)
#history["ipacc"].append(ipacc)
#history["npacc"].append(npacc)
history["confmat"].append(confmat)
stop = stopcrit.check(valacc, mod.state_dict())
if split != 1.0:
history["best_avg"] = stopcrit.last_avg
torch.save(
stopcrit.best_model_dict, model_save_dir + name + "/" + n +
'/epoch_' + str(stopcrit.best_check))
plot_run(name, n, history, folder)
best_acc = stopcrit.best_val
best_epoch = stopcrit.best_check
stopcrit.reset()
else:
torch.save(mod.state_dict(),
model_save_dir + name + "/" + n + "/epoch_" + str(EPOCHS))
best_acc = None
best_epoch = None
return history, best_acc, best_epoch