def train_lm(n_cycles: int = 3,
cycle_len: int = 1,
cycle_mult: int = 2,
momentum: float = 0.8,
bptt: int = 40,
lr: float = 1e-3,
wd: float = 1e-7):
datasets = create_or_restore(DATA_ROOT)
lm_data = [
fastai_patch(ds)
for ds in (datasets['train_unsup'], datasets['test_unsup'])
]
bunch = TextLMDataBunch.create(lm_data, path=LM_PATH, bptt=bptt)
n = sum(len(ds) for ds in lm_data)
phases = [
TrainingPhase(n * (cycle_len * cycle_mult**i),
lrs=lr,
moms=momentum,
lr_anneal=annealing_cos) for i in range(n_cycles)
]
learner = RNNLearner.language_model(bunch, bptt)
cbs = [
EarlyStopping(learner, patience=2),
GeneralScheduler(learner, phases),
SaveModel(learner)
]
if cycle_mult == 1:
total_epochs = n_cycles * cycle_len
else:
total_epochs = int(cycle_len * (1 - cycle_mult**n_cycles) /
(1 - cycle_mult))
print(f'Total number of epochs: {total_epochs:d}')
try:
learner.fit(total_epochs, wd=wd, callbacks=cbs)
except RuntimeError as e:
print(f'Model training error: {e}')
finally:
folder = learner.path / learner.model_dir
print(f'Saving latest model state into {folder}')
learner.save('lm_final')
learner.save_encoder('lm_final_enc')
'PersonalStories', 'PossiblyFeedback', 'SentimentNegative',
'SentimentNeutral', 'SentimentPositive'
]
fold = str(args.fold)
cat = all_cats[args.cl]
model_id = '2019_ 4_04_20_02_39_766228'
model_id = '2019_ 4_05_02_05_04_069084'
model_id = '2019_ 4_05_11_54_23_327097'
model_id = '2019_ 4_06_18_53_42_028549'
model_id = '2019_ 4_07_20_36_23_662822'
model_id = '2019_ 4_07_20_35_19_533996'
model_id = '2019_ 4_07_23_55_08_571313'
exp_path = '/mnt/data/group07/johannes/ompc/ppexp_short4/' + cat + '_' + fold
data_lm_ft = TextLMDataBunch.load(
Path('/mnt/data/group07/johannes/ompc/pplmexp_short4'))
if True or not Path('/mnt/data/group07/johannes/ompc/ppexp_short4/' + cat +
'_' + fold + '/models/enc5.pth').is_file():
if True or not Path(
'/mnt/data/group07/johannes/ompc/pplmexp_short/models/enc5.pth'
).is_file():
print('need to save enc')
learn_lm = language_model_learner(data_lm_ft)
learn_lm.load(model_id)
learn_lm.save_encoder('enc5')
del learn_lm
print('need to copy enc')
os.makedirs(
os.path.dirname('/mnt/data/group07/johannes/ompc/ppexp_short4/' + cat +
'_' + fold + '/models/enc5.pth'),
import news_utils.fastai
import news_utils.clean.german
# In[ ]:
parser = argparse.ArgumentParser()
parser.add_argument("--cl", type=int)
parser.add_argument("--fold", type=int)
args = parser.parse_args()
bpemb_de = BPEmb(lang="de", vs=25000, dim=300)
data_lm_ft = TextLMDataBunch.load(path='/mnt/data/group07/johannes/ompc/lmexp', cache_name='whatever')
all_cats = ['ArgumentsUsed', 'Discriminating', 'Inappropriate', 'OffTopic', 'PersonalStories', 'PossiblyFeedback', 'SentimentNegative', 'SentimentNeutral', 'SentimentPositive']
fold = str(args.fold)
cat = all_cats[args.cl]
model_id = '2019_ 4_01_00_11_32_066215'
exp_path = '/mnt/data/group07/johannes/ompc/exp/' + cat + '_' + fold
os.makedirs(os.path.dirname('/mnt/data/group07/johannes/ompc/exp/' + cat + '_' + fold +'/models/enc5.pth'), exist_ok=True)
shutil.copy('/mnt/data/group07/johannes/ompc/lmexp/models/enc5.pth', '/mnt/data/group07/johannes/ompc/exp/' + cat + '_' + fold +'/models/enc5.pth')
def run_for_class(it=1):
train_df = pd.read_pickle(Path('/mnt/data/group07/johannes/ompc/data_ann')/cat/fold/'train.pkl')
test_df = pd.read_pickle(Path('/mnt/data/group07/johannes/ompc/data_ann')/cat/fold/'test.pkl')