def get_model():
d121 = models.densenet121(pretrained=True)
model = SentinelDenseNet(d121, get_data().targets)
wrapper = Model(model)
return wrapper
def load_model():
wrapper, db = get_model(), get_data()
model = wrapper.model
mapper = db.targets
model.load_state_dict(torch.load(CP))
model.eval()
return model, mapper
def load_model(metric):
CP = 'single_d121_ur_phase3_bal_4_' + f'{metric}.pt'
wrapper, db = get_model(metric), get_data(metric)
model = wrapper.model
mapper = db.targets
model.load_state_dict(torch.load(CP))
model.eval()
return model, mapper
def main():
wrapper = get_model()
opt_fn = optim.Adam
cycle_fn = partial(optim.lr_scheduler.CyclicLR,
mode='triangular2',
cycle_momentum=False)
# step_fn = partial(optim.lr_scheduler.StepLR, step_size=1, gamma=0.1)
learn = Learner(wrapper)
# # IMG_SZ = 224, BS: 32
# db = get_data(img_sz=224, bs=64)
# # Freeze: Features, 3 epochs
# learn.wrapper.freeze_features()
# print('\n\nPhase 1, 224, 64', learn.wrapper.grads)
# learn.fit(2, 1e-2, db, opt_fn, cycle_fn)
# learn.fit(2, 1e-3, db, opt_fn, cycle_fn)
# torch.save(learn.wrapper.model.state_dict(), 'single_d121_ur_phase1_bal_4.pt')
# # learn.wrapper.model.load_state_dict(torch.load('single_d121_ur_phase1_bal_4.pt'))
# # Freeze: Partial(0.7), 3 epochs
# learn.wrapper.partial_freeze_features(0.7)
# print('\n\nPhase 2, 224, 64', learn.wrapper.grads)
# learn.fit(2, 1e-4, db, opt_fn, cycle_fn)
# torch.save(learn.wrapper.model.state_dict(), 'single_d121_ur_phase2_bal_4.pt')
# # learn.wrapper.model.load_state_dict(torch.load('single_d121_ur_phase2.pt'))
db = get_data(img_sz=224, bs=32)
# Freee: None, 3 Epochs
learn.wrapper.freeze_features(False)
print('\n\nPhase 3, 224, 32', learn.wrapper.grads)
learn.fit(2, 1e-4, db, opt_fn, cycle_fn)
learn.fit(2, 1e-5, db, opt_fn, cycle_fn)
torch.save(learn.wrapper.model.state_dict(),
'single_d121_ur_phase3_bal_4.pt')
# # learn.wrapper.model.load_state_dict(torch.load('single_d121_ur_phase3.pt'))
# IMG_SZ = 224, BS: 32
# db = get_data(img_sz=224, bs=32)
# print('\n\nPhase 4, 224, 32', learn.wrapper.grads)
# learn.fit(1, 0.001, db, opt_fn, cycle_fn)
# learn.fit(3, 0.1, db, opt_fn, cycle_fn)
# torch.save(learn.wrapper.model.state_dict(), 'single_d121_phase4.pt')
# learn.wrapper.model.load_state_dict(torch.load('single_d121_phase4.pt'))
# learn.wrapper.freeze_features(False)
# # IMG_SZ = 299, BS: 32
# db = get_data(img_sz=299, bs=16)
# print('\n\nPhase 4, 299, 16', learn.wrapper.grads)
# learn.fit(2, 0.00001, db, opt_fn, cycle_fn)
# torch.save(learn.wrapper.model.state_dict(), 'single_d121_ur_phase4.pt')
db = get_data(img_sz=224, bs=64)
learn.wrapper.model.load_state_dict(
torch.load('single_d121_ur_phase3_bal_4.pt'))
# Freee: None, 3 Epochs
learn.wrapper.freeze_features(True)
print('\n\nPhase 4, 224, 64', learn.wrapper.grads)
learn.fit(2, 1e-5, db, opt_fn, cycle_fn)
torch.save(learn.wrapper.model.state_dict(),
'single_d121_ur_phase3_bal_4_final.pt')