From c96cfa909d2a1357a0dc1bba9fbb5e7704a79717 Mon Sep 17 00:00:00 2001
From: anurag2506 <anuraldo@gmail.com>
Date: Mon, 14 Oct 2024 01:06:38 +0530
Subject: [PATCH] finishing the sgd algorithm on a custom dataset

---
 ...stocastic_gradient_descent_from_scratch.py | 64 +++++++++++++++++++
 1 file changed, 64 insertions(+)

diff --git a/ML_algos_from_scratch/stocastic_gradient_descent_from_scratch/stocastic_gradient_descent_from_scratch.py b/ML_algos_from_scratch/stocastic_gradient_descent_from_scratch/stocastic_gradient_descent_from_scratch.py
index e69de29..21ea3a0 100644
--- a/ML_algos_from_scratch/stocastic_gradient_descent_from_scratch/stocastic_gradient_descent_from_scratch.py
+++ b/ML_algos_from_scratch/stocastic_gradient_descent_from_scratch/stocastic_gradient_descent_from_scratch.py
@@ -0,0 +1,64 @@
+'''
+This program performs SGD on a custom dataset that is produced using the random function. 
+It can be extrapolated to any type of dataset in a Regression model
+'''
+
+import tensorflow as tf
+import random
+import numpy as np
+
+
+#Create the random dataset to test the model on the regression task
+
+#This is a very abstract dataset and is just used for testing the model
+x_train = np.random.uniform(0, 10, size=(100, 1)).astype(np.float32)
+y_train=x_train+1
+#Define the model
+model=tf.keras.Sequential([
+    tf.keras.layers.Dense(64, activation='relu',input_shape=(1,)),
+    tf.keras.layers.Dense(32,activation='relu'),
+    tf.keras.layers.Dense(32,activation='relu'),
+    tf.keras.layers.Dense(16,activation='relu'),
+    tf.keras.layers.Dense(16,activation='relu'),
+    tf.keras.layers.Dense(1)
+    ])
+
+#Compile the model with the SGD opimizer and the MSE loss function
+model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=0.01), loss='mean_squared_error')
+
+#Fitting the model to the training data
+model.fit(x_train,y_train, epochs=100)
+
+# Evaluate the model
+loss = model.evaluate(x_train, y_train)
+print(f"Loss: {loss}")
+
+
+# Make predictions
+predictions = model.predict(x_train)
+print("Predictions:", predictions)
+
+
+#Explanation of what SGD does in comparison to traditional Gradient Descent:
+'''
+Gradient Descent is an algorithm that is used to minimize the loss function to increase
+the accuracy of the predictions by the model. This is done by modifiying the weights of the model
+SGD and GD differ in how they update these parameters during each iteration. 
+
+GD computes the gradient of the descent by evaluating the entire dataset at once and there is
+no recursion or iteration
+
+This is not at all optimal as we are not aware of the specifics of the dataset. 
+And also is computationally inefficient as it requires evaluating the entire dataset at every iteration
+
+But SGD does the same thing iteratively or in other words for each and every data point.
+Updates the params using only one training example at a time. SGD introduces noise into the training data
+which can help it to avoid shallow minimas and look for better, more optimal solutions. 
+
+But both of them have their own pros and cons.
+
+There is a third type of Gradient Descent called Mini-Batch Gradient Descent 
+which is a combination of both GD and SGD and takes a fixed 'k' data points at a time
+and calculates the loss function iteratively for (n/k) times(n data points in the entire dataset)
+
+'''
\ No newline at end of file