def policy_decoder(augment, num_policy, num_op):
op_list = augment_list(False)
policies = []
for i in range(num_policy):
ops = []
for j in range(num_op):
op_idx = augment['policy_%d_%d' % (i, j)]
op_prob = augment['prob_%d_%d' % (i, j)]
op_level = augment['level_%d_%d' % (i, j)]
ops.append((op_list[op_idx][0].__name__, op_prob, op_level))
policies.append(ops)
return policies
})
del corrects, corrects_max
except StopIteration:
pass
del model
metrics = metrics / 'cnt'
# This is what it keeps track of for hyperopting.
tune.track.log(top_1_valid=metrics['correct'],
minus_loss=metrics['minus_loss'],
plus_loss=metrics['plus_loss'])
return metrics['minus_loss']
ops = augment_list(False) # Get the default augmentation set.
# Define the space of our augmentations.
space = {}
for i in range(args.num_policy):
for j in range(args.num_op):
space['policy_%d_%d' % (i, j)] = hp.choice('policy_%d_%d' % (i, j),
list(range(0, len(ops))))
space['prob_%d_%d' % (i, j)] = hp.uniform('prob_%d_%d' % (i, j), 0.0,
1.0)
space['level_%d_%d' % (i, j)] = hp.uniform('level_%d_%d' % (i, j), 0.0,
1.0)
final_policy_set = []
reward_attr = 'minus_loss'