train=train_conv)
l1 = Convolution(input_sizes=[batch_size, 224, 224, 64],
filter_sizes=[3, 3, 64, 64],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=learning_rate,
activation=Relu(),
bias=0.0,
last_layer=False,
name="conv2",
load=weights_conv,
train=train_conv)
l2 = MaxPool(size=[batch_size, 224, 224, 64],
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="VALID")
l3 = Convolution(input_sizes=[batch_size, 112, 112, 64],
filter_sizes=[3, 3, 64, 128],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=learning_rate,
activation=Relu(),
bias=0.0,
last_layer=False,
name="conv3",
load=weights_conv,
train=train_conv)
optimizer = tf.train.AdamOptimizer(learning_rate=0.01, beta1=0.9, beta2=0.999, epsilon=1).minimize(loss)
'''
###############################################################
l0 = Convolution(input_sizes=[batch_size, 256, 256, 3],
filter_sizes=[3, 3, 3, 16],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Relu(),
last_layer=False)
l1 = MaxPool(size=[batch_size, 256, 256, 16],
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="SAME")
l2 = FeedbackConv(size=[batch_size, 128, 128, 16],
num_classes=num_classes,
sparse=sparse,
rank=rank)
l3 = Convolution(input_sizes=[batch_size, 128, 128, 16],
filter_sizes=[3, 3, 16, 16],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Relu(),
last_layer=False)
l1 = FeedbackConv(size=[batch_size, 28, 28, 32],
num_classes=10,
sparse=sparse,
rank=args.rank)
l2 = Convolution(input_sizes=[batch_size, 28, 28, 32],
filter_sizes=[3, 3, 32, 64],
num_classes=10,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Tanh(),
last_layer=False)
l3 = MaxPool(size=[batch_size, 28, 28, 64],
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="VALID")
l4 = FeedbackConv(size=[batch_size, 14, 14, 64],
num_classes=10,
sparse=sparse,
rank=args.rank)
l5 = ConvToFullyConnected(shape=[14, 14, 64])
l6 = FullyConnected(size=[14 * 14 * 64, 128],
num_classes=10,
init_weights=args.init,
alpha=ALPHA,
activation=Tanh(),
last_layer=False)
l7 = FeedbackFC(size=[14 * 14 * 64, 128],
num_classes=10,
get_conv5_bias = tf.get_default_graph().get_tensor_by_name(
os.path.split(conv5.name)[0] + '/bias:0')
else:
l0 = Convolution(input_sizes=[batch_size, 227, 227, 3],
filter_sizes=[11, 11, 3, 96],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 4, 4, 1],
padding="VALID",
alpha=ALPHA,
activation=Relu(),
bias=0.0,
last_layer=False)
l1 = MaxPool(size=[batch_size, 55, 55, 96],
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding="VALID")
l2 = Convolution(input_sizes=[batch_size, 27, 27, 96],
filter_sizes=[5, 5, 96, 256],
num_classes=num_classes,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Relu(),
bias=0.0,
last_layer=False)
l3 = MaxPool(size=[batch_size, 27, 27, 256],
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
############################################## tf.set_random_seed(0) tf.reset_default_graph() batch_size = tf.placeholder(tf.int32, shape=()) XTRAIN = tf.placeholder(tf.float32, [None, 32, 32, 3]) YTRAIN = tf.placeholder(tf.float32, [None, 100]) XTRAIN = tf.map_fn(lambda frame: tf.image.per_image_standardization(frame), XTRAIN) XTEST = tf.placeholder(tf.float32, [None, 32, 32, 3]) YTEST = tf.placeholder(tf.float32, [None, 100]) XTEST = tf.map_fn(lambda frame1: tf.image.per_image_standardization(frame1), XTEST) l0 = Convolution(input_sizes=[batch_size, 32, 32, 3], filter_sizes=[5, 5, 3, 96], num_classes=100, init_filters=args.init, strides=[1, 1, 1, 1], padding="SAME", alpha=ALPHA, activation=Tanh(), last_layer=False) l1 = MaxPool(size=[batch_size, 32, 32, 96], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME") l2 = FeedbackConv(size=[batch_size, 16, 16, 96], num_classes=100, sparse=sparse, rank=rank) l3 = Convolution(input_sizes=[batch_size, 16, 16, 96], filter_sizes=[5, 5, 96, 128], num_classes=100, init_filters=args.init, strides=[1, 1, 1, 1], padding="SAME", alpha=ALPHA, activation=Tanh(), last_layer=False) l4 = MaxPool(size=[batch_size, 16, 16, 128], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME") l5 = FeedbackConv(size=[batch_size, 8, 8, 128], num_classes=100, sparse=sparse, rank=rank) l6 = Convolution(input_sizes=[batch_size, 8, 8, 128], filter_sizes=[5, 5, 128, 256], num_classes=100, init_filters=args.init, strides=[1, 1, 1, 1], padding="SAME", alpha=ALPHA, activation=Tanh(), last_layer=False) l7 = MaxPool(size=[batch_size, 8, 8, 256], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME") l8 = FeedbackConv(size=[batch_size, 4, 4, 256], num_classes=100, sparse=sparse, rank=rank) l9 = ConvToFullyConnected(shape=[4, 4, 256]) l10 = FullyConnected(size=[4*4*256, 2048], num_classes=100, init_weights=args.init, alpha=ALPHA, activation=Tanh(), last_layer=False) l11 = FeedbackFC(size=[4*4*256, 2048], num_classes=100, sparse=sparse, rank=rank) l12 = FullyConnected(size=[2048, 2048], num_classes=100, init_weights=args.init, alpha=ALPHA, activation=Tanh(), last_layer=False)
XTEST = tf.placeholder(tf.float32, [None, 32, 32, 3])
YTEST = tf.placeholder(tf.float32, [None, 10])
XTEST = tf.map_fn(lambda frame1: tf.image.per_image_standardization(frame1),
XTEST)
l0 = Convolution(input_sizes=[batch_size, 32, 32, 3],
filter_sizes=[5, 5, 3, 96],
num_classes=10,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Tanh(),
last_layer=False)
l1 = MaxPool(size=[batch_size, 32, 32, 96],
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding="VALID")
l2 = FeedbackConv(size=[batch_size, 15, 15, 96],
num_classes=10,
sparse=sparse,
rank=rank)
l3 = Convolution(input_sizes=[batch_size, 15, 15, 96],
filter_sizes=[5, 5, 96, 128],
num_classes=10,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Tanh(),
last_layer=False)
XTEST = tf.placeholder(tf.float32, [None, 28, 28, 1])
YTEST = tf.placeholder(tf.float32, [None, 10])
XTEST = tf.map_fn(lambda frame1: tf.image.per_image_standardization(frame1),
XTEST)
l0 = Convolution(input_sizes=[batch_size, 28, 28, 1],
filter_sizes=[3, 3, 1, 32],
num_classes=10,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Tanh(),
last_layer=False)
l1 = MaxPool(size=[batch_size, 28, 28, 32],
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="SAME")
l2 = FeedbackConv(size=[batch_size, 14, 14, 32],
num_classes=10,
sparse=sparse,
rank=rank)
l3 = Convolution(input_sizes=[batch_size, 14, 14, 32],
filter_sizes=[3, 3, 32, 64],
num_classes=10,
init_filters=args.init,
strides=[1, 1, 1, 1],
padding="SAME",
alpha=ALPHA,
activation=Tanh(),
last_layer=False)