def get_model(sparse_data, train_labels, is_training, tensor_in_sizes, num_classes = 10, scope = "mn16-", initializer = None, regularizer = None):
strides = [1,1,1,1,1]
padding = "SAME"
dim = 5
pooling_sizes = [1,2,2,2,1]
batch_size = tensor_in_sizes[0]
total_size = 1
for i in range(1, len(tensor_in_sizes)): #skip batch size
total_size = total_size * tensor_in_sizes[i]
sd_converted = ld.create_sparse_data_to_direct_sparse(sparse_data, dim)
ops = [None]*6
d1 = 0.33
net, tensor_in_sizes = ld.create_sparse_conv_layer(sd_converted, [3,3,3,1,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc1", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc2", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc3", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, 1)
d2 = 0.33
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc4", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc5", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-ABS", name = scope + "sc6", initializer=initializer)
sd = ld.create_direct_sparse_to_dense(net, dim)
sd_flat = tf.reshape(sd, [batch_size, total_size * 2])
conv_out = tf.layers.dropout(sd_flat, 0.5, name="dropout", training=is_training)
fc512 = tf.layers.dense(conv_out, 1024, name="dense2")
fc10 = tf.layers.dense(fc512, num_classes, name="dense1")
#if train:
sd_out = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fc10, labels=train_labels, name = "softmax_loss"))
p_sd_out = tf.nn.softmax(logits=fc10)
return [sd_out, p_sd_out, ops]
def get_model(sparse_data, train_labels, is_training, tensor_in_sizes, num_classes = 10, scope = "mn128-", initializer = None, regularizer = None):
strides = [1,1,1,1,1]
padding = "SAME"
dim = 5
pooling_sizes = [1,2,2,2,1]
dpooling_sizes = [2,2,2]
batch_size = tensor_in_sizes[0]
total_size = 1
for i in range(1, len(tensor_in_sizes)): #skip batch size
total_size = total_size * tensor_in_sizes[i]
ops = [None]*9
sd_converted = ld.create_sparse_data_to_direct_sparse(sparse_data, dim)
d1 = 0.02
net, tensor_in_sizes = ld.create_sparse_conv_layer(sd_converted, [3,3,3,1,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc1", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc2", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc3", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, 0.06)
d2 = 0.06
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc4", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc5", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc6", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, 0.18)
d3 = 0.14
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,24], tensor_in_sizes, strides, padding, dim, d3, "K-RELU", name = scope + "sc7", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,24,24], tensor_in_sizes, strides, padding, dim, d3, "K-RELU", name = scope + "sc8", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,24,24], tensor_in_sizes, strides, padding, dim, d3, "K-RELU", name = scope + "sc9", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, 0.50)
net = ld.create_direct_sparse_to_dense(net, dim)
net = tf.reshape(net, [batch_size, 16, 16, 16, 24])
#dense layers
net = tf.layers.conv3d(inputs=net, filters=32, kernel_size=[3, 3, 3], padding="same", activation=tf.nn.relu, name = scope + "sc10", kernel_initializer=initializer, kernel_regularizer=regularizer)
net = tf.layers.conv3d(inputs=net, filters=32, kernel_size=[3, 3, 3], padding="same", activation=tf.nn.relu, name = scope + "sc11", kernel_initializer=initializer, kernel_regularizer=regularizer)
net = tf.layers.conv3d(inputs=net, filters=32, kernel_size=[3, 3, 3], padding="same", activation=tf.nn.relu, name = scope + "sc12", kernel_initializer=initializer, kernel_regularizer=regularizer)
net = tf.layers.max_pooling3d(inputs=net, pool_size=dpooling_sizes, strides=2, padding="same", name="dp1")
net = tf.layers.conv3d(inputs=net, filters=40, kernel_size=[3, 3, 3], padding="same", activation=tf.nn.relu, name = scope + "sc13", kernel_initializer=initializer, kernel_regularizer=regularizer)
net = tf.layers.conv3d(inputs=net, filters=40, kernel_size=[3, 3, 3], padding="same", activation=tf.nn.relu, name = scope + "sc14", kernel_initializer=initializer, kernel_regularizer=regularizer)
n3 = net = tf.layers.conv3d(inputs=net, filters=40, kernel_size=[3, 3, 3], padding="same", name = scope + "sc15")
net = tf.layers.dropout(net, 0.5, name="dropout", training=is_training)
net = tf.reshape(net, [batch_size, -1])
net = tf.layers.dense(net, 1024)
net = tf.layers.dense(net, num_classes)
sd_out = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=net, labels=train_labels, name = "softmax_loss"))
p_sd_out = tf.nn.softmax(logits=net)
return [sd_out, p_sd_out, ops]
def get_model(sparse_data, train_labels, is_training, tensor_in_sizes, num_classes = 10, scope = "mn128-", initializer = None, regularizer = None):
strides = [1,1,1,1,1]
padding = "SAME"
dim = 5
pooling_sizes = [1,2,2,2,1]
dpooling_sizes = [2,2,2]
batch_size = tensor_in_sizes[0]
total_size = 1
for i in range(1, len(tensor_in_sizes)): #skip batch size
total_size = total_size * tensor_in_sizes[i]
ops = [None]*9
sd_converted = ld.create_sparse_data_to_direct_sparse(sparse_data, dim)
d1 = 0.14
net, tensor_in_sizes = ld.create_sparse_conv_layer(sd_converted, [3,3,3,1,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc1", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc2", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,8], tensor_in_sizes, strides, padding, dim, d1, "K-RELU", name = scope + "sc3", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, min(1, 6 * d1))
d2 = 0.33
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,8,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc4", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc5", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,16], tensor_in_sizes, strides, padding, dim, d2, "K-RELU", name = scope + "sc6", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_pooling_layer(net, pooling_sizes, tensor_in_sizes, dim, min(1, 6 * d2))
d3 = 0.66
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,16,24], tensor_in_sizes, strides, padding, dim, d3, "K-RELU", name = scope + "sc7", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,24,24], tensor_in_sizes, strides, padding, dim, d3, "K-RELU", name = scope + "sc8", initializer=initializer)
net, tensor_in_sizes = ld.create_sparse_conv_layer(net, [3,3,3,24,24], tensor_in_sizes, strides, padding, dim, d3, "K-ABS", name = scope + "sc9", initializer=initializer)
net = ld.create_direct_sparse_to_dense(net, dim)
net = tf.reshape(net, [batch_size, 8, 8, 8, 24])
#dense layers
net = tf.layers.dropout(net, 0.5, name="dropout", training=is_training)
net = tf.reshape(net, [batch_size, -1])
net = tf.layers.dense(net, 1024)
net = tf.layers.dense(net, num_classes)
sd_out = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=net, labels=train_labels, name = "softmax_loss"))
p_sd_out = tf.nn.softmax(logits=net)
return [sd_out, p_sd_out, ops]
def model_mnist(sparse_data,
tensor_in_sizes,
train_labels=None,
num_classes=10,
scope='mn256-',
initializer=None,
d1=0.1,
d2=0.3,
d3=0.4,
rscale=0.02,
max_bias=0.1):
dim = 5
strides = [1, 1, 1, 1, 1]
padding = 'SAME'
pooling_sizes = [1, 1, 2, 2, 1]
batch_size = tensor_in_sizes[0]
total_size = np.prod(tensor_in_sizes)
net = {}
ops = [None] * 6
tmp_tin = tensor_in_sizes
net['sd_converted'] = ld.create_sparse_data_to_direct_sparse(
sparse_data, dim)
net['conv1_1'], tmp_tin, ops[0] = ld.create_sparse_conv_layer_reg(
net['sd_converted'], [1, 3, 3, 1, 8],
tmp_tin,
strides,
padding,
dim,
d1,
'K-RELU',
name=scope + 'sc1',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['conv1_2'], tmp_tin, ops[1] = ld.create_sparse_conv_layer_reg(
net['conv1_1'], [1, 3, 3, 8, 8],
tmp_tin,
strides,
padding,
dim,
d1,
'K-RELU',
name=scope + 'sc2',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['conv1_3'], tmp_tin, ops[2] = ld.create_sparse_conv_layer_reg(
net['conv1_2'], [1, 3, 3, 8, 8],
tmp_tin,
strides,
padding,
dim,
d1,
'K-RELU',
name=scope + 'sc3',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['pool1'], tmp_tin = ld.create_sparse_pooling_layer(
net['conv1_3'], pooling_sizes, tmp_tin, dim, d3)
net['conv2_1'], tmp_tin, ops[3] = ld.create_sparse_conv_layer_reg(
net['pool1'], [1, 3, 3, 8, 16],
tmp_tin,
strides,
padding,
dim,
d2,
'K-RELU',
name=scope + 'sc4',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['conv2_2'], tmp_tin, ops[4] = ld.create_sparse_conv_layer_reg(
net['conv2_1'], [1, 3, 3, 16, 16],
tmp_tin,
strides,
padding,
dim,
d2,
'K-RELU',
name=scope + 'sc5',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['conv2_3'], tmp_tin, ops[5] = ld.create_sparse_conv_layer_reg(
net['conv2_2'], [1, 3, 3, 16, 16],
tmp_tin,
strides,
padding,
dim,
d2,
'K-ABS',
name=scope + 'sc6',
initializer=initializer,
scale=rscale,
bias_offset=max_bias)
net['sparse_to_dense'] = ld.create_direct_sparse_to_dense(
net['conv2_3'], dim)
net['dense_reshaped1'] = tf.reshape(net['sparse_to_dense'],
[batch_size, 1, 14, 14, 16])
net['dense_reshaped2'] = tf.reshape(net['dense_reshaped1'],
[batch_size, -1])
net['dense1'] = tf.layers.dense(net['dense_reshaped2'], 512)
net['dense2'] = tf.layers.dense(net['dense1'], num_classes)
predictions = {
'classes': tf.argmax(net['dense2'], axis=1),
'probabilities': tf.nn.softmax(net['dense2'])
}
loss = tf.losses.softmax_cross_entropy(onehot_labels=train_labels,
logits=net['dense2'])
loss += tf.reduce_sum(tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES))
accuracy = tf.metrics.accuracy(tf.argmax(train_labels, axis=1),
predictions['classes'])
return loss, predictions, accuracy, net, ops
def model_mnist(sparse_data,
tensor_in_sizes,
train_labels=None,
num_classes=10,
scope='mn256-',
initializer=None,
d1=0.1,
d2=0.3,
d3=0.4):
dim = 5
strides = [1, 1, 1, 1, 1]
padding = 'SAME'
pooling_sizes = [1, 1, 2, 2, 1]
batch_size = tensor_in_sizes[0]
total_size = np.prod(tensor_in_sizes)
net = {}
net['sd_converted'] = ld.create_sparse_data_to_direct_sparse(
sparse_data, dim)
net['conv1_1'] = ld.create_sparse_conv_layer(net['sd_converted'],
[1, 3, 3, 1, 8],
strides,
padding,
dim,
d1,
'K-ABS',
name=scope + 'sc1',
initializer=initializer)
net['conv1_2'] = ld.create_sparse_conv_layer(net['conv1_1'],
[1, 3, 3, 8, 8],
strides,
padding,
dim,
d1,
'K-RELU',
name=scope + 'sc2',
initializer=initializer)
net['conv1_3'] = ld.create_sparse_conv_layer(net['conv1_2'],
[1, 3, 3, 8, 8],
strides,
padding,
dim,
d1,
'K-RELU',
name=scope + 'sc3',
initializer=initializer)
net['pool1'] = ld.create_sparse_pooling_layer(net['conv1_3'],
pooling_sizes, dim, d3)
net['conv2_1'] = ld.create_sparse_conv_layer(net['pool1'],
[1, 3, 3, 8, 16],
strides,
padding,
dim,
d2,
'K-ABS',
name=scope + 'sc4',
initializer=initializer)
net['conv2_2'] = ld.create_sparse_conv_layer(net['conv2_1'],
[1, 3, 3, 16, 16],
strides,
padding,
dim,
d2,
'K-RELU',
name=scope + 'sc5',
initializer=initializer)
net['conv2_3'] = ld.create_sparse_conv_layer(net['conv2_2'],
[1, 3, 3, 16, 16],
strides,
padding,
dim,
d2,
'K-ABS',
name=scope + 'sc6',
initializer=initializer)
net['sparse_to_dense'] = ld.create_direct_sparse_to_dense(
net['conv2_3'], dim)
net['dense_reshaped1'] = tf.reshape(net['sparse_to_dense'],
[batch_size, 1, 14, 14, 16])
net['dense_reshaped2'] = tf.reshape(net['dense_reshaped1'],
[batch_size, -1])
net['dense1'] = tf.layers.dense(net['dense_reshaped2'], 512)
net['dense2'] = tf.layers.dense(net['dense1'], num_classes)
predictions = {
'classes': tf.argmax(net['dense2'], axis=1),
'probabilities': tf.nn.softmax(net['dense2'])
}
loss = tf.losses.softmax_cross_entropy(onehot_labels=train_labels,
logits=tf.clip_by_value(
net['dense2'], 1e-7, 1e8))
accuracy = tf.metrics.accuracy(tf.argmax(train_labels, axis=1),
predictions['classes'])
return loss, predictions, accuracy, net
