Module reclab.recommenders.llorma.llorma_lib.train_utils
LLORMA training utils
Expand source code
""" LLORMA training utils
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import tensorflow as tf
def init_session():
"""Initializes TF session
Returns
-------
obj: tf.Session
Returns TF Session
"""
# gpu_options = tf.GPUOptions(
# per_process_gpu_memory_fraction=GPU_MEMORY_FRAC)
# gpu_config = tf.ConfigProto(gpu_options=gpu_options)
# session = tf.Session(config=gpu_config)
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.compat.v1.Session(config=config)
session.run(tf.compat.v1.global_variables_initializer())
return session
def get_train_op(optimizer, loss, var_list):
""" Get a train operation
Parameters
----------
optimizer : obj
Valid TensorFlow optimizer,
e.g. tf.train.GradientDescentOptimizer
loss : obj
TF variable
var_list : obj
List of TF variables
"""
gvs = optimizer.compute_gradients(loss, var_list=var_list)
# capped_gvs = [(tf.clip_by_value(grad, -100.0, 100.0), var)
# for grad, var in gvs]
capped_gvs = gvs
train_op = optimizer.apply_gradients(capped_gvs)
return train_op
def init_latent_mat(n, rank, mu_val, std_val):
"""Initialize a matrix for the latent factors
Parameters
----------
n : int
Number of user/items
rank : int
Size of the latent dimension
mu_val : float
Unscaled mean value
std_val : float
Unscaled standard deviation value
"""
_mu = math.sqrt(mu_val / rank)
_std = math.sqrt((math.sqrt(mu_val * mu_val + std_val * std_val) - mu_val) / rank)
return tf.Variable(
tf.compat.v1.truncated_normal([n, rank], _mu, _std, dtype=tf.float64))Functions
def get_train_op(optimizer, loss, var_list)-
Get a train operation
Parameters
optimizer:obj- Valid TensorFlow optimizer, e.g. tf.train.GradientDescentOptimizer
loss:obj- TF variable
var_list:obj- List of TF variables
Expand source code
def get_train_op(optimizer, loss, var_list): """ Get a train operation Parameters ---------- optimizer : obj Valid TensorFlow optimizer, e.g. tf.train.GradientDescentOptimizer loss : obj TF variable var_list : obj List of TF variables """ gvs = optimizer.compute_gradients(loss, var_list=var_list) # capped_gvs = [(tf.clip_by_value(grad, -100.0, 100.0), var) # for grad, var in gvs] capped_gvs = gvs train_op = optimizer.apply_gradients(capped_gvs) return train_op def init_latent_mat(n, rank, mu_val, std_val)-
Initialize a matrix for the latent factors
Parameters
n:int- Number of user/items
rank:int- Size of the latent dimension
mu_val:float- Unscaled mean value
std_val:float- Unscaled standard deviation value
Expand source code
def init_latent_mat(n, rank, mu_val, std_val): """Initialize a matrix for the latent factors Parameters ---------- n : int Number of user/items rank : int Size of the latent dimension mu_val : float Unscaled mean value std_val : float Unscaled standard deviation value """ _mu = math.sqrt(mu_val / rank) _std = math.sqrt((math.sqrt(mu_val * mu_val + std_val * std_val) - mu_val) / rank) return tf.Variable( tf.compat.v1.truncated_normal([n, rank], _mu, _std, dtype=tf.float64)) def init_session()-
Initializes TF session
Returns
obj:tf.Session- Returns TF Session
Expand source code
def init_session(): """Initializes TF session Returns ------- obj: tf.Session Returns TF Session """ # gpu_options = tf.GPUOptions( # per_process_gpu_memory_fraction=GPU_MEMORY_FRAC) # gpu_config = tf.ConfigProto(gpu_options=gpu_options) # session = tf.Session(config=gpu_config) config = tf.compat.v1.ConfigProto() config.gpu_options.allow_growth = True session = tf.compat.v1.Session(config=config) session.run(tf.compat.v1.global_variables_initializer()) return session