def configure_next_level(lvl: int, depth: int, budget: int = 50):
new_exp_key = f'covid-{lvl}'
src_trials = MongoTrials('mongo://localhost:1234/covid/jobs',
exp_key=f'covid-{lvl-1}')
all_trials = MongoTrials('mongo://localhost:1234/covid/jobs')
dest_trials = MongoTrials('mongo://localhost:1234/covid/jobs',
exp_key=new_exp_key)
hist_length = {2: 3, 3: 5, 5: 8}.get(depth, 10)
forward_losses = []
for trial, loss in zip(src_trials.trials, src_trials.losses()):
if loss is None:
forward_losses.append(None)
continue
v_x, vloss, _, _ = zip(*trial['result']['validation_stats'])
slope, intercept, _, _, _ = linregress(v_x[-hist_length:],
vloss[-hist_length:])
forward_losses.append(
min(
0.5 * (loss + intercept + slope * v_x[-1] + slope *
(1 - 0.8**(depth - v_x[-1])) / (1 - 0.8)), loss))
ordered_idxs = list(
np.argsort([x if x is not None else np.inf for x in forward_losses]))
last_tid = 0 if len(all_trials.tids) == 0 else max(all_trials.tids)
available_tids = []
result_docs = []
while len(ordered_idxs) > 0:
idx = ordered_idxs.pop(0)
if src_trials.losses()[idx] is None:
continue
epochs_completed = src_trials.trials[idx]['result'].get(
'training_loss_hist', [(0, np.inf)])[-1][0]
spec = None
result = {'status': 'new'}
misc = copy.deepcopy(src_trials.trials[idx]['misc'])
result_docs.append((spec, result, misc))
budget -= (depth - epochs_completed)
if budget <= 0:
break
while len(ordered_idxs) > 0:
idx = ordered_idxs.pop()
if src_trials.losses()[idx] is None:
continue
if len(available_tids) == 0:
available_tids = dest_trials.new_trial_ids(last_tid)
tid = available_tids.pop(0)
last_tid = tid
# copy in the ones that aren't worth exploring further
cpy = copy.deepcopy(src_trials.trials[idx])
cpy['exp_key'] = new_exp_key
cpy['tid'] = tid
cpy['misc']['tid'] = tid
cpy['misc']['idxs'] = {k: [tid] for k in cpy['misc']['idxs'].keys()}
del cpy['_id']
dest_trials.insert_trial_doc(cpy)
return result_docs
