from learner import Learner
from sklearn import model_selection as ms
from classifier import Classifier
if __name__ == '__main__':
device = torch.device("cpu")
net = Classifier('tf_efficientnet_b4_ns', 5, pretrained=True)
transform = T.Compose([
T.ToTensor(),
T.Resize((380, 380)),
T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
data_root = "/home/namnd/personal-workspace/cassava-leaf-disease-classification"
df = pd.read_csv(os.path.join(data_root, 'train.csv'))
# train_df, val_df = ms.train_test_split(df, test_size=0.2, random_state=42, stratify=df.label.values)
#
# train_dataset = CassavaLeafDiseaseDataset(data_root, df=train_df, transform=transform)
# val_dataset = CassavaLeafDiseaseDataset(data_root, df=val_df, transform=transform)
#
# train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=8, shuffle=True, num_workers=mp.cpu_count())
# val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=4, shuffle=False, num_workers=mp.cpu_count())
dataloader = CassavaLeafDiseaseDataset(data_root, df, transform=transform)
learner = Learner(net, dataloader, device)
lr_finder = LRFinder(learner)
lr_finder.find()
lr_finder.plot()
#Model: define inputs and outputs
model = Model(inputs=[in_1, in_2, in_3], outputs=out_vals) #probabilities)#
opt = optimizers.Adam(
clipnorm=1.,
lr=lrate) #remove clipnorm and add loss penalty - clipnorm works better
model.compile(loss=bin_loss, optimizer=opt)
if find_lr == True:
lr_finder = LRFinder(model)
X_train = [X1_train, X2_train, X3_train]
lr_finder.find(X_train,
y_train,
start_lr=0.00000001,
end_lr=1,
batch_size=batch_size,
epochs=1)
losses = lr_finder.losses
lrs = lr_finder.lrs
l_l = np.asarray([lrs, losses])
np.savetxt(out_dir + 'lrs_losses.txt', l_l)
num_epochs = 0
#LR schedule
class LRschedule(Callback):
'''lr scheduel according to one-cycle policy.
'''
def __init__(self, interval=1):
super(Callback, self).__init__()