def load_cv_folds(model_name):
models = []
for i in range(5):
net = np.load(paths.predictions + model_name +
'-split_{}.npz'.format(i))
net = {
'train': np.average(net['train'], axis=0),
'val': np.average(net['val'], axis=0),
'test': np.average(net['test'], axis=0)
}
models.append(net)
labels_df = labels.get_labels_df()
kf = sklearn.model_selection.KFold(n_splits=5,
shuffle=True,
random_state=1)
split = kf.split(labels_df)
folds = []
for i, ((train_idx, val_idx), net) in enumerate(zip(split, models)):
val = labels_df.ix[val_idx]
train = labels_df.ix[train_idx]
folds.append((net, val, train))
print(i)
return folds
def predict_kfold(model_name, pre_transforms=[]):
model = locate(model_name + '.generate_model')()
random_state = locate(model_name + '.random_state')
print('Random state: {}'.format(random_state))
labels_df = labels.get_labels_df()
kf = sklearn.model_selection.KFold(n_splits=5,
shuffle=True,
random_state=random_state)
split = kf.split(labels_df)
for i, (train_idx, val_idx) in enumerate(split):
split_name = model_name + '-split_' + str(i)
best_epoch = util.find_epoch_val(split_name)
print('Using epoch {} for predictions'.format(best_epoch))
epoch_name = split_name + '-epoch_' + str(best_epoch)
train = labels_df.ix[train_idx]
val = labels_df.ix[val_idx]
state = torch.load(os.path.join(paths.models, split_name, epoch_name))
predict_model(model,
state,
train,
val,
output_file=split_name,
pre_transforms=pre_transforms)
pretrained_model.avgpool = nn.AvgPool2d(8)
classifier = [
nn.Linear(pretrained_model.fc.in_features, 17),
]
self.classifier = nn.Sequential(*classifier)
pretrained_model.fc = self.classifier
def forward(self, x):
return F.sigmoid(self.pretrained_model(x))
return MyModel(torchvision.models.resnet50(pretrained=True))
random_state = 1
labels_df = labels.get_labels_df()
kf = sklearn.model_selection.KFold(n_splits=5, shuffle=True, random_state=random_state)
split = kf.split(labels_df)
def train_net(train, val, model, name):
transformations_train = transforms.apply_chain([
transforms.random_fliplr(),
transforms.random_flipud(),
transforms.augment(),
torchvision.transforms.ToTensor()
])
transformations_val = transforms.apply_chain([
torchvision.transforms.ToTensor(),
])