keep_prob = mod.ConstantPlaceholderModule("keep_prob",
shape=(),
dtype=tf.float32)
activations = [tf.nn.relu, tf.nn.relu, tf.nn.relu, tf.identity]
filter_shapes = [[8, 8, 1, 6], [8, 8, 6, 16]]
bias_shapes = [[1, 28, 28, 6], [1, 14, 14, 16], [1, 120], [1, 10]]
ksizes = [[1, 4, 4, 1], [1, 4, 4, 1]]
pool_strides = [[1, 2, 2, 1], [1, 2, 2, 1]]
network = recurrentLenet5("rlenet5", activations, filter_shapes, bias_shapes,
ksizes, pool_strides, keep_prob.placeholder)
one_time_error = mod.ErrorModule("cross_entropy", cross_entropy)
error = mod.TimeAddModule("add_error")
accuracy = mod.BatchAccuracyModule("accuracy")
optimizer = mod.OptimizerModule("adam", tf.train.AdamOptimizer())
network.add_input(inp)
one_time_error.add_input(network)
one_time_error.add_input(labels)
error.add_input(one_time_error, 0)
error.add_input(error, -1)
accuracy.add_input(network)
accuracy.add_input(labels)
optimizer.add_input(error)
optimizer.create_output(TIME_DEPTH)
accuracy.create_output(TIME_DEPTH)
def train_batch(sess, i):
batch = train_mnist[i * BATCH_SIZE:(i + 1) * BATCH_SIZE]
# parse data from files using get_digits
dataset = get_mnist.MNIST(n_occ=FLAGS.n_occluders)
# initialize classes with parameters
# -----
network = networks.ReCoNNet("ReCoNNet", is_training.placeholder, activations,
conv_filter_shapes, bias_shapes, ksizes,
pool_strides, topdown_filter_shapes,
topdown_output_shapes, keep_prob.placeholder,
FLAGS)
one_time_error = m.ErrorModule("cross_entropy", CROSSENTROPY_FN)
error = m.TimeAddModule("add_error")
optimizer = m.OptimizerModule("adam", tf.train.AdamOptimizer(lrate))
accuracy = m.BatchAccuracyModule("accuracy")
network.add_input(inp)
one_time_error.add_input(network)
one_time_error.add_input(labels)
error.add_input(one_time_error, 0)
error.add_input(error, -1)
optimizer.add_input(error)
accuracy.add_input(network)
accuracy.add_input(labels)
error.create_output(TIME_DEPTH + TIME_DEPTH_BEYOND)
optimizer.create_output(TIME_DEPTH)
for time in range(0, (TIME_DEPTH + TIME_DEPTH_BEYOND + 1)):
