model.l2_penalty: l2,
model.drop_penalty: regularize_drop_penalty
}
lstm_feed.update(char_embedding_feed)
if batch_dataset == 'A':
run_list = [train_op_A, model.loss_A]
elif batch_dataset == 'B':
run_list = [train_op_B, model.loss_B]
elif batch_dataset == 'C':
run_list = [train_op_C, model.loss_C]
_, loss = sess.run(run_list, feed_dict=lstm_feed)
return loss
test_batches_B = dp.get_batches('cdr_test', batch_size, random=False)
test_batches_C = dp.get_batches('bc_test', batch_size, random=False)
model_path = '/mnt/nfs/work1/mccallum/tbansal/nathan/disjoint-dataset/models/'
save_path = model_path + 'exp1_%d_%.1f.ckpt' % (hidden_size, args.dropout)
saver.restore(sess, save_path)
print('\nModel loaded from %s' % save_path)
print('Best model results:')
run_evaluation(test_batches_B, 'B', 'cdr_test')
run_evaluation(test_batches_C, 'C', 'bc_test')
if batch_dataset == 'A':
run_list = [train_op_A, model.loss_A]
elif batch_dataset == 'B':
run_list = [train_op_B, model.loss_B]
elif batch_dataset == 'C':
run_list = [train_op_C, model.loss_C]
_, loss = sess.run(run_list, feed_dict=lstm_feed)
return loss
epochs = 6000
train_batches_A = dp.get_batches('cdr_train_weak', batch_size, random=False)
train_batches_B = dp.get_batches('cdr_train_gold', batch_size, random=False)
train_batches_C = dp.get_batches('bc_train', batch_size, random=False)
dev_batches_B = dp.get_batches('cdr_dev', batch_size, random=False)
dev_batches_C = dp.get_batches('bc_dev', batch_size, random=False)
test_batches_B = dp.get_batches('cdr_test', batch_size, random=False)
test_batches_C = dp.get_batches('bc_test', batch_size, random=False)
max_f1 = 0
epoch_last_saved = 0
last_saved = 'N/A'
for i in range(epochs):
batches_A = dp.get_batches('cdr_train_weak', batch_size)
batches_B = dp.get_batches('cdr_train_gold', batch_size)
model.input_dropout_keep_prob: 1 - input_dropout,
model.l2_penalty: l2,
model.drop_penalty: regularize_drop_penalty
}
lstm_feed.update(char_embedding_feed)
if A:
_, loss = sess.run([train_op_A, model.loss_A], feed_dict=lstm_feed)
else:
_, loss = sess.run([train_op_B, model.loss_B], feed_dict=lstm_feed)
return loss
epochs = 6000
dev_batches_A = dp.get_batches('cdr_dev', batch_size, random=False)
dev_batches_B = dp.get_batches('bc_dev', batch_size, random=False)
max_f1 = 0
last_saved = 'N/A'
for i in range(epochs):
batches_A = dp.get_batches('cdr_train_gold', batch_size) + dp.get_batches(
'cdr_train_weak', batch_size)
np.random.shuffle(batches_A)
batches_B = dp.get_batches('bc_train', batch_size)
if dataset == 'joint':
l = min(len(batches_A), len(batches_B))
elif dataset == 'A':
l = len(batches_A)
elif dataset == 'B':
l = len(batches_B)
model.input_dropout_keep_prob: 1 - input_dropout,
model.l2_penalty: l2,
model.drop_penalty: regularize_drop_penalty
}
lstm_feed.update(char_embedding_feed)
if A:
_, loss = sess.run([train_op_A, model.loss_A], feed_dict=lstm_feed)
else:
_, loss = sess.run([train_op_B, model.loss_B], feed_dict=lstm_feed)
return loss
epochs = 600
dev_batches_A = dp.get_batches('cdr_dev', batch_size, random=False)
dev_batches_B = dp.get_batches('bc_dev', batch_size, random=False)
max_f1 = 0
last_saved = 0
for i in range(epochs):
f1_A, _ = run_evaluation(dev_batches_A,
'cdr_dev (iteration %d)' % i,
A=True)
f1_B, _ = run_evaluation(dev_batches_B,
'bc_dev (iteration %d)' % i,
A=False)
# if np.mean((f1_A, f1_B)) > max_f1:
if f1_B > max_f1:
model_path = 'modelpath/bc/'
save_path = saver.save(
if batch_dataset == 'A':
run_list = [train_op_A, model.loss_A]
elif batch_dataset == 'B':
run_list = [train_op_B, model.loss_B]
elif batch_dataset == 'C':
run_list = [train_op_C, model.loss_C]
_, loss = sess.run(run_list, feed_dict=lstm_feed)
return loss
epochs = 6000
train_batches_A = dp.get_batches('A_train', batch_size, random=False)
train_batches_B = dp.get_batches('B_train', batch_size, random=False)
dev_batches = dp.get_batches('dev', batch_size, random=False)
test_batches = dp.get_batches('test', batch_size, random=False)
max_f1 = 0
epoch_last_saved = 0
last_saved = 'N/A'
for i in range(epochs):
batches_A = dp.get_batches('A_train', batch_size)
batches_B = dp.get_batches('B_train', batch_size)
l = min(len(batches_A), len(batches_B))
model.middle_dropout_keep_prob: 1 - middle_dropout,
model.input_dropout_keep_prob: 1 - input_dropout,
model.l2_penalty: l2,
model.drop_penalty: regularize_drop_penalty
}
lstm_feed.update(char_embedding_feed)
if batch_dataset == 'A':
run_list = [train_op_A, model.loss_A]
elif batch_dataset == 'B':
run_list = [train_op_B, model.loss_B]
elif batch_dataset == 'C':
run_list = [train_op_C, model.loss_C]
_, loss = sess.run(run_list, feed_dict=lstm_feed)
return loss
test_batches = dp.get_batches('test', batch_size, random=False)
model_path = 'modelpath/3_sets/'
save_path = model_path + 'exp_split_%d_%.1f.ckpt' % (hidden_size, args.dropout)
saver.restore(sess, save_path)
print('\nModel loaded from %s' % save_path)
print('Best model results:')
run_evaluation(test_batches, 'full', 'test')