def data_gen(input_dir, output_dir):
output_direc = merging_datafile(input_dir, output_dir)
input_tensor, target_tensor, inp_lang, targ_lang = load_dataset(
output_direc)
# Calculate max_length of the target tensors
max_length_targ, max_length_inp = target_tensor.shape[
1], input_tensor.shape[1]
input_tensor_train, input_tensor_val, target_tensor_train, target_tensor_val = train_test_split(
input_tensor, target_tensor, test_size=0.2)
# Show length
print(len(input_tensor_train), len(target_tensor_train),
len(input_tensor_val), len(target_tensor_val))
print("Input Language; index to word mapping")
convert(inp_lang, input_tensor_train[0])
print()
print("Target Language; index to word mapping")
convert(targ_lang, target_tensor_train[0])
buffer_size = len(input_tensor_train)
batch_size = 16
steps_per_epoch = len(input_tensor_train) // batch_size
embedding_dim = 256
units = 1024
vocab_inp_size = len(inp_lang.word_index) + 1
vocab_tar_size = len(targ_lang.word_index) + 1
dataset = tf.data.Dataset.from_tensor_slices(
(input_tensor_train, target_tensor_train)).shuffle(buffer_size)
dataset = dataset.batch(batch_size, drop_remainder=True)
example_input_batch, example_target_batch = next(iter(dataset))
return dataset, vocab_inp_size, vocab_tar_size, embedding_dim, units, batch_size, example_input_batch, steps_per_epoch, targ_lang, max_length_targ, max_length_inp, inp_lang, targ_lang
os.path.basename(train_tarball), n_units)))
else:
model.intersect_word2vec_format(pre_trained_path, binary=True)
model.train(sentences,
total_words=None,
epochs=model.iter,
total_examples=model.corpus_count,
queue_factor=2,
report_delay=report_delay)
model.save(
os.path.join(
out_folder, 'finetuned_gigaword_{}_{}.model'.format(
os.path.basename(train_tarball), n_units)))
if bins:
vocab, pos2id, n_classes, n_participants, train, val, test = pd.load_dataset(
model.wv.vocab, gensim=True, bins=True)
else:
vocab, pos2id, train, val, test, mean, std = pd.load_dataset(
model.wv.vocab, gensim=True)
print('Data samples eyetracking: %d' % len(train))
train_iter = EyetrackingBatchIterator(train,
window_eyetracking,
batchsize_eyetracking,
repeat=True,
shuffle=True,
wlen=wlen,
pos=pos,
prev_fix=prev_fix,
freq=freq,
elif incorrect_visualized_counter < visualize_incorrect and corrects_vector[i] == False:
visualize.visualize_with_correct(batch[inputs_placeholder][i], predictions[i], true_label,
name + "_incorrect")
incorrect_visualized_counter += 1
precision = correct_num / number_of_examples
summary = tf.Summary()
summary.value.add(tag='Accuracy_' + name, simple_value=precision)
summary_writer.add_summary(summary, learning_step)
print("Accuracy %.3f" % precision)
if __name__ == '__main__':
# Load dataset
train = load_dataset("train.p", False)
validation = load_dataset("validation.p", False)
test = load_dataset("test.p", False)
with tf.Graph().as_default():
# Wiring
# model = Sequence()
# model = SequenceBiggerOutput()
# model = SequenceReshapedConvolution()
model = SequenceReshapedConvolutionBatchnorm()
# model = SequenceReshapedConvolutionDeeper()
inputs_placeholder, labels_placeholder, keep_prob_placeholder, is_training_placeholder = model.input_placeholders()
logits = model.inference(inputs_placeholder, keep_prob_placeholder, is_training_placeholder)
loss = model.loss(logits, labels_placeholder)
training = model.training(loss, 0.0001)
num_workers = args.num_workers
print('dataset: ', dataset_name)
print('model: ', model_name)
print('pretrained: ', pretrained)
print('epochs: ', epochs)
print('batch_size: ', batch_size)
print('lr: ', lr)
print('seed: ', seed)
print('num_workers: ', num_workers)
print('-' * 30)
torch.manual_seed(seed)
# load dataset
train_data, test_data, in_channels, num_classes = load_dataset(
dataset_name)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
# load model
model = getattr(target_models, model_name)(in_channels, num_classes)
model.cuda()
optimizer = optim.SGD(model.parameters(),
lr=lr,
chainer.cuda.get_device_from_id(args.gpu).use()
cuda.check_cuda_available()
print('GPU: {}'.format(args.gpu))
print('# unit: {}'.format(args.unit))
#print('Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('Output type: {}'.format(args.out_type))
assert (args.surprisal_order in range(2, 6))
if args.gpu >= 0:
cuda.get_device_from_id(args.gpu).use()
if args.bins:
vocab, pos2id, n_classes, n_participants, train, val, test = pd.load_dataset(
bins=True, surprisal_order=args.surprisal_order)
else:
vocab, pos2id, train, val, test, mean, std = pd.load_dataset(
surprisal_order=args.surprisal_order)
print('')
print('Mean dataset times: {}'.format(mean))
print('Std_dev dataset times: {}'.format(std))
index2word = {v: k for k, v in vocab.items()}
n_vocab = len(vocab)
n_pos = len(pos2id)
print('n_vocab: %d' % n_vocab)
print('data length: %d' % len(train))
print('n_pos: %d' % n_pos)
pos = True
prev_fix = True
freq = True
surprisal = True
n_pos_units = 50
out = F.tanh
n_hidden = 200
configs = [(False, 0, 0), (False, 1, 0), (False, 0, 2),
(True, 0, 0), (True, 1, 0), (True, 0, 2)]
for config in configs:
bins, window, n_layers = config
if bins:
vocab, pos2id, n_classes, n_participants, train, val, test = pd.load_dataset(bins=True)
else:
vocab, pos2id, train, val, test, mean, std = pd.load_dataset()
n_vocab = len(vocab)
n_pos = len(pos2id)
if bins:
loss_func = F.softmax_cross_entropy
else:
loss_func = F.mean_squared_error
if bins:
model = EyetrackingClassifier(n_vocab, n_units, n_participants, n_classes, loss_func, out, n_hidden=n_hidden, window=window, n_layers=n_layers, wlen=wlen, pos=pos, prev_fix=prev_fix, freq=freq, surprisal=surprisal, n_pos=n_pos, n_pos_units=50)
else:
model = EyetrackingLinreg(n_vocab, n_units, loss_func, out, n_hidden=n_hidden, window=window, n_layers=n_layers, wlen=wlen, pos=pos, prev_fix=prev_fix, freq=freq, surprisal=surprisal, n_pos=n_pos, n_pos_units=50)
'--output', dest='output', metavar='outputDirectory', help='dataset directory', required=True)
requiredNamed.add_argument(
'--batch-size', dest='BatchSize', help='Input Batch Size for dataset according to data size', required=True)
requiredNamed.add_argument(
'--epochs', dest='Epochs', help='Input string for translation', required=True)
args = parser.parse_args()
input_dir = args.input
output_dir = args.output
BATCH_SIZE = int(args.BatchSize)
Epoc=int(args.Epochs)
output_direc = merging_datafile(input_dir,output_dir)
pic_dir=input_dir+'/pickle_objects'
os.mkdir(pic_dir)
num_examples = None
input_tensor, target_tensor, inp_lang, targ_lang = load_dataset(output_direc, num_examples)
max_length_targ, max_length_inp = target_tensor.shape[1], input_tensor.shape[1]
input_tensor_train, input_tensor_val, target_tensor_train, target_tensor_val = train_test_split(input_tensor, target_tensor, test_size=0.2)
BUFFER_SIZE = len(input_tensor_train)
steps_per_epoch = len(input_tensor_train)//BATCH_SIZE
embedding_dim = 256
units = 1024
vocab_inp_size = len(inp_lang.word_index)+1
vocab_tar_size = len(targ_lang.word_index)+1
dataset = tf.data.Dataset.from_tensor_slices((input_tensor_train, target_tensor_train)).shuffle(BUFFER_SIZE)
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
example_input_batch, example_target_batch = next(iter(dataset))
encoder = Encoder(vocab_inp_size, embedding_dim, units, BATCH_SIZE)
gpu = -1
if out_type_eyetracking == 'tanh':
out_eyetracking = F.tanh
elif out_type_eyetracking == 'relu':
out_eyetracking = F.relu
elif out_type_eyetracking == 'sigmoid':
out_eyetracking = F.sigmoid
elif out_type_eyetracking == 'id':
out_eyetracking = F.identity
else:
raise Exception('Unknown output type: {}'.format(out_type))
if os.path.exists(sys.argv[1] + '.model'):
model = gensim.models.Word2Vec.load(sys.argv[1] + '.model')
vocab, pos2id, train, val, test, mean, std = pd.load_dataset(
model.wv.vocab, gensim=True)
n_vocab = len(model.wv.vocab)
n_pos = len(pos2id)
else:
vocab, pos2id, train, val, test, mean, std = pd.load_dataset()
n_vocab = len(vocab)
n_pos = len(pos2id)
loss_func = F.mean_squared_error
model_eyetracking = EyetrackingLinreg(n_vocab,
n_units,
loss_func,
out_eyetracking,
window=window_eyetracking,
n_layers=n_layers,