def main(_):
# Create model.
batch_size = 2
image_size = 256
vid_placeholder = tf.placeholder(tf.float32,
(batch_size*FLAGS.num_frames, image_size, image_size, 3)) # pylint: disable=line-too-long
object_placeholder = tf.placeholder(tf.float32,
(batch_size*FLAGS.num_frames, image_size, image_size, FLAGS.num_object_classes)) # pylint: disable=line-too-long
input_placeholder = (vid_placeholder, object_placeholder)
# We are using the full_asnp50_structure, since we feed both video and object.
FLAGS.model_structure = json.dumps(model_structures.full_asnp50_structure) # pylint: disable=line-too-long
FLAGS.model_edge_weights = json.dumps(model_structures.full_asnp_structure_weights) # pylint: disable=line-too-long
network = assemblenet_plus.assemblenet_plus(
assemblenet_depth=50,
num_classes=FLAGS.num_classes,
data_format='channels_last')
# The model function takes the inputs and is_training.
outputs = network(input_placeholder, False)
with tf.Session() as sess:
# Generate a random video to run on.
# This should be replaced by a real video.
sess.run(tf.global_variables_initializer())
vid = np.random.rand(*vid_placeholder.shape)
obj = np.random.rand(*object_placeholder.shape)
logits = sess.run(outputs, feed_dict={input_placeholder: (vid, obj)})
print(logits)
print(np.argmax(logits, axis=1))
def main(_):
# Create model.
batch_size = 2
image_size = 256
vid_placeholder = tf.placeholder(tf.float32,
(batch_size, FLAGS.num_frames, image_size, image_size, 3)) # pylint: disable=line-too-long
if FLAGS.assemblenet_mode == 'assemblenet_plus_lite':
FLAGS.model_structure = json.dumps(
model_structures.asnp_lite_structure)
FLAGS.model_edge_weights = json.dumps(model_structures.asnp_lite_structure_weights) # pylint: disable=line-too-long
network = assemblenet_plus_lite.assemblenet_plus_lite(
num_layers=[3, 5, 11, 7],
num_classes=FLAGS.num_classes,
data_format='channels_last')
else:
vid_placeholder = tf.reshape(vid_placeholder,
[batch_size * FLAGS.num_frames, image_size, image_size, 3]) # pylint: disable=line-too-long
if FLAGS.assemblenet_mode == 'assemblenet_plus':
# Here, we are using model_structures.asn50_structure for AssembleNet++
# instead of full_asnp50_structure. By using asn50_structure, it
# essentially becomes AssembleNet++ without objects, only requiring RGB
# inputs (and optical flow to be computed inside the model).
FLAGS.model_structure = json.dumps(
model_structures.asn50_structure)
FLAGS.model_edge_weights = json.dumps(model_structures.asn_structure_weights) # pylint: disable=line-too-long
network = assemblenet_plus.assemblenet_plus(
assemblenet_depth=50,
num_classes=FLAGS.num_classes,
data_format='channels_last')
else:
FLAGS.model_structure = json.dumps(
model_structures.asn50_structure)
FLAGS.model_edge_weights = json.dumps(model_structures.asn_structure_weights) # pylint: disable=line-too-long
network = assemblenet.assemblenet_v1(assemblenet_depth=50,
num_classes=FLAGS.num_classes,
data_format='channels_last')
# The model function takes the inputs and is_training.
outputs = network(vid_placeholder, False)
with tf.Session() as sess:
# Generate a random video to run on.
# This should be replaced by a real video.
vid = np.random.rand(*vid_placeholder.shape)
sess.run(tf.global_variables_initializer())
logits = sess.run(outputs, feed_dict={vid_placeholder: vid})
print(logits)
print(np.argmax(logits, axis=1))