def restore(self, **solver_args):
"""
Restore from a previous restore point fixed via set_restore_point
If set_restore_point was not previously called, restores the
state of the solver from when it was initialised.
Optionally, one or more keyword arguments may be supplied, which
then override the arguments used to restore the solver, when
recreating the solver via solver.create .
"""
restore_args = dict(self.restore_args)
restore_args.update(solver_args)
self._last_p_0 = restore_args['p_0']
self.solver = solver.create(
self.model,
self.sink,
**restore_args
)
def create_jobs():
Net.Params = {
'model_name': job_name,
'train_lmdb': train_lmdb,
'test_lmdb': test_lmdb,
'mean_file': mean_file,
'batch_size_per_device': 128,
'test_batch_size': 50,
'num_classes': 2,
'num_test_image': 1000,
}
# train/test/deploy.prototxt
for phase in ['train', 'test', 'deploy']:
proto_file = os.path.join(job_dir, '{}.prototxt'.format(phase))
with open(proto_file, 'w') as f:
print(Net.create_net(phase), file=f)
# solver.prototxt
Solver.Params['train_net'] = os.path.join(job_dir, 'train.prototxt')
Solver.Params['test_net'] = [os.path.join(job_dir, 'test.prototxt')]
Solver.Params['snapshot_prefix'] = os.path.join(job_dir, job_name)
Solver.Params['base_lr'] = 0.01
Solver.Params['test_iter'] = [Net.get_testiter()]
Solver.Params['test_interval'] = 2500
Solver.Params['stepsize'] = 30000
Solver.Params['weight_decay'] = 0.008
Solver.Params['momentum'] = 0.9
Solver.Params['max_iter'] = 400000
Solver.Params['snapshot'] = 50000
Solver.Params['display'] = 500
Solver.Params['lr_policy'] = 'step'
solver_file = os.path.join(job_dir, 'solver.prototxt')
with open(solver_file, 'w') as f:
print(Solver.create(), file=f)