+ +

Polynomial Regression Using TensorFlow 2.x

+ +

I have a similar post titled Polynomial Regression Using Tensorflow that used tensorflow.compat.v1 (Which still works as of TF 2.16). But, I thought it would be nicer to redo it with newer TF versions.

+ +

I will be skipping all the introductions about polynomial regression and jumping straight to the code. Personally, I prefer using scikit-learn for this task.

+ +

Position vs Salary Dataset

+ +

Again, we will be using https://drive.google.com/file/d/1tNL4jxZEfpaP4oflfSn6pIHJX7Pachm9/view (Salary vs Position Dataset)

+ +

If you are in a Python Notebook environment like Kaggle or Google Colaboratory, you can simply run:

+ +

!wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=1tNL4jxZEfpaP4oflfSn6pIHJX7Pachm9' -O data.csv
+

+ +

Code

+ +

If you just want to copy-paste the code, scroll to the bottom for the entire snippet. Here I will try and walk through setting up code for a 3rd-degree (cubic) polynomial

+ +

Imports

+ +

import pandas as pd
+import tensorflow as tf
+import matplotlib.pyplot as plt
+import numpy as np
+

+ +

Reading the Dataset

+ +

df = pd.read_csv("data.csv")
+

+ +

Variables and Constants

+ +

Here, we initialize the X and Y values as constants, since they are not going to change. The coefficients are defined as variables.

+ +

X = tf.constant(df["Level"], dtype=tf.float32)
+Y = tf.constant(df["Salary"], dtype=tf.float32)
+
+coefficients = [tf.Variable(np.random.randn() * 0.01, dtype=tf.float32) for _ in range(4)]
+

+ +

Here, X and Y are the values from our dataset. We initialize the coefficients for the equations as small random values.

+ +

These coefficients are evaluated by Tensorflow's tf.math.poyval function which returns the n-th order polynomial based on how many coefficients are passed. Since our list of coefficients contains 4 different variables, it will be evaluated as:

+ +

y = (x**3)*coefficients[3] + (x**2)*coefficients[2] + (x**1)*coefficients[1] (x**0)*coefficients[0]
+

+ +

Which is equivalent to the general cubic equation:

+ +

+ +$$ +y = ax^3 + bx^2 + cx + d +$$ + +### Optimizer Selection & Training +

+ +

optimizer = tf.keras.optimizers.Adam(learning_rate=0.3)
+num_epochs = 10_000
+
+for epoch in range(num_epochs):
+    with tf.GradientTape() as tape:
+        y_pred = tf.math.polyval(coefficients, X)
+        loss = tf.reduce_mean(tf.square(y - y_pred))
+    grads = tape.gradient(loss, coefficients)
+    optimizer.apply_gradients(zip(grads, coefficients))
+    if (epoch+1) % 1000 == 0:
+        print(f"Epoch: {epoch+1}, Loss: {loss.numpy()}"
+

+ +

+ + +In TensorFlow 1, we would have been using `tf.Session` instead. + +Here we are using `GradientTape()` instead, to keep track of the loss evaluation and coefficients. This is crucial, as our optimizer needs these gradients to be able to optimize our coefficients. + +Our loss function is Mean Squared Error (MSE) + +$$ += \frac{1}{n}\sum_{i=1}^{n} (Y_i - \^{Y_i}) +$$ + +Where $\^{Y_i}$ is the predicted value and $Y_i$ is the actual value + +### Plotting Final Coefficients +

+ +

final_coefficients = [c.numpy() for c in coefficients]
+print("Final Coefficients:", final_coefficients)
+
+plt.plot(df["Level"], df["Salary"], label="Original Data")
+plt.plot(df["Level"],[tf.math.polyval(final_coefficients, tf.constant(x, dtype=tf.float32)).numpy() for x in df["Level"]])
+plt.ylabel('Salary')
+plt.xlabel('Position')
+plt.title("Salary vs Position")
+plt.show()
+

+ +

+ + + +## Code Snippet for a Polynomial of Degree N + +### Using Gradient Tape + +This should work regardless of the Keras backend version (2 or 3) +

+ +

import tensorflow as tf
+import numpy as np
+import pandas as pd
+import matplotlib.pyplot as plt
+
+df = pd.read_csv("data.csv")
+
+############################
+## Change Parameters Here ##
+############################
+x_column = "Level"         #
+y_column = "Salary"        #
+degree = 2                 #
+learning_rate = 0.3        #
+num_epochs = 25_000        #
+############################
+
+X = tf.constant(df[x_column], dtype=tf.float32)
+Y = tf.constant(df[y_column], dtype=tf.float32)
+
+coefficients = [tf.Variable(np.random.randn() * 0.01, dtype=tf.float32) for _ in range(degree + 1)]
+
+optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
+
+for epoch in range(num_epochs):
+    with tf.GradientTape() as tape:
+        y_pred = tf.math.polyval(coefficients, X)
+        loss = tf.reduce_mean(tf.square(Y - y_pred))
+    grads = tape.gradient(loss, coefficients)
+    optimizer.apply_gradients(zip(grads, coefficients))
+    if (epoch+1) % 1000 == 0:
+        print(f"Epoch: {epoch+1}, Loss: {loss.numpy()}")
+
+final_coefficients = [c.numpy() for c in coefficients]
+print("Final Coefficients:", final_coefficients)
+
+print("Final Equation:", end=" ")
+for i in range(degree+1):
+  print(f"{final_coefficients[i]} * x^{degree-i}", end=" + " if i < degree else "\n")
+
+plt.plot(X, Y, label="Original Data")
+plt.plot(X,[tf.math.polyval(final_coefficients, tf.constant(x, dtype=tf.float32)).numpy() for x in df[x_column]]), label="Our Poynomial"
+plt.ylabel(y_column)
+plt.xlabel(x_column)
+plt.title(f"{x_column} vs {y_column}")
+plt.legend()
+plt.show()
+

+ +

+ + +### Without Gradient Tape + +This relies on the Optimizer's `minimize` function and uses the `var_list` parameter to update the variables. + +This will not work with Keras 3 backend in TF 2.16.0 and above unless you switch to the legacy backend. +

+ +

import tensorflow as tf
+import numpy as np
+import pandas as pd
+import matplotlib.pyplot as plt
+
+df = pd.read_csv("data.csv")
+
+############################
+## Change Parameters Here ##
+############################
+x_column = "Level"         #
+y_column = "Salary"        #
+degree = 2                 #
+learning_rate = 0.3        #
+num_epochs = 25_000        #
+############################
+
+X = tf.constant(df[x_column], dtype=tf.float32)
+Y = tf.constant(df[y_column], dtype=tf.float32)
+
+coefficients = [tf.Variable(np.random.randn() * 0.01, dtype=tf.float32) for _ in range(degree + 1)]
+
+optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
+
+def loss_function():
+  pred_y = tf.math.polyval(coefficients, X)
+  return tf.reduce_mean(tf.square(pred_y - Y))
+
+for epoch in range(num_epochs):
+    optimizer.minimize(loss_function, var_list=coefficients)
+    if (epoch+1) % 1000 == 0:
+        current_loss = loss_function().numpy()
+        print(f"Epoch {epoch+1}: Training Loss: {current_loss}")
+
+final_coefficients = coefficients.numpy()
+print("Final Coefficients:", final_coefficients)
+
+print("Final Equation:", end=" ")
+for i in range(degree+1):
+  print(f"{final_coefficients[i]} * x^{degree-i}", end=" + " if i < degree else "\n")
+
+plt.plot(X, Y, label="Original Data")
+plt.plot(X,[tf.math.polyval(final_coefficients, tf.constant(x, dtype=tf.float32)).numpy() for x in df[x_column]], label="Our Polynomial")
+plt.ylabel(y_column)
+plt.xlabel(x_column)
+plt.legend()
+plt.title(f"{x_column} vs {y_column}")
+plt.show()
+

+ +

+ + + +As always, remember to tweak the parameters and choose the correct model for the job. A polynomial regression model might not even be the best model for this particular dataset. + +## Further Programming + +How would you modify this code to use another type of nonlinear regression? Say, $ y = ab^x $ + +Hint: Your loss calculation would be similar to: +

+ +

bx = tf.pow(coefficients[1], X)
+pred_y = tf.math.multiply(coefficients[0], bx)
+loss = tf.reduce_mean(tf.square(pred_y - Y))
+

+ +

If you have scrolled this far, consider subscribing to my mailing list here. You can subscribe to either a specific type of post you are interested in, or subscribe to everything with the "Everything" list.

+ +

Polynomial Regression Using TensorFlow 2.x

I will be skipping all the introductions about polynomial regression and jumping straight to the code. Personally, I prefer using scikit-learn for this task.

Position vs Salary Dataset

Again, we will be using https://drive.google.com/file/d/1tNL4jxZEfpaP4oflfSn6pIHJX7Pachm9/view (Salary vs Position Dataset)

@@ -61,11 +61,11 @@ -

Code

If you just want to copy-paste the code, scroll to the bottom for the entire snippet. Here I will try and walk through setting up code for a 3rd-degree (cubic) polynomial

Imports

import pandas as pd
@@ -75,14 +75,14 @@

Reading the Dataset

df = pd.read_csv("data.csv")

Variables and Constants

Here, we initialize the X and Y values as constants, since they are not going to change. The coefficients are defined as variables.

@@ -109,7 +109,7 @@

y = a x^{3} + b x^{2} + c x + d

Optimizer Selection & Training

optimizer = tf.keras.optimizers.Adam(learning_rate=0.3)
@@ -136,7 +136,7 @@
 
 Where  $\hat{Y_{i}}$  is the predicted value and  $Y_{i}$  is the actual value
 
-Plotting Final Coefficients
+Plotting Final Coefficients
 
 
 final_coefficients = [c.numpy() for c in coefficients]
@@ -151,9 +151,9 @@
 
 
 
-Code Snippet for a Polynomial of Degree N
+Code Snippet for a Polynomial of Degree N
 
-Using Gradient Tape
+Using Gradient Tape
 
 This should work regardless of the Keras backend version (2 or 3)
 
@@ -208,7 +208,7 @@

Without Gradient Tape

This relies on the Optimizer's minimize function and uses the var_list parameter to update the variables.

@@ -268,7 +268,7 @@

As always, remember to tweak the parameters and choose the correct model for the job. A polynomial regression model might not even be the best model for this particular dataset.

Further Programming

How would you modify this code to use another type of nonlinear regression? Say,

-- cgit v1.2.3 From 9e620084e57378952c1a7f8e0a772ebebd18932b Mon Sep 17 00:00:00 2001 From: Navan Chauhan Date: Wed, 27 Mar 2024 20:35:09 -0600 Subject: quick fix --- docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html') diff --git a/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html b/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html index 0d958b2..e2edc19 100644 --- a/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html +++ b/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html @@ -6,13 +6,13 @@ - id="polynomial-regression-using-tensorflow-2x">Polynomial Regression Using TensorFlow 2.x + Polynomial Regression Using TensorFlow 2.x - Polynomial Regression Using TensorFlow 2.x" /> - Polynomial Regression Using TensorFlow 2.x" /> + + -- cgit v1.2.3 From 01ff93c9c16867216f2d249664803860e1d6d5eb Mon Sep 17 00:00:00 2001 From: Navan Chauhan Date: Wed, 27 Mar 2024 22:49:40 -0600 Subject: generate new theme --- ...3-21-Polynomial-Regression-in-TensorFlow-2.html | 55 +++++++++++++++------- 1 file changed, 37 insertions(+), 18 deletions(-) (limited to 'docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html') diff --git a/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html b/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html index e2edc19..6dcd62b 100644 --- a/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html +++ b/docs/posts/2024-03-21-Polynomial-Regression-in-TensorFlow-2.html @@ -2,14 +2,26 @@ - + + + + + Polynomial Regression Using TensorFlow 2.x + + + + + + + - Polynomial Regression Using TensorFlow 2.x - @@ -29,21 +41,27 @@ - - - + + +

- +

Polynomial Regression Using TensorFlow 2.x

@@ -283,14 +301,15 @@

If you have scrolled this far, consider subscribing to my mailing list here. You can subscribe to either a specific type of post you are interested in, or subscribe to everything with the "Everything" list.