def stance_fine_tune():
# importing the required module
import matplotlib.pyplot as plt
for lr in [1e-3, 5e-4, 2e-4, 1e-4]:
hp = HPFineTunePair()
topic = "hillary"
e = Experiment(hp)
hp.lr = lr
hp.num_epochs = 100
preload_id = ("LM_pair_tweets_hillary_run2", 1247707)
setting = shared_setting.TopicTweets2Stance(topic)
stance_data = stance_detection.FineLoader(topic, hp.seq_max,
setting.vocab_filename,
hp.sent_max)
valid_history = e.train_stance(setting.vocab_size, stance_data,
preload_id)
e.clear_run()
l_acc, l_f1 = zip(*valid_history)
plt.plot(l_acc, label="{} / ACC".format(lr))
plt.plot(l_f1, label="{} / F1".format(lr))
plt.legend(loc='lower right')
# giving a title to my graph
plt.title('learning rate - dev !')
# function to show the plot
plt.show()
def predict_rf_tune():
hp = hyperparams.HPBert()
hp.batch_size = 256
e = Experiment(hp)
nli_setting = NLI()
nli_setting.vocab_size = 30522
nli_setting.vocab_filename = "bert_voca.txt"
target_label = 'match'
#data_id = 'test_conflict'
data_id = "{}_1000".format(target_label)
e_config = ExperimentConfig()
l = [(0.1, 12039), (0.2, 12245), (0.3, 12063), (0.4, 12250), (0.6, 12262),
(0.7, 12253)]
l = [(0.5, 12175), (0.8, 12269), (0.9, 12259)]
for del_g, step in l:
e_config.name = "X_match_del_{}".format(del_g)
e_config.load_names = ['bert', 'cls_dense', 'aux_conflict']
data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename,
True)
load_id = ("NLIEx_match_del_{}".format(del_g), "model-{}".format(step))
e.clear_run()
e.predict_rf(nli_setting, e_config, data_loader, load_id, data_id, 5)
