def mode_dot(tensor, matrix, mode):
new_shape = tensor.get_shape().as_list()
if matrix.get_shape().as_list()[1] != tensor.get_shape().as_list()[mode]:
raise ValueError("Shape error. {0}(matrix's 2nd dimension) is not as same as {1} (dimension of the tensor)".format(matrix.get_shape().as_list()[1], tensor.get_shape().as_list()[mode]))
new_shape[mode] = matrix.get_shape().as_list()[0]
res = tf.matmul(matrix, TNSR.unfold(tensor, mode))
return TNSR.fold(res, mode, new_shape)
def mode_dot(tensor, matrix, mode):
new_shape = tensor.get_shape().as_list()
if matrix.get_shape().as_list()[1] != tensor.get_shape().as_list()[mode]:
raise ValueError("Shape error. {0}(matrix's 2nd dimension) is not as same as {1} (dimension of the tensor)".format(matrix.get_shape().as_list()[1], tensor.get_shape().as_list()[mode]))
new_shape[mode] = matrix.get_shape().as_list()[0]
res = tf.matmul(matrix, TNSR.unfold(tensor, mode))
return TNSR.fold(res, mode, new_shape)
def ttrl(x, ranks, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
if(type(n_outputs) == type([])):
#n_outputs [2,2,2,2,2,2]
#ranks [1,1,2,2,3,4,3,2,2,1,1]
#x [50,14,14,32]
#input_shape[14,14,32,2,2,2,2,2,2]
suffix = np.prod(n_outputs)
input_shape = x.get_shape().as_list()[1:]+n_outputs
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
cores = []
for i in range(len(input_shape)-1):
cores.append(tf.get_variable("core_{0}_output_{1}".format(i,suffix),
shape = (ranks[i], input_shape[i], ranks[i+1]),
initializer = weight_initializer))
cores.append(tf.get_variable("core_{0}_last_output_{1}".format(i,suffix),
shape = (ranks[-2], input_shape[-1], ranks[-1]),
initializer = weight_initializer))
#for c in cores:
# print(c.get_shape().as_list())
regression_weights = TNSR.tt_to_tensor(cores)
w_minus1 = tf.reshape(regression_weights, [np.prod(input_shape[:3]), np.prod(n_outputs)])
x_0 = tf.reshape(x, [-1, np.prod(x.get_shape().as_list()[1:])])
return tf.reshape(tf.add(tf.matmul(x_0, w_minus1) ,bias), [-1]+n_outputs)
else:
suffix = n_outputs
input_shape = x.get_shape().as_list()[1:]
#bias = tf.get_variable("bias", shape=(1, n_outputs))
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
cores = []
for i in range(1, len(ranks)-1):
cores.append(tf.get_variable("core_{0}_output_{1}".format(i,suffix),
shape = (ranks[i-1], input_shape[i-1], ranks[i]),
initializer = weight_initializer))
cores.append(tf.get_variable("core_{0}_last_output_{1}".format(i,suffix),
shape=(ranks[-2],n_outputs,ranks[-1]),
initializer = weight_initializer))
regression_weights = TNSR.tt_to_tensor(cores)
return regression(x, regression_weights, input_shape, bias, n_outputs)
def cprl(x, rank, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
input_shape = x.get_shape().as_list()[1:]
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
rank1_tnsrs = []
print(rank)
for i in range(rank):
rank1_tnsr = []
for j in range(len(input_shape)):
rank1_tnsr.append(tf.get_variable("rank1_tnsr_{0}_{1}_{2}".format(i,j,np.prod(n_outputs)),
shape = (input_shape[j]),
initializer = weight_initializer))
rank1_tnsr.append(tf.get_variable("rank1_tnsr_{0}_output_{1}".format(i,np.prod(n_outputs)),
shape = (n_outputs),
initializer = weight_initializer))
rank1_tnsrs.append(rank1_tnsr)
regression_weights = TNSR.cp_to_tensor(rank1_tnsrs)
return regression(x, regression_weights, input_shape, bias, n_outputs)
def cprl(x, rank, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
input_shape = x.get_shape().as_list()[1:]
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
rank1_tnsrs = []
for i in range(rank):
rank1_tnsr = []
for j in range(len(input_shape)):
rank1_tnsr.append(tf.get_variable("rank1_tnsr_{0}_{1}_{2}".format(i,j,np.prod(n_outputs)),
shape = (input_shape[j]),
initializer = weight_initializer))
rank1_tnsr.append(tf.get_variable("rank1_tnsr_{0}_output_{1}".format(i,np.prod(n_outputs)),
shape = (n_outputs),
initializer = weight_initializer))
rank1_tnsrs.append(rank1_tnsr)
regression_weights = TNSR.cp_to_tensor(rank1_tnsrs)
return regression(x, regression_weights, input_shape, bias, n_outputs)
def trl(x, ranks, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
input_shape = x.get_shape().as_list()[1:]
core,factors = None, None
core = tf.get_variable("core_last", shape=ranks, initializer = weight_initializer)
factors = [ tf.get_variable("basic_factor_{0}_{1}".format(i,e),
shape=(input_shape[i],ranks[i]),
initializer = weight_initializer)
for (i, e) in enumerate(input_shape)
]
bias = tf.get_variable("bias_trl", shape=(1, n_outputs))
factors.append(tf.get_variable("factor_{}".format(len(ranks)-1),
shape=(n_outputs, ranks[-1]),
initializer = weight_initializer))
regression_weights = TNSR.tucker_to_tensor(core, factors)
x_0 = tf.reshape(x, [-1, np.prod(input_shape)])
w_minus1 = tf.reshape(regression_weights, [np.prod(input_shape),n_outputs])
return tf.add(tf.matmul(x_0, w_minus1) ,bias)
def trl(x, ranks, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
input_shape = x.get_shape().as_list()[1:]
core, factors = None, None
core = tf.get_variable("core_last",
shape=ranks,
initializer=weight_initializer)
factors = [
tf.get_variable("basic_factor_{0}_{1}".format(i, e),
shape=(input_shape[i], ranks[i]),
initializer=weight_initializer)
for (i, e) in enumerate(input_shape)
]
bias = tf.get_variable("bias_trl", shape=(1, n_outputs))
factors.append(
tf.get_variable("factor_{}".format(len(ranks) - 1),
shape=(n_outputs, ranks[-1]),
initializer=weight_initializer))
regression_weights = TNSR.tucker_to_tensor(core, factors)
x_0 = tf.reshape(x, [-1, np.prod(input_shape)])
w_minus1 = tf.reshape(regression_weights,
[np.prod(input_shape), n_outputs])
return tf.add(tf.matmul(x_0, w_minus1), bias)
def restore_2(self):
md = 2
t = TNSR.np_mode_dot(self.cores[0],
np.transpose(TNSR.np_unfold(self.cores[1], 0)),
md).reshape(
list(self.cores[0].shape)[:-1] +
list(self.cores[1].shape)[1:])
for i in range(1, len(self.tensor_size) - 1):
print(t.shape)
print(i)
md = md + 1
t = TNSR.np_mode_dot(
t, np.transpose(TNSR.np_unfold(self.cores[i + 1], 0)),
md).reshape(
list(t.shape)[:-1] + list(self.cores[i + 1].shape)[1:])
return t.reshape(self.tensor_size)
def ttrl(x, ranks, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
suffix = n_outputs
input_shape = x.get_shape().as_list()[1:]
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
cores = []
for i in range(1, len(ranks)-1):
cores.append(tf.get_variable("core_{0}_output_{1}".format(i,suffix),
shape = (ranks[i-1], input_shape[i-1], ranks[i]),
initializer = weight_initializer))
cores.append(tf.get_variable("core_{0}_last_output_{1}".format(i,suffix),
shape=(ranks[-2],n_outputs,ranks[-1]),
initializer = weight_initializer))
regression_weights = TNSR.tt_to_tensor(cores)
return regression(x, regression_weights, input_shape, bias, n_outputs)
def ttrl(x, ranks, n_outputs):
weight_initializer = tf.contrib.layers.xavier_initializer()
suffix = n_outputs
input_shape = x.get_shape().as_list()[1:]
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)),
shape=(1, np.prod(n_outputs)))
cores = []
for i in range(1, len(ranks) - 1):
cores.append(
tf.get_variable("core_{0}_output_{1}".format(i, suffix),
shape=(ranks[i - 1], input_shape[i - 1], ranks[i]),
initializer=weight_initializer))
cores.append(
tf.get_variable("core_{0}_last_output_{1}".format(i, suffix),
shape=(ranks[-2], n_outputs, ranks[-1]),
initializer=weight_initializer))
regression_weights = TNSR.tt_to_tensor(cores)
return regression(x, regression_weights, input_shape, bias, n_outputs)
def trl(x, ranks, n_outputs):
# n_outputs = [2,2,2,2,2,2]
# ranks = [1,2,2,1]
#FIXED_RANK = 3
weight_initializer = tf.contrib.layers.xavier_initializer()
input_shape = x.get_shape().as_list()[1:]
core,factors = None, None
"""
if(type(n_outputs) == type([])):
core = tf.get_variable("core_higher_{}".format(np.prod(n_outputs)), shape=ranks[1:]+[FIXED_RANK]*len(n_outputs),
initializer = weight_initializer)
factors = [tf.get_variable("basic_factor_higher_map_{0}_{1}".format(i,e),
shape=(input_shape[i],e),
initializer = weight_initializer)
for (i, e) in enumerate(ranks[1:])
]
else:
"""
core = tf.get_variable("core_last", shape=ranks, initializer = weight_initializer)
factors = [ tf.get_variable("basic_factor_{0}_{1}".format(i,e),
shape=(input_shape[i],ranks[i]),
initializer = weight_initializer)
for (i, e) in enumerate(input_shape)
]
bias = tf.get_variable("bias_{}".format(np.prod(n_outputs)), shape=(1, np.prod(n_outputs)))
#bias = tf.get_variable("bias", shape=(1, n_outputs))
"""
if(type(n_outputs) == type([])):
for i in range(len(n_outputs)):
factors.append(tf.get_variable("higher_factor_{0}_{1}".format(i,np.prod(n_outputs)),
shape=(n_outputs[i], FIXED_RANK),
initializer = weight_initializer
))
else:
"""
# append the last N+1 factor matrix to the list
factors.append(tf.get_variable("factor_{}".format(len(ranks)-1),
shape=(n_outputs, ranks[-1]),
initializer = weight_initializer))
regression_weights = TNSR.tucker_to_tensor(core, factors)
x_0 = tf.reshape(x, [-1, np.prod(input_shape)])
w_minus1 = None
"""
if type(n_outputs) == type([]):
w_minus1 = tf.reshape(regression_weights, [np.prod(input_shape), np.prod(n_outputs)])
return tf.reshape(tf.add(tf.matmul(x_0, w_minus1) ,bias), [-1]+n_outputs)
else:
"""
print(core.get_shape().as_list())
for f in factors:
print(f.get_shape().as_list())
w_minus1 = tf.reshape(regression_weights, [np.prod(input_shape),n_outputs])
return tf.add(tf.matmul(x_0, w_minus1) ,bias)