sys.path.append('../../..')
#sys.path.append('../../../../DeepRL')
sys.path.append('../../../../../transformer_pytorch')
import pickle
from generative_playground.dependency_trees.train.main_train_dependencies import train_dependencies
with open('../../ud_utils/meta.pickle','rb') as f:
meta = pickle.load(f)
batch_size = 100
drop_rate = 0.5
max_steps = meta['maxlen']
model, fitter1 = train_dependencies(EPOCHS=1000,
BATCH_SIZE=batch_size,
max_steps=max_steps,
lr=3e-5,
drop_rate=drop_rate,
decoder_type='attention',
plot_prefix='lr 3e-5 no attn vae ',
dashboard ='dependencies_vae',
#save_file='dependencies_test.h5',
use_self_attention=None,
include_predefined_embedding=True,
vae=True,
target_names=['token', 'head', 'upos', 'deprel'],
meta=meta)
#preload_file='policy_gradient_run.h5')
while True:
next(fitter1)
missing_embed_to_zeros = True
meta_filename = get_meta_filename(max_len, cutoff, missing_embed_to_zeros)
with open('../data/processed/' + meta_filename, 'rb') as f:
meta = pickle.load(f)
batch_size = 100
drop_rate = 0.05
max_steps = meta['maxlen']
model, fitter1 = train_dependencies(
EPOCHS=1000,
BATCH_SIZE=batch_size,
max_steps=max_steps,
lr=3e-5,
drop_rate=drop_rate,
decoder_type='attention',
plot_prefix='lr 3e-5 both roman',
dashboard='dependencies_novae',
#save_file='dependencies_test.h5',
include_predefined_embedding=True,
use_self_attention='both', # None, True, False or Both
vae=False,
plot_ignore_initial=0,
target_names=['head', 'upos', 'deprel'], #'token' ,
meta=meta,
languages=['ca', 'fr', 'gl', 'ro', 'es', 'pt',
'it'], #['pt'],##['en', 'de', 'fr'],
ignore_padding=True)
#preload_file='policy_gradient_run.h5')
while True:
next(fitter1)