def main(unused_args):
''' Trains model from data '''
if not FLAGS.input_data:
raise ValueError("Must set --input_data to the filename of input dataset")
if not FLAGS.train_dir:
raise ValueError("Must set --train_dir to the directory where training files will be saved")
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
with tf.Graph().as_default(), tf.Session() as session:
''' To make tf.train.Saver write parameters as part of the saved file, add params to the graph as variables (hackish? - MK)'''
with tf.variable_scope("params", reuse=None):
num_layers_var = tf.Variable(FLAGS.num_layers, trainable=False, name='num_layers')
hidden_size_var = tf.Variable(FLAGS.hidden_size, trainable=False, name='hidden_size')
''' If pre-trained model loaded from file, use loaded vocabulary and NN geometry. Else, compute vocabulary and use command-line params for num_layers and hidden_size '''
if FLAGS.load_model:
vocab_size_var = tf.Variable(0, trainable=False, name='vocab_size')
tf.train.Saver([num_layers_var, hidden_size_var, vocab_size_var]).restore(session, FLAGS.load_model)
vocab_var = tf.Variable([0] * vocab_size_var.eval(), trainable=False, name='vocab')
tf.train.Saver([vocab_var]).restore(session, FLAGS.load_model)
FLAGS.num_layers = np.asscalar(num_layers_var.eval()) # need np.asscalar to upcast np.int32 to Python int
FLAGS.hidden_size = np.asscalar(hidden_size_var.eval())
vocab = Vocab.from_array(vocab_var.eval())
train_data, valid_data, test_data, vocab = reader.read_datasets(FLAGS.input_data, FLAGS.train_fraction, FLAGS.valid_fraction, vocab=vocab)
else:
train_data, valid_data, test_data, vocab = reader.read_datasets(FLAGS.input_data, FLAGS.train_fraction, FLAGS.valid_fraction, vocab_size=FLAGS.vocab_size)
vocab_size_var = tf.Variable(vocab.size, trainable=False, name='vocab_size')
vocab_var = tf.Variable(vocab.to_array(), trainable=False, name='vocab')
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.init_scale, FLAGS.init_scale)
with tf.variable_scope("model", initializer=initializer):
m = graph.inference_graph(vocab.size, FLAGS.num_layers, FLAGS.hidden_size, FLAGS.batch_size, FLAGS.num_steps, FLAGS.dropout_rate)
m.update(graph.cost_graph(m.logits, FLAGS.batch_size, FLAGS.num_steps, vocab.size))
m.update(graph.training_graph(m.cost, FLAGS.grad_clip))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("model", reuse=True):
mvalid = graph.inference_graph(vocab.size, FLAGS.num_layers, FLAGS.hidden_size, FLAGS.batch_size, FLAGS.num_steps)
mvalid.update(graph.cost_graph(mvalid.logits, FLAGS.batch_size, FLAGS.num_steps, vocab.size))
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAGS.load_model)
else:
print('Created model')
print('\tnum_layers:', FLAGS.num_layers)
print('\thidden_size:', FLAGS.hidden_size)
print('\tvocab_size:', vocab.size)
print()
print('Training parameters')
print('\tbatch_size:', FLAGS.batch_size)
print('\tnum_steps:', FLAGS.num_steps)
print('\tlearning_rate:', FLAGS.learning_rate)
print('\tbeta1:', FLAGS.beta1)
print('\tbeta2:', FLAGS.beta2)
print()
print('Datasets')
print('\ttraining dataset size:', len(train_data))
print('\tvalidation dataset size:', len(valid_data))
print('\ttest dataset size:', len(test_data))
print()
''' create two summaries: training cost and validation cost '''
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, graph=session.graph)
summary_train = summary_graph('Training cost', ema_decay=0.95)
summary_valid = summary_graph('Validation cost')
session.run([
m.lr.initializer,
m.beta1.initializer,
m.beta2.initializer,
])
tf.initialize_all_variables().run()
session.run([
tf.assign(m.lr, FLAGS.learning_rate),
tf.assign(m.beta1, FLAGS.beta1),
tf.assign(m.beta2, FLAGS.beta2),
])
state = session.run(m.initial_state)
iterations = len(train_data) // FLAGS.batch_size // FLAGS.num_steps * FLAGS.max_epochs
for i, (x, y) in enumerate(reader.next_batch(train_data, FLAGS.batch_size, FLAGS.num_steps)):
if i >= iterations:
break
start_time = time.time()
cost, state, _ = session.run([m.cost, m.final_state, m.train_op], {
m.input_data: x,
m.targets: y,
m.initial_state: state
})
epoch = float(i) / (len(train_data) // FLAGS.batch_size // FLAGS.num_steps)
time_elapsed = time.time() - start_time
print('%d/%d (epoch %.3f), train_loss = %6.8f, time/batch = %.4fs' % (i+1, iterations, epoch, cost, time_elapsed))
session.run([summary_train.update], {summary_train.x: cost})
if (i+1) % FLAGS.eval_val_every == 0 or i == iterations-1:
# evaluate loss on validation data
cost = run_test(session, mvalid, valid_data, FLAGS.batch_size, FLAGS.num_steps)
print("validation cost = %6.8f" % cost)
save_as = '%s/epoch%.2f_%.4f.model' % (FLAGS.train_dir, epoch, cost)
saver.save(session, save_as)
''' write out summary events '''
buffer, = session.run([summary_train.summary])
summary_writer.add_summary(buffer, i)
session.run([summary_valid.update], {summary_valid.x: cost})
buffer, = session.run([summary_valid.summary])
summary_writer.add_summary(buffer, i)
summary_writer.flush()
if len(test_data) > FLAGS.batch_size * FLAGS.num_steps:
cost = run_test(session, mvalid, test_data, FLAGS.batch_size, FLAGS.num_steps)
print("Test cost: %.3f" % test_loss)
choices=('age', 'gender', 'both', 'joint'), default='both')
# parser.add_argument('--trainer', help='which training algorithm to use',
# choices=('adagrad', 'sgd', 'adam', 'adadelta', 'momentum'), default='adam')
parser.add_argument('--num-epochs', help='Number of epochs', default=50)
# parser.add_argument('--status', help='number of processed instances between status updates', default=10000, type=int)
# parser.add_argument('--noise', help='amount of noise added to embeddings', default=0.1, type=float)
parser.add_argument('--dim-rnn', help='dimensionality of hidden RNN layer', default=50, type=int)
parser.add_argument('--dim-emb', help='dimensionality of word embeddings', default=100, type=int)
parser.add_argument('--dim-out', help='dimensionality of output layer', default=32, type=int)
# parser.add_argument('--dropout', help='dropout probability for final sentence representation', default=0.0, type=float)
parser.add_argument('--batch-size', help='batch size', default=1, type=int)
parser.add_argument('--max-len', help='Maximum length of input', default=100, type=int)
args = parser.parse_args()
# Read in data, mapping words to integers
datasets = read_datasets(args.train, args.test, args.dev, args.target)
train_dataset = datasets['train']
dev_dataset = datasets['dev']
test_dataset = datasets['test']
word_mapper = datasets['word_mapper']
label_mapper = datasets['label_mapper']
word_embs = tf.Variable(tf.random_uniform([len(word_mapper), args.dim_emb], -1, 1, tf.float32),
name='word_embs')
# Create dense representations of datasets
X_train, y_train = make_dense(train_dataset['sentences'], args.max_len,
train_dataset['labels'], len(label_mapper))
train_sent_lens = np.array([len(sent) for sent in train_dataset['sentences']])
X_dev, y_dev = make_dense(dev_dataset['sentences'], args.max_len,
dev_dataset['labels'], len(label_mapper))
def pretrain_mdnet(datasets,
init_model_path,
result_dir,
load_path=None,
shuffle=True,
norm=False,
dropout=True,
regularization=True):
config = Config()
# print parameters
print('shuffle', shuffle)
print('norm', norm)
print('dropout', dropout)
print('regularization', regularization)
print('init_model_path', init_model_path)
print('result_dir', result_dir)
# create directory
if not os.path.exists(result_dir):
os.makedirs(result_dir)
# load sequences
train_data = reader.read_datasets(datasets)
K = len(train_data.data)
# create session and saver
gpu_config = tf.ConfigProto(allow_soft_placement=True)
sess = tf.InteractiveSession(config=gpu_config)
# load model, weights
model = MDNet(config)
model.build_trainer(K,
config.batch_size,
dropout=dropout,
regularization=regularization)
tf.global_variables_initializer().run()
model.load(init_model_path, sess)
sess.run(model.lr_rate.assign(config.lr_rate))
# create saver
saver = tf.train.Saver(
[v for v in tf.global_variables() if 'fc6' not in v.name])
# restore from model
if load_path is not None:
saver.restore(sess, load_path)
# prepare roidb and frame list
train_loss_file = open(os.path.join(result_dir, 'train_loss.txt'), 'w')
n_frames = config.batch_frames * config.num_cycle
for i in range(config.num_cycle):
loss_total = 0
print('###### training cycle ' + str(i) + '/' + str(config.num_cycle) +
'...')
seq_i = 0
for seq, seq_data in train_data.data.iteritems():
print('### training video "' + seq + '"...')
seq_n_frames = len(seq_data.frames)
## prepare roidb
print('- preparing roidb...')
seq_data.rois = proc.seq2roidb(seq_data, config)
## prepare frame list
print('- shuffle frames...')
seq_data.frame_lists = []
while len(seq_data.frame_lists) < n_frames:
seq_data.frame_lists = np.r_[
seq_data.frame_lists,
np.random.permutation(seq_n_frames)]
seq_data.frame_lists = seq_data.frame_lists[:n_frames]
## start training
# extract batch_size frames
frame_inds = seq_data.frame_lists[config.batch_frames *
i:config.batch_frames *
(i + 1)].astype(np.int)
# sample boxes
pos_boxes = np.concatenate([
seq_data.rois[frame_ind].pos_boxes for frame_ind in frame_inds
],
axis=0)
neg_boxes = np.concatenate([
seq_data.rois[frame_ind].neg_boxes for frame_ind in frame_inds
],
axis=0)
pos_inds = np.random.permutation(
config.posPerFrame * config.batch_frames)[:config.batch_pos]
neg_inds = np.random.permutation(
config.negPerFrame * config.batch_frames)[:config.batch_neg]
# pack as boxes, paths
pos_boxes = pos_boxes[pos_inds]
neg_boxes = neg_boxes[neg_inds]
boxes = np.r_[pos_boxes, neg_boxes]
box_relinds = np.r_[pos_inds // config.posPerFrame,
neg_inds // config.negPerFrame]
paths = [seq_data.frames[ind] for ind in frame_inds[box_relinds]]
gts = np.repeat(np.identity(2),
[config.batch_pos, config.batch_neg],
axis=0)
patches = proc.load_patch(paths, boxes, norm=False)
# shuffle
if shuffle:
inds = np.random.permutation(config.batch_size)
patches = patches[inds]
gts = gts[inds]
# training
_, loss, score, weight, bias = sess.run(
[
model.trainable[seq_i], model.losses['loss-' + str(seq_i)],
model.layers['fc6-' + str(seq_i)],
model.weights['fc6-' + str(seq_i)],
model.biases['fc6-' + str(seq_i)]
],
feed_dict={
model.layers['input']: patches,
model.layers['y-' + str(seq_i)]: gts
})
print(seq_i)
print(score.reshape(-1, 2)[:5])
print(gts[:5])
print(np.mean(loss))
print(weight)
print(bias)
loss_total += np.mean(loss)
# update seq_i
seq_i += 1
## save the model
train_loss_file.write('Epoch ' + str(i) + ', Loss: ' +
str(np.mean(loss)))
saver.save(sess,
os.path.join(result_dir, 'model_e' + str(i) + '.ckpt'),
global_step=i + 1)
train_loss_file.close()
parser.add_argument('--dev', help='dev files', required=True)
parser.add_argument('--test', help='test files', required=True)
parser.add_argument('--architecture', '-a', help="Model architecture", choices=['lstm', 'cnn'])
parser.add_argument('--target', help='predict age, gender, both of them, or the joint cross-product',
choices=('age', 'gender', 'both', 'joint'), default='both')
parser.add_argument('--num-epochs', help='Number of epochs', default=50)
parser.add_argument('--dim-rnn', help='dimensionality of hidden RNN layer', default=50, type=int)
parser.add_argument('--dim-emb', help='dimensionality of word embeddings', default=100, type=int)
parser.add_argument('--dim-out', help='dimensionality of output layer', default=32, type=int)
# parser.add_argument('--dropout', help='dropout probability for final sentence representation', default=0.0, type=float)
parser.add_argument('--batch-size', help='batch size', default=1, type=int)
parser.add_argument('--max-len', help='Maximum length of input', default=100, type=int)
args = parser.parse_args()
# Read in data, mapping words to integers
datasets = read_datasets(args.train, args.test, args.dev, "gender") # for now just gender
train_dataset = datasets['train']
dev_dataset = datasets['dev']
word_mapper = datasets['word_mapper']
label_mapper = datasets['label_mapper']
# Create dense representations of datasets
X_train, y_train = make_dense(train_dataset['sentences'], args.max_len,
train_dataset['labels'], len(label_mapper))
train_sent_lens = np.array([len(sent) for sent in train_dataset['sentences']])
X_dev, y_dev = make_dense(dev_dataset['sentences'], args.max_len,
dev_dataset['labels'], len(label_mapper))
dev_sent_lens = np.array([len(sent) for sent in dev_dataset['sentences']])
assert len(X_train) == len(y_train)
