I trying to understand linear regression... here is script that I tried to understand:
'''
A linear regression learning algorithm example using TensorFlow library.
Author: Aymeric Damien
Project: https://github.com/aymericdamien/TensorFlow-Examples/
'''
from __future__ import print_function
import tensorflow as tf
from numpy import *
import numpy
import matplotlib.pyplot as plt
rng = numpy.random
# Parameters
learning_rate = 0.0001
training_epochs = 1000
display_step = 50
# Training Data
train_X = numpy.asarray([3.3,4.4,5.5,6.71,6.93,4.168,9.779,6.182,7.59,2.167,
7.042,10.791,5.313,7.997,5.654,9.27,3.1])
train_Y = numpy.asarray([1.7,2.76,2.09,3.19,1.694,1.573,3.366,2.596,2.53,1.221,
2.827,3.465,1.65,2.904,2.42,2.94,1.3])
train_X=numpy.asarray(train_X)
train_Y=numpy.asarray(train_Y)
n_samples = train_X.shape[0]
# tf Graph Input
X = tf.placeholder("float")
Y = tf.placeholder("float")
# Set model weights
W = tf.Variable(rng.randn(), name="weight")
b = tf.Variable(rng.randn(), name="bias")
# Construct a linear model
pred = tf.add(tf.multiply(X, W), b)
# Mean squared error
cost = tf.reduce_sum(tf.pow(pred-Y, 2))/(2*n_samples)
# Gradient descent
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# Initializing the variables
init = tf.global_variables_initializer()
# Launch the graph
with tf.Session() as sess:
sess.run(init)
# Fit all training data
for epoch in range(training_epochs):
for (x, y) in zip(train_X, train_Y):
sess.run(optimizer, feed_dict={X: x, Y: y})
# Display logs per epoch step
if (epoch+1) % display_step == 0:
c = sess.run(cost, feed_dict={X: train_X, Y:train_Y})
print("Epoch:", '%04d' % (epoch+1), "cost=", "{:.9f}".format(c), \
"W=", sess.run(W), "b=", sess.run(b))
print("Optimization Finished!")
training_cost = sess.run(cost, feed_dict={X: train_X, Y: train_Y})
print("Training cost=", training_cost, "W=", sess.run(W), "b=", sess.run(b), '\n')
# Graphic display
plt.plot(train_X, train_Y, 'ro', label='Original data')
plt.plot(train_X, sess.run(W) * train_X + sess.run(b), label='Fitted line')
plt.legend()
plt.show()
Question is what this part represent:
# Set model weights
W = tf.Variable(rng.randn(), name="weight")
b = tf.Variable(rng.randn(), name="bias")
And why are there random float numbers?
Also could you show me some math with formals represents cost, pred, optimizer variables?
let's try to put up some intuition&sources together with the tfapproach.
General intuition:
Regression as presented here is a supervised learning problem. In it, as defined in Russel&Norvig's Artificial Intelligence, the task is:
given a training set (X, y) of m input-output pairs (x1, y1), (x2, y2), ... , (xm, ym), where each output was generated by an unknown function y = f(x), discover a function h that approximates the true function f
For that sake, the h hypothesis function combines somehow each x with the to-be-learned parameters, in order to have an output that is as close to the corresponding y as possible, and this for the whole dataset. The hope is that the resulting function will be close to f.
But how to learn this parameters? in order to be able to learn, the model has to be able to evaluate. Here comes the cost (also called loss, energy, merit...) function to play: it is a metric function that compares the output of h with the corresponding y, and penalizes big differences.
Now it should be clear what is exactly the "learning" process here: alter the parameters in order to achieve a lower value for the cost function.
Linear Regression:
The example that you are posting performs a parametric linear regression, optimized with gradient descent based on the mean squared error as cost function. Which means:
Parametric: The set of parameters is fixed. They are held in the exact same memory placeholders thorough the learning process.
Linear: The output of h is merely a linear (actually, affine) combination between the input x and your parameters. So if x and w are real-valued vectors of the same dimensionality, and b is a real number, it holds that h(x,w, b)= w.transposed()*x+b. Page 107 of the Deep Learning Book brings more quality insights and intuitions into that.
Cost function: Now this is the interesting part. The average squared error is a convex function. This means it has a single, global optimum, and furthermore, it can be directly found with the set of normal equations (also explained in the DLB). In the case of your example, the stochastic (and/or minibatch) gradient descent method is used: this is the preferred method when optimizing non-convex cost functions (which is the case in more advanced models like neural networks) or when your dataset has a huge dimensionality (also explained in the DLB).
Gradient descent: tf deals with this for you, so it is enough to say that GD minimizes the cost function by following its derivative "downwards", in small steps, until reaching a saddle point. If you totally need to know, the exact technique applied by TF is called automatic differentiation, kind of a compromise between the numeric and symbolic approaches. For convex functions like yours this point will be the global optimum, and (if your learning rate is not too big) it will always converge to it, so it doesn't matter which values you initialize your Variables with. The random initialization is necessary in more complex architectures like neural networks. There is some extra code regarding the management of the minibatches, but I won't get into that because it is not the main focus of your question.
The TensorFlow approach:
Deep Learning frameworks are nowadays about nesting lots of functions by building computational graphs (you may want to take a look at the presentation on DL frameworks that I did some weeks ago). For constructing and running the graph, TensoFlow follows a declarative style, which means that the graph has to be first completely defined and compiled, before it is deployed and executed. It is very reccommended to read this short wiki article, if you haven't yet. In this context, the setup is split in two parts:
Firstly, you define your computational Graph, where you put your dataset and parameters in memory placeholders, define the hypothesis and cost functions building on them, and tell tf which optimization technique to apply.
Then you run the computation in a Session and the library will be able to (re)load the data placeholders and perform the optimization.
The code:
The code of the example follows this approach closely:
Define the test data X and labels Y, and prepare a placeholder in the Graph for them (which is fed in the feed_dict part).
Define the 'W' and 'b' placeholders for the parameters. They have to be Variables because they will be updated during the Session.
Define pred (our hypothesis) and cost as explained before.
From this, the rest of the code should be clearer. Regarding the optimizer, as I said, tf already knows how to deal with this but you may want to look into gradient descent for more details (again, the DLB is a pretty good reference for that)
Cheers!
Andres
CODE EXAMPLES: GRADIENT DESCENT VS. NORMAL EQUATIONS
This small snippets generate simple multi-dimensional datasets and test both approaches. Notice that the normal equations approach doesn't require looping, and brings better results. For small dimensionality (DIMENSIONS<30k) is probably the preferred approach:
from __future__ import absolute_import, division, print_function
import numpy as np
import tensorflow as tf
####################################################################################################
### GLOBALS
####################################################################################################
DIMENSIONS = 5
f = lambda(x): sum(x) # the "true" function: f = 0 + 1*x1 + 1*x2 + 1*x3 ...
noise = lambda: np.random.normal(0,10) # some noise
####################################################################################################
### GRADIENT DESCENT APPROACH
####################################################################################################
# dataset globals
DS_SIZE = 5000
TRAIN_RATIO = 0.6 # 60% of the dataset is used for training
_train_size = int(DS_SIZE*TRAIN_RATIO)
_test_size = DS_SIZE - _train_size
ALPHA = 1e-8 # learning rate
LAMBDA = 0.5 # L2 regularization factor
TRAINING_STEPS = 1000
# generate the dataset, the labels and split into train/test
ds = [[np.random.rand()*1000 for d in range(DIMENSIONS)] for _ in range(DS_SIZE)] # synthesize data
# ds = normalize_data(ds)
ds = [(x, [f(x)+noise()]) for x in ds] # add labels
np.random.shuffle(ds)
train_data, train_labels = zip(*ds[0:_train_size])
test_data, test_labels = zip(*ds[_train_size:])
# define the computational graph
graph = tf.Graph()
with graph.as_default():
# declare graph inputs
x_train = tf.placeholder(tf.float32, shape=(_train_size, DIMENSIONS))
y_train = tf.placeholder(tf.float32, shape=(_train_size, 1))
x_test = tf.placeholder(tf.float32, shape=(_test_size, DIMENSIONS))
y_test = tf.placeholder(tf.float32, shape=(_test_size, 1))
theta = tf.Variable([[0.0] for _ in range(DIMENSIONS)])
theta_0 = tf.Variable([[0.0]]) # don't forget the bias term!
# forward propagation
train_prediction = tf.matmul(x_train, theta)+theta_0
test_prediction = tf.matmul(x_test, theta) +theta_0
# cost function and optimizer
train_cost = (tf.nn.l2_loss(train_prediction - y_train)+LAMBDA*tf.nn.l2_loss(theta))/float(_train_size)
optimizer = tf.train.GradientDescentOptimizer(ALPHA).minimize(train_cost)
# test results
test_cost = (tf.nn.l2_loss(test_prediction - y_test)+LAMBDA*tf.nn.l2_loss(theta))/float(_test_size)
# run the computation
with tf.Session(graph=graph) as s:
tf.initialize_all_variables().run()
print("initialized"); print(theta.eval())
for step in range(TRAINING_STEPS):
_, train_c, test_c = s.run([optimizer, train_cost, test_cost],
feed_dict={x_train: train_data, y_train: train_labels,
x_test: test_data, y_test: test_labels })
if (step%100==0):
# it should return bias close to zero and parameters all close to 1 (see definition of f)
print("\nAfter", step, "iterations:")
#print(" Bias =", theta_0.eval(), ", Weights = ", theta.eval())
print(" train cost =", train_c); print(" test cost =", test_c)
PARAMETERS_GRADDESC = tf.concat(0, [theta_0, theta]).eval()
print("Solution for parameters:\n", PARAMETERS_GRADDESC)
####################################################################################################
### NORMAL EQUATIONS APPROACH
####################################################################################################
# dataset globals
DIMENSIONS = 5
DS_SIZE = 5000
TRAIN_RATIO = 0.6 # 60% of the dataset isused for training
_train_size = int(DS_SIZE*TRAIN_RATIO)
_test_size = DS_SIZE - _train_size
f = lambda(x): sum(x) # the "true" function: f = 0 + 1*x1 + 1*x2 + 1*x3 ...
noise = lambda: np.random.normal(0,10) # some noise
# training globals
LAMBDA = 1e6 # L2 regularization factor
# generate the dataset, the labels and split into train/test
ds = [[np.random.rand()*1000 for d in range(DIMENSIONS)] for _ in range(DS_SIZE)]
ds = [([1]+x, [f(x)+noise()]) for x in ds] # add x[0]=1 dimension and labels
np.random.shuffle(ds)
train_data, train_labels = zip(*ds[0:_train_size])
test_data, test_labels = zip(*ds[_train_size:])
# define the computational graph
graph = tf.Graph()
with graph.as_default():
# declare graph inputs
x_train = tf.placeholder(tf.float32, shape=(_train_size, DIMENSIONS+1))
y_train = tf.placeholder(tf.float32, shape=(_train_size, 1))
theta = tf.Variable([[0.0] for _ in range(DIMENSIONS+1)]) # implicit bias!
# optimum
optimum = tf.matrix_solve_ls(x_train, y_train, LAMBDA, fast=True)
# run the computation: no loop needed!
with tf.Session(graph=graph) as s:
tf.initialize_all_variables().run()
print("initialized")
opt = s.run(optimum, feed_dict={x_train:train_data, y_train:train_labels})
PARAMETERS_NORMEQ = opt
print("Solution for parameters:\n",PARAMETERS_NORMEQ)
####################################################################################################
### PREDICTION AND ERROR RATE
####################################################################################################
# generate test dataset
ds = [[np.random.rand()*1000 for d in range(DIMENSIONS)] for _ in range(DS_SIZE)]
ds = [([1]+x, [f(x)+noise()]) for x in ds] # add x[0]=1 dimension and labels
test_data, test_labels = zip(*ds)
# define hypothesis
h_gd = lambda(x): PARAMETERS_GRADDESC.T.dot(x)
h_ne = lambda(x): PARAMETERS_NORMEQ.T.dot(x)
# define cost
mse = lambda pred, lab: ((pred-np.array(lab))**2).sum()/DS_SIZE
# make predictions!
predictions_gd = np.array([h_gd(x) for x in test_data])
predictions_ne = np.array([h_ne(x) for x in test_data])
# calculate and print total error
cost_gd = mse(predictions_gd, test_labels)
cost_ne = mse(predictions_ne, test_labels)
print("total cost with gradient descent:", cost_gd)
print("total cost with normal equations:", cost_ne)
Variables allow us to add trainable parameters to a graph. They are constructed with a type and initial value:
W = tf.Variable([.3], tf.float32)
b = tf.Variable([-.3], tf.float32)
x = tf.placeholder(tf.float32)
linear_model = W * x + b
The variable with type tf.Variable is the parameter which we will learn use TensorFlow. Assume you use the gradient descent to minimize the loss function. You need initial these parameter first. The rng.randn() is used to generate a random value for this purpose.
I think the Getting Started With TensorFlow is a good start point for you.
I'll first define the variables:
W is a multidimensional line that spans R^d (same dimensionality as X)
b is a scalar value (bias)
Y is also a scalar value i.e. the value at X
pred = W (dot) X + b # dot here refers to dot product
# cost equals the average squared error
cost = ((pred - Y)^2) / 2*num_samples
#finally optimizer
# optimizer computes the gradient with respect to each variable and the update
W += learning_rate * (pred - Y)/num_samples * X
b += learning_rate * (pred - Y)/num_samples
Why are W and b set to random well this updates based on gradients from the error calculated from the cost so W and b could have been initialized to anything. It isn't performing linear regression via least squares method although both will converge to the same solution.
Look here for more information: Getting Started
Related
I'm trying to implement and train a neural network using the JAX library and its little neural network submodule, "Stax". Since this library doesn't come with an implementation of binary cross entropy, I wrote my own:
def binary_cross_entropy(y_hat, y):
bce = y * jnp.log(y_hat) + (1 - y) * jnp.log(1 - y_hat)
return jnp.mean(-bce)
I implemented a simple neural network and trained it on MNIST, and started to get suspicious of some of the results I was getting. So I implemented the same setup in Keras, and I immediately got wildly different results! The same model, trained in the same way on the same data, was getting 90% training accuracy in Keras instead of around 50% in JAX. Eventually I tracked down part of the issue to my naive implementation of cross-entropy, which is supposedly numerically unstable. Following this post and this code I found, I wrote the following new version:
def binary_cross_entropy_stable(y_hat, y):
y_hat = jnp.clip(y_hat, 0.000001, 0.9999999)
logits = jnp.log(y_hat/(1 - y_hat))
max_logit = jnp.clip(logits, 0, None)
bces = logits - logits * y + max_logit + jnp.log(jnp.exp(-max_logit) + jnp.exp(-logits - max_logit))
return jnp.mean(bces)
This works a little better. Now my JAX implementation gets up to 80% train accuracy, but that's still a lot less than the 90% Keras gets. What I want to know is what is going on? Why are my two implementations not behaving the same way?
Below, I condensed my two implementations down to a single script. In this script, I implement the same model in JAX and in Keras. I initialize both with the same weights, and train them using full-batch gradient descent for 10 steps on 1000 datapoints from MNIST, the same data for each model. JAX finishes with 80% training accuracy, while Keras finishes with 90%. Specifically, I get this output:
Initial Keras accuracy: 0.4350000023841858
Initial JAX accuracy: 0.435
Final JAX accuracy: 0.792
Final Keras accuracy: 0.9089999794960022
JAX accuracy (Keras weights): 0.909
Keras accuracy (JAX weights): 0.7919999957084656
And actually, when I vary the conditions a little (using different random initial weights or a different training set), sometimes I get back the 50% JAX accuracy and 90% Keras accuracy.
I swap the weights at the end to verify that the weights obtained from training are indeed the issue, not something to do with the actual computation of the network predictions, or the way I calculate accuracy.
The code:
import numpy as np
import jax
from jax import jit, grad
from jax.experimental import stax, optimizers
import jax.numpy as jnp
import keras
import keras.datasets.mnist
def binary_cross_entropy(y_hat, y):
bce = y * jnp.log(y_hat) + (1 - y) * jnp.log(1 - y_hat)
return jnp.mean(-bce)
def binary_cross_entropy_stable(y_hat, y):
y_hat = jnp.clip(y_hat, 0.000001, 0.9999999)
logits = jnp.log(y_hat/(1 - y_hat))
max_logit = jnp.clip(logits, 0, None)
bces = logits - logits * y + max_logit + jnp.log(jnp.exp(-max_logit) + jnp.exp(-logits - max_logit))
return jnp.mean(bces)
def binary_accuracy(y_hat, y):
return jnp.mean((y_hat >= 1/2) == (y >= 1/2))
########################################
# #
# Create dataset #
# #
########################################
input_dimension = 784
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data(path="mnist.npz")
xs = np.concatenate([x_train, x_test])
xs = xs.reshape((70000, 784))
ys = np.concatenate([y_train, y_test])
ys = (ys >= 5).astype(np.float32)
ys = ys.reshape((70000, 1))
train_xs = xs[:1000]
train_ys = ys[:1000]
########################################
# #
# Create JAX model #
# #
########################################
jax_initializer, jax_model = stax.serial(
stax.Dense(1000),
stax.Relu,
stax.Dense(1),
stax.Sigmoid
)
rng_key = jax.random.PRNGKey(0)
_, initial_jax_weights = jax_initializer(rng_key, (1, input_dimension))
########################################
# #
# Create Keras model #
# #
########################################
initial_keras_weights = [*initial_jax_weights[0], *initial_jax_weights[2]]
keras_model = keras.Sequential([
keras.layers.Dense(1000, activation="relu"),
keras.layers.Dense(1, activation="sigmoid")
])
keras_model.compile(
optimizer=keras.optimizers.SGD(learning_rate=0.01),
loss=keras.losses.binary_crossentropy,
metrics=["accuracy"]
)
keras_model.build(input_shape=(1, input_dimension))
keras_model.set_weights(initial_keras_weights)
if __name__ == "__main__":
########################################
# #
# Compare untrained models #
# #
########################################
initial_keras_predictions = keras_model.predict(train_xs, verbose=0)
initial_jax_predictions = jax_model(initial_jax_weights, train_xs)
_, keras_initial_accuracy = keras_model.evaluate(train_xs, train_ys, verbose=0)
jax_initial_accuracy = binary_accuracy(jax_model(initial_jax_weights, train_xs), train_ys)
print("Initial Keras accuracy:", keras_initial_accuracy)
print("Initial JAX accuracy:", jax_initial_accuracy)
########################################
# #
# Train JAX model #
# #
########################################
L = jit(binary_cross_entropy_stable)
gradL = jit(grad(lambda w, x, y: L(jax_model(w, x), y)))
opt_init, opt_apply, get_params = optimizers.sgd(0.01)
network_state = opt_init(initial_jax_weights)
for _ in range(10):
wT = get_params(network_state)
gradient = gradL(wT, train_xs, train_ys)
network_state = opt_apply(
0,
gradient,
network_state
)
final_jax_weights = get_params(network_state)
final_jax_training_predictions = jax_model(final_jax_weights, train_xs)
final_jax_accuracy = binary_accuracy(final_jax_training_predictions, train_ys)
print("Final JAX accuracy:", final_jax_accuracy)
########################################
# #
# Train Keras model #
# #
########################################
for _ in range(10):
keras_model.fit(
train_xs,
train_ys,
epochs=1,
batch_size=1000,
verbose=0
)
final_keras_loss, final_keras_accuracy = keras_model.evaluate(train_xs, train_ys, verbose=0)
print("Final Keras accuracy:", final_keras_accuracy)
########################################
# #
# Swap weights #
# #
########################################
final_keras_weights = keras_model.get_weights()
final_keras_weights_in_jax_format = [
(final_keras_weights[0], final_keras_weights[1]),
tuple(),
(final_keras_weights[2], final_keras_weights[3]),
tuple()
]
jax_accuracy_with_keras_weights = binary_accuracy(
jax_model(final_keras_weights_in_jax_format, train_xs),
train_ys
)
print("JAX accuracy (Keras weights):", jax_accuracy_with_keras_weights)
final_jax_weights_in_keras_format = [*final_jax_weights[0], *final_jax_weights[2]]
keras_model.set_weights(final_jax_weights_in_keras_format)
_, keras_accuracy_with_jax_weights = keras_model.evaluate(train_xs, train_ys, verbose=0)
print("Keras accuracy (JAX weights):", keras_accuracy_with_jax_weights)
Try changing the PRNG seed at line 57 to a value other than 0 to run the experiment using different initial weights.
Your binary_cross_entropy_stable function does not match the output of keras.binary_crossentropy; for example:
x = np.random.rand(10)
y = np.random.rand(10)
print(keras.losses.binary_crossentropy(x, y))
# tf.Tensor(0.8134677734043875, shape=(), dtype=float64)
print(binary_cross_entropy_stable(x, y))
# 0.9781515
That is where I would start if you're trying to exactly duplicate the model.
You can view the source of the keras loss function here: keras/losses.py#L1765-L1810, with the main part of the implementation here: keras/backend.py#L4972-L5017
One detail: it appears that with a sigmoid activation function, Keras re-uses some cached logits to compute the binary cross entropy while avoiding problematic values: keras/backend.py#L4988-L4997. I'm not sure how to easily replicate that behavior using JAX & stax.
context: I am trying to create a generic function to optimize the cost of any regression problem using polynomial regression (of any specified degree).
I am trying to fit my model to the load_boston dataset (with the house price as the label and 13 features).
I used multiple degrees of polynomials, and multiple learning rates and epochs (with gradient descent) and the MSE is coming out to be so high even on the training dataset (I am using 100% of the data to train the model, and I am checking the cost on the same data, but the MSE cost is still very high).
import tensorflow as tf
from sklearn.datasets import load_boston
def polynomial(x, coeffs):
y = 0
for i in range(len(coeffs)):
y += coeffs[i]*x**i
return y
def initial_parameters(dimensions, data_type, degree): # list number of dims/features and degree
thetas = [tf.Variable(0, dtype=data_type)] # the constant theta/bias
for i in range(degree):
thetas.append(tf.Variable( tf.zeros([dimensions, 1], dtype=data_type)))
return thetas
def regression_error(x, y, thetas):
hx = thetas[0] # constant thetas - no need to have 1 for each variable (e.g x^0*th + y^0*th...)
for i in range(1, len(thetas)):
hx = tf.add(hx, tf.matmul( tf.pow(x, i), thetas[i]))
return tf.reduce_mean(tf.squared_difference(hx, y))
def polynomial_regression(x, y, data_type, degree, learning_rate, epoch): #features=dimensions=variables
thetas = initial_parameters(x.shape[1], data_type, degree)
cost = regression_error(x, y, thetas)
init = tf.initialize_all_variables()
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
with tf.Session() as sess:
sess.run(init)
for epoch in range(epoch):
sess.run(optimizer)
return cost.eval()
x, y = load_boston(True) # yes just use the entire dataset
for deg in range(1, 2):
for lr in range(-8, -5):
error = polynomial_regression(x, y, tf.float64, deg, 10**lr, 100 )
print (deg, lr, error)
It outputs 97.3 even though most of the labels are around 30 (degree = 1, learning rate = 10^-6).
what is wrong with the code?
The problem is that the different features are on different orders of magnitude and hence are not compatible with the learning rate which is the same for all features. Even more, when using a non-zero variable initialization, one has to make sure that these initial values are as well compatible with the feature values.
In [1]: from sklearn.datasets import load_boston
In [2]: x, y = load_boston(True)
In [3]: x.std(axis=0)
Out[3]:
array([8.58828355e+00, 2.32993957e+01, 6.85357058e+00, 2.53742935e-01,
1.15763115e-01, 7.01922514e-01, 2.81210326e+01, 2.10362836e+00,
8.69865112e+00, 1.68370495e+02, 2.16280519e+00, 9.12046075e+01,
7.13400164e+00])
In [4]: x.mean(axis=0)
Out[4]:
array([3.59376071e+00, 1.13636364e+01, 1.11367787e+01, 6.91699605e-02,
5.54695059e-01, 6.28463439e+00, 6.85749012e+01, 3.79504269e+00,
9.54940711e+00, 4.08237154e+02, 1.84555336e+01, 3.56674032e+02,
1.26530632e+01])
A common approach is to normalize the input data (e.g. to have zero mean and unit variance) and to choose the initial weights randomly (e.g. normal distribution, std.dev. = 1). sklearn.preprocessing offers various functionality for these cases.
PolynomialFeatures can be used to generate the polynomial features automatically.
StandardScaler scales the data to zero mean and unit variance.
pipeline.Pipeline can be used for convenience to combine these preprocessing steps.
The polynomial_regression function then reduces to:
pipeline = Pipeline([
('poly', PolynomialFeatures(degree)),
('scaler', StandardScaler())
])
x = pipeline.fit_transform(x)
thetas = tf.Variable(tf.random_normal([x.shape[1], 1], dtype=data_type))
cost = tf.reduce_mean(tf.squared_difference(tf.matmul(x, thetas), y))
# Perform variable initialization and optimizer instantiation here.
# Run optimization over epochs.
i'm struggling to understand tensorflow, and I can't find good basic examples that don't rely on the MNIST dataset. I've tried to create a classification nn for some public datasets where they provide a number of (unknown) features, and a label for each sample. There's one where they provide around 90 features of audio analysis, and the year of publication as the label. (https://archive.ics.uci.edu/ml/datasets/yearpredictionmsd)
Needless to say, I didn't manage to train the network, and little could I do for understanding the provided features.
I'm now trying to generate artificial data, and try to train a network around it. The data are pairs of number (position), and the label is 1 if that position is inside a circle of radius r around an arbitrary point (5,5).
numrows=10000
circlex=5
circley=5
circler=3
data = np.random.rand(numrows,2)*10
labels = [ math.sqrt( math.pow(x-circlex, 2) + math.pow(y-circley, 2) ) for x,y in data ]
labels = list( map(lambda x: x<circler, labels) )
If tried many combinations of network shape, parameters, optimizers, learning rates, etc (I admit the math is not strong on this one), but eithere there's no convergence, or it sucks (70% accuracy on last test).
Current version (labels converted to one_hot encoding [1,0] and [0,1] (outside, inside).
# model creation
graph=tf.Graph()
with graph.as_default():
X = tf.placeholder(tf.float32, [None, 2] )
layer1 = tf.layers.dense(X, 2)
layer2 = tf.layers.dense(layer1, 2)
Y = tf.nn.softmax(layer2)
y_true = tf.placeholder(tf.float32, [None, 2] )
loss=tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits_v2(logits=Y, labels=y_true) )
optimizer = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
/ predictions.shape[0])
# training
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
for step in range(1000):
_, l, predictions = session.run([optimizer,loss,Y], feed_dict={X:data, y_true:labels})
if step % 100 == 0:
print("Loss at step %d: %f" % (step, l)
print("Accuracy %f" % accuracy(predictions, labels))
The acuracy in this example is around 70% (loss around 0.6).
The question is... what am I doing wrong?
UPDATE
I'm leaving the question as originally asked. Main lessons I learned:
Normalize your input data. The mean should be around 0, and the range ~ between -1 and 1.
Blue: normalized data, Red: raw input data as created above
Batch your input data. If the subsets used are random enough, it decreases the number of iterations needed without hurting accuracy too much.
Don't forget activation functions between layers :)
The input:
Plotting the synthetic data with two classes.
Output from the code above:
All outputs are classified as a single class and because of class imbalance, accuracy is high 70%.
Issues with the code
Even though there are two layers defined, no activation function defined between the two. So tf.softmax( ((x*w1)+b1) * w2 + b2) squashes down to a single layer. There is just a single hyperplane trying to separate this input and the hyperplane lies outside the input space, thats why you get all inputs classified as a single class.
Bug: Softmax is applied twice: on the logits as well as during entropy_loss.
The entire input is given as a single batch, instead of mini-batches.
Inputs need to be normalized.
Fixing the above issues and the output becomes:
The above output makes sense, as the model has two hidden layers and so we have two hyperplanes trying to separate the data. The final layer then combines these two hyperplanes in such a way to minimize error.
Increasing the hidden layer from 2 to 3:
With 3 hidden layers, we get 3 hyperplanes and we can see the final layer adjusts these hyperplanes to separate the data well.
Code:
# Normalize data
data = (data - np.mean(data)) /np.sqrt(np.var(data))
n_hidden = 3
batch_size = 128
# Feed batch data
def get_batch(inputX, inputY, batch_size):
duration = len(inputX)
for i in range(0,duration//batch_size):
idx = i*batch_size
yield inputX[idx:idx+batch_size], inputY[idx:idx+batch_size]
# Create the graph
tf.reset_default_graph()
graph=tf.Graph()
with graph.as_default():
X = tf.placeholder(tf.float32, [None, 2] )
layer1 = tf.layers.dense(X, n_hidden, activation=tf.nn.sigmoid)
layer2 = tf.layers.dense(layer1, 2)
Y = tf.nn.softmax(layer2)
y_true = tf.placeholder(tf.int32, [None] )
loss = tf.losses.sparse_softmax_cross_entropy(logits=layer2, labels=y_true)
optimizer = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(Y, 1),tf.argmax(tf.one_hot(y_true,2), 1)), tf.float32))
# training
with tf.Session(graph=graph) as session:
session.run(tf.global_variables_initializer())
for epoch in range(10):
acc_avg = 0.
loss_avg = 0.
for step in range(10000//batch_size):
for inputX, inputY in get_batch(data, labels, batch_size):
_, l, acc = session.run([optimizer,loss,accuracy], feed_dict={X:inputX, y_true:inputY})
acc_avg += acc
loss_avg += l
print("Loss at step %d: %f" % (step, loss_avg*batch_size/10000))
print("Accuracy %f" % (acc_avg*batch_size/10000))
#Get prediction
pred = session.run(Y, feed_dict={X:data})
# Plotting function
import matplotlib.pylab as plt
plt.scatter(data[:,0], data[:,1], s=20, c=np.argmax(pred,1), cmap='jet', vmin=0, vmax=1)
plt.show()
I am trying to implement a linear regression model in Tensorflow, with additional constraints (coming from the domain) that the W and b terms must be non-negative.
I believe there are a couple of ways to do this.
We can modify the cost function to penalize negative weights [Lagrangian approach] [See:TensorFlow - best way to implement weight constraints
We can compute the gradients ourselves and project them on [0, infinity] [Projected gradient approach]
Approach 1: Lagrangian
When I tried the first approach, I would often end up with negative b.
I had modified the cost function from:
cost = tf.reduce_sum(tf.pow(pred-Y, 2))/(2*n_samples)
to:
cost = tf.reduce_sum(tf.pow(pred-Y, 2))/(2*n_samples)
nn_w = tf.reduce_sum(tf.abs(W) - W)
nn_b = tf.reduce_sum(tf.abs(b) - b)
constraint = 100.0*nn_w + 100*nn_b
cost_with_constraint = cost + constraint
Keeping the coefficient of nn_b and nn_w to be very high leads to instability and very high cost.
Here is the complete code.
import numpy as np
import tensorflow as tf
n_samples = 50
train_X = np.linspace(1, 50, n_samples)
train_Y = 10*train_X + 6 +40*np.random.randn(50)
X = tf.placeholder("float")
Y = tf.placeholder("float")
# Set model weights
W = tf.Variable(np.random.randn(), name="weight")
b = tf.Variable(np.random.randn(), name="bias")
# Construct a linear model
pred = tf.add(tf.multiply(X, W), b)
# Gradient descent
learning_rate=0.0001
# Initializing the variables
init = tf.global_variables_initializer()
# Mean squared error
cost = tf.reduce_sum(tf.pow(pred-Y, 2))/(2*n_samples)
nn_w = tf.reduce_sum(tf.abs(W) - W)
nn_b = tf.reduce_sum(tf.abs(b) - b)
constraint = 1.0*nn_w + 100*nn_b
cost_with_constraint = cost + constraint
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost_with_constraint)
training_epochs=200
with tf.Session() as sess:
sess.run(init)
# Fit all training data
cost_array = np.zeros(training_epochs)
W_array = np.zeros(training_epochs)
b_array = np.zeros(training_epochs)
for epoch in range(training_epochs):
for (x, y) in zip(train_X, train_Y):
sess.run(optimizer, feed_dict={X: x, Y: y})
W_array[epoch] = sess.run(W)
b_array[epoch] = sess.run(b)
cost_array[epoch] = sess.run(cost, feed_dict={X: train_X, Y: train_Y})
The following is the mean of b across 10 different runs.
0 -1.101268
1 0.169225
2 0.158363
3 0.706270
4 -0.371205
5 0.244424
6 1.312516
7 -0.069609
8 -1.032187
9 -1.711668
Clearly, the first approach is not optimal. Further, there is a lot of art involved in choosing the coefficient of penalty terms.
Approach 2: Projected gradient
I then thought to use the second approach, which is more guaranteed to work.
gr = tf.gradients(cost, [W, b])
We manually compute the gradients and update the W and b.
with tf.Session() as sess:
sess.run(init)
for epoch in range(training_epochs):
for (x, y) in zip(train_X, train_Y):
W_del, b_del = sess.run(gr, feed_dict={X: x, Y: y})
W = max(0, (W - W_del)*learning_rate) #Project the gradient on [0, infinity]
b = max(0, (b - b_del)*learning_rate) # Project the gradient on [0, infinity]
This approach seems to be very slow.
I am wondering if there is a better way to run the second approach, or guarantee the results with the first approach. Can we somehow allow the optimizer to ensure that the learnt weights are non-negative?
Edit: How to do this in Autograd
https://github.com/HIPS/autograd/issues/207
If you modify your linear model to:
pred = tf.add(tf.multiply(X, tf.abs(W)), tf.abs(b))
it will have the same effect as using only positive W and b values.
The reason your second approach is slow is that you clip the W and b values outside of the tensorflow graph. (Also it will not converge because (W - W_del)*learning_rate must instead be W - W_del*learning_rate)
edit:
You can implement the clipping using tensorflow graph like this:
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
with tf.control_dependencies([train_step]):
clip_W = W.assign(tf.maximum(0., W))
clip_b = b.assign(tf.maximum(0., b))
train_step_with_clip = tf.group(clip_W, clip_b)
In this case W and b values will be clipped to 0 and not to small positive numbers.
Here is a small mnist example with clipping:
import tensorflow as tf
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
x = tf.placeholder(tf.uint8, [None, 28, 28])
x_vec = tf.cast(tf.reshape(x, [-1, 784]), tf.float32) / 255.
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.matmul(x_vec, W) + b
y_target = tf.placeholder(tf.uint8, [None])
y_target_one_hot = tf.one_hot(y_target, 10)
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_target_one_hot, logits=y))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
with tf.control_dependencies([train_step]):
clip_W = W.assign(tf.maximum(0., W))
clip_b = b.assign(tf.maximum(0., b))
train_step_with_clip = tf.group(clip_W, clip_b)
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_target_one_hot, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
with tf.Session() as sess:
tf.global_variables_initializer().run()
for i in range(1000):
sess.run(train_step_with_clip, feed_dict={
x: x_train[(i*100)%len(x_train):((i+1)*100)%len(x_train)],
y_target: y_train[(i*100)%len(x_train):((i+1)*100)%len(x_train)]})
if not i%100:
print("Min_W:", sess.run(tf.reduce_min(W)))
print("Min_b:", sess.run(tf.reduce_min(b)))
print("Accuracy:", sess.run(accuracy, feed_dict={
x: x_test,
y_target: y_test}))
I actually was not able to reproduce your problem of getting negative bs with your first approach.
But I do agree that this is not optimal for your use case and can result in negative values.
You should be able to constrain your parameters to non-negative values like so:
W *= tf.cast(W > 0., tf.float32)
b *= tf.cast(b > 0., tf.float32)
(exchange > with >= if necessary, the cast is necessary as the comparison operators will produce boolean values.
You then would optimize for the "standard cost" without the additional constraints.
However, this does not work in every case. For example, it should be avoided to initialize W or b with negative values in the beginning.
Your second (and probably better) approach can be accelerated by defining the update logic in the general computational graph, i.e. after the definition of cost
params = [W, b]
grads = tf.gradients(cost, params)
optimizer = [tf.assign(param, tf.maximum(0., param - grad*learning_rate))
for param, grad in zip(params, grads)]
I think your solution is slow because it creates new computation nodes every time which is probably very costly and repeated a lot inside the loops.
update using tensorflow optimizer
In my solution above not the gradients are clipped but rather the resulting update values.
Along the lines of this answer you could clip the gradients to be at most the value of the updated parameter like so:
params = [W, b]
opt = tf.train.GradientDescentOptimizer(learning_rate)
grads_and_vars = opt.compute_gradients(cost, params)
clipped_grads_vars = [(tf.clip_by_value(grad, -np.inf, var), var) for grad, var in grads_and_vars]
optimizer = opt.apply_gradients(clipped_grads_vars)
This way an update will never decrease a parameter to a value below 0.
However, I think this will not work in the case the updated variable is already negative.
Also, if the optimizing algorithm somehow multiplies the clipped gradient by a value greater than 1.
The latter might actually never happen, but I'm not 100% sure.
I am learning TensorFlow.
I have a question about the code in Introduction:
import tensorflow as tf
import numpy as np
# Create 100 phony x, y data points in NumPy, y = x * 0.1 + 0.3
x_data = np.random.rand(100).astype(np.float32)
y_data = x_data * 0.1 + 0.3
# Try to find values for W and b that compute y_data = W * x_data + b
# (We know that W should be 0.1 and b 0.3, but TensorFlow will
# figure that out for us.)
W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))
b = tf.Variable(tf.zeros([1]))
y = W * x_data + b
# Minimize the mean squared errors.
loss = tf.reduce_mean(tf.square(y - y_data))
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss)
# Before starting, initialize the variables. We will 'run' this first.
init = tf.global_variables_initializer()
# Launch the graph.
sess = tf.Session()
sess.run(init)
# Fit the line.
for step in range(201):
sess.run(train)
if step % 20 == 0:
print(step, sess.run(W), sess.run(b))
# Learns best fit is W: [0.1], b: [0.3]
This program learns best fit of W and b.
If I don't know the formula (y = W * x_data + b), how can I train a model?
For example, this is a training set:
{input = {{1,1}, {1,2}, {2,3}, ... }, target = {2, 3, 5, ...}}
How to train a function(a, b) ~= (a+b)?
In most cases, we do not know the exact form of the objective formula. Thus, we have to design a function and try to approximate the objective formula by this function.
In neural network, the formula is defined by the network architecture (for example, Multilayer perceptron or Recurrent Neural network) and hyper-parameters (for example, the number of hidden layer, the number of neuron in the hidden layers).
In this particular case for example, you can assume the approximate function has the form of (y = Wa+Ub+C -- a linear perceptron) and train the parameters of this function (W,U,C) to approximate the parameters of the objective formula (y=a+b) using the data given.
A neural network is a universal function approximator: that is, for any function (linear, polynomial, etc.), a neural network can approximate it given enough nodes in hidden layers and an activation function. A non-linear activation function (e.g. sigmoid, tanh, ReLU) will "bend" the linear boundary produced by Wx+b to be non-linear.