def train():
g = tf.Graph()
with g.as_default():
# load data get iterator
data_loader = data_utils_mean.DataLoader(SEQUENCE_LENGTH, BATCH_SIZE,
NUM_EPOCHS)
iterator = data_loader.load_data(TRAIN_TFR_PATH, True)
with tf.Session(graph=g) as sess:
frameNo, image, label = iterator.get_next()
# VGG FACE network
VGGFace_network = vgg_face.VGGFace(SEQUENCE_LENGTH * BATCH_SIZE)
image_batch = reshape_to_cnn(image)
VGGFace_network.setup(image_batch, trainable=False)
face_output = VGGFace_network.get_face_fc0()
# RNN part
rnn_in = reshape_to_rnn(face_output)
prediction = models.get_prediction_highway(rnn_in,
useHighway=False)
prediction = tf.reshape(prediction,
[BATCH_SIZE, SEQUENCE_LENGTH, 2])
label_batch = tf.reshape(label, [BATCH_SIZE, SEQUENCE_LENGTH, 2])
# compute losses using slim
compute_loss(prediction, label_batch)
total_loss = slim.losses.get_total_loss()
optimizer = tf.train.AdamOptimizer(LEARNING_RATE)
# restore VGG-FACE model at the beginning
restore_names = VGGFace_network.get_restore_vars()
variables_to_restore = tf.contrib.framework.get_variables_to_restore(
include=restore_names)
init_fn = slim.assign_from_checkpoint_fn(VGG_RESTORE_PATH,
variables_to_restore,
ignore_missing_vars=False)
# summarize_gradients : Whether or not add summaries for each gradient.
# variables_to_train: an optional list of variables to train. If None, it will default to all tf.trainable_variables().
train_op = slim.learning.create_train_op(
total_loss,
optimizer,
summarize_gradients=
True # Whether or not add summaries for each gradient.
)
loggingTF.set_verbosity(1)
# keep 10000 ckpts
saver = tf.train.Saver(max_to_keep=10000)
# including initialize local and global variables
slim.learning.train(
train_op,
TRAIN_DIR,
init_fn=init_fn,
save_summaries_secs=60 *
15, # How often, in seconds, to save summaries.
log_every_n_steps=
500, # The frequency, in terms of global steps, that the loss and global step are logged.
save_interval_secs=60 *
15, # How often, in seconds, to save the model to `logdir`.
saver=saver)
def evaluate():
g = tf.Graph()
with g.as_default():
# load data get iterator
data_loader = data_utils_mean.DataLoader(SEQUENCE_LENGTH, BATCH_SIZE, NUM_EPOCHS)
iterator = data_loader.load_data(TEST_TFR_PATH, False)
frameNo, image, label = iterator.get_next()
# define model graph
# VGG FACE network
VGGFace_network = vgg_face.VGGFace(SEQUENCE_LENGTH * BATCH_SIZE)
image_batch = tf.reshape(image, [-1, 96, 96, 3])
VGGFace_network.setup(image_batch, trainable=False) # image_batch is a tensor of shape (batch_size*seq_length,image_dim,image_dim,3)
face_output = VGGFace_network.get_face_fc0()
# RNN part
rnn_in = reshape_to_rnn(face_output)
prediction = models.get_prediction_highway(rnn_in, useHighway=False)
prediction = tf.reshape(prediction, [BATCH_SIZE, SEQUENCE_LENGTH, 2])
label_batch = tf.reshape(label, [BATCH_SIZE, SEQUENCE_LENGTH, 2])
# Computing MSE and Concordance values, and adding them to summary
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
'eval/mse_valence': slim.metrics.streaming_mean_squared_error(prediction[:, :, 0], label_batch[:, :, 0]),
'eval/mse_arousal': slim.metrics.streaming_mean_squared_error(prediction[:, :, 1], label_batch[:, :, 1]),
})
summary_ops = []
conc_total = 0
mse_total = 0
for i, name in enumerate(['valence', 'arousal']):
with tf.name_scope(name) as scope:
concordance_cc2, values, updates = metrics.concordance_cc2(
tf.reshape(prediction[:, :, i], [-1]),
tf.reshape(label_batch[:, :, i], [-1]))
for n, v in updates.items():
names_to_updates[n + '/' + name] = v
op = tf.summary.scalar('eval/concordance_' + name, concordance_cc2)
op = tf.Print(op, [concordance_cc2], 'eval/concordance_' + name)
summary_ops.append(op)
mse_eval = 'eval/mse_' + name
op = tf.summary.scalar(mse_eval, names_to_values[mse_eval])
op = tf.Print(op, [names_to_values[mse_eval]], mse_eval)
summary_ops.append(op)
mse_total += names_to_values[mse_eval]
conc_total += concordance_cc2
conc_total = conc_total / 2
mse_total = mse_total / 2
op = tf.summary.scalar('eval/concordance_total', conc_total)
op = tf.Print(op, [conc_total], 'eval/concordance_total')
summary_ops.append(op)
op = tf.summary.scalar('eval/mse_total', mse_total)
op = tf.Print(op, [mse_total], 'eval/mse_total')
summary_ops.append(op)
num_batches = int(NUM_BATCHES)
loggingTF.set_verbosity(1)
if not os.path.exists(SUMMARY_PATH):
os.makedirs(SUMMARY_PATH)
# always check latest ckpt and wait for next.
slim.evaluation.evaluate_once('',
MODEL_PATH,
SUMMARY_PATH,
num_evals=num_batches,
eval_op=list(names_to_updates.values()),
summary_op=tf.summary.merge(summary_ops),
)