Loss Functions in Linear Regression
The choice of loss function can significantly impact the convergence of TensorFlow algorithms. We will compare and contrast using the L1 and L2 loss functions for linear regression.
L1 Loss
The L1 loss is the absolute value of the difference between the target (y) and the prediction (p):
L1(y,p) = |y - p|
L2 Loss
The L2 loss is the squared difference between the target (y) and the prediction (p):
L2(y,p) = (y - p)^2
Summary
L2 Loss | L1 Loss |
---|---|
More stable | Less Stable |
Not very robust | Robust |
Graph of L1 vs L2 Loss Functions
Graph of L1 vs L2 Loss Functions (L2 not converging)
Here is an example of the L2 function not converging. Despite a large learning rate, L1 has converged but L2 has not.
Graphical Summary of L1 and L2 with Learning Rates
Here is a plot of a 1D example of L1 and L2 loss with a small and large learning rate.
To note:
Top Left
- L1 loss with small learning rate: Robust, converges, but may take a while to converge.
Top Right
- L2 loss with small learning rate: Robust, converges, but may take a while to converge.
Bottom Left
- L1 loss with large learning rate: More robust, and less likely to explode, but may bounce around the optimum at the end.
Bottom Right
- L2 loss with large learning rate: Not robust, explodes because of large learning rate. Very sensitive to learning rate.
Moral of the story: When your algorithm isn’t converging, try decreasing the learning rate first.
04_lin_reg_l1_vs_l2.py
# Linear Regression: L1 vs L2
#----------------------------------
#
# This function shows how to use TensorFlow to
# solve linear regression via the matrix inverse.
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from sklearn import datasets
from tensorflow.python.framework import ops
ops.reset_default_graph()
# Create graph
#修改位置
config = tf.ConfigProto(allow_soft_placement= True, log_device_placement= True)
sess = tf.Session(config= config)
# Create graph
# sess = tf.Session()
# Load the data
# iris.data = [(Sepal Length, Sepal Width, Petal Length, Petal Width)]
iris = datasets.load_iris()
x_vals = np.array([x[3] for x in iris.data])
y_vals = np.array([y[0] for y in iris.data])
# Declare batch size and number of iterations
batch_size = 25
learning_rate = 0.4 # Will not converge with learning rate at 0.4
iterations = 50
# Initialize placeholders
x_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
y_target = tf.placeholder(shape=[None, 1], dtype=tf.float32)
# Create variables for linear regression
A = tf.Variable(tf.random_normal(shape=[1,1]))
b = tf.Variable(tf.random_normal(shape=[1,1]))
# Declare model operations
model_output = tf.add(tf.matmul(x_data, A), b)
# Declare loss functions
loss_l1 = tf.reduce_mean(tf.abs(y_target - model_output))
# Declare optimizers
my_opt_l1 = tf.train.GradientDescentOptimizer(learning_rate)
train_step_l1 = my_opt_l1.minimize(loss_l1)
# Initialize variables
init = tf.global_variables_initializer()
sess.run(init)
# Training loop
loss_vec_l1 = []
for i in range(iterations):
rand_index = np.random.choice(len(x_vals), size=batch_size)
rand_x = np.transpose([x_vals[rand_index]])
rand_y = np.transpose([y_vals[rand_index]])
sess.run(train_step_l1, feed_dict={x_data: rand_x, y_target: rand_y})
temp_loss_l1 = sess.run(loss_l1, feed_dict={x_data: rand_x, y_target: rand_y})
loss_vec_l1.append(temp_loss_l1)
if (i+1)%25==0:
print('Step #' + str(i+1) + ' A = ' + str(sess.run(A)) + ' b = ' + str(sess.run(b)))
# L2 Loss
# Reinitialize graph
ops.reset_default_graph()
# Create graph
#修改位置
config = tf.ConfigProto(allow_soft_placement= True, log_device_placement= True)
sess = tf.Session(config= config)
# Create graph
# sess = tf.Session()
# Initialize placeholders
x_data = tf.placeholder(shape=[None, 1], dtype=tf.float32)
y_target = tf.placeholder(shape=[None, 1], dtype=tf.float32)
# Create variables for linear regression
A = tf.Variable(tf.random_normal(shape=[1,1]))
b = tf.Variable(tf.random_normal(shape=[1,1]))
# Declare model operations
model_output = tf.add(tf.matmul(x_data, A), b)
# Declare loss functions
loss_l2 = tf.reduce_mean(tf.square(y_target - model_output))
# Declare optimizers
my_opt_l2 = tf.train.GradientDescentOptimizer(learning_rate)
train_step_l2 = my_opt_l2.minimize(loss_l2)
# Initialize variables
init = tf.global_variables_initializer()
sess.run(init)
loss_vec_l2 = []
for i in range(iterations):
rand_index = np.random.choice(len(x_vals), size=batch_size)
rand_x = np.transpose([x_vals[rand_index]])
rand_y = np.transpose([y_vals[rand_index]])
sess.run(train_step_l2, feed_dict={x_data: rand_x, y_target: rand_y})
temp_loss_l2 = sess.run(loss_l2, feed_dict={x_data: rand_x, y_target: rand_y})
loss_vec_l2.append(temp_loss_l2)
if (i+1)%25==0:
print('Step #' + str(i+1) + ' A = ' + str(sess.run(A)) + ' b = ' + str(sess.run(b)))
# Plot loss over time
plt.plot(loss_vec_l1, 'k-', label='L1 Loss')
plt.plot(loss_vec_l2, 'r--', label='L2 Loss')
plt.title('L1 and L2 Loss per Generation')
plt.xlabel('Generation')
plt.ylabel('L1 Loss')
plt.legend(loc='upper right')
plt.show()
InternalError: Blas SGEMM launch failed : a.shape=(25, 1), b.shape=(1, 1), m=25, n=1, k=1
[[Node: MatMul = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false, _device="/job:localhost/replica:0/task:0/gpu:0"](_recv_Placeholder_0/_7, Variable/read)]]