Full gradient descent in keras

Any of those help At a local minimum (or maximum) x, the derivative of the target function f vanishes: f'(x) = 0 (assuming sufficient smoothness of f).
Gradient descent tries to find such a minimum x by using information from the first derivative of f: It simply follows the steepest descent from the current point. This is like rolling a ball down the graph of f until it comes to rest (while neglecting inertia).

should help you out I have a working implementation of multivariable linear regression using gradient descent in R. I'd like to see if I can use what I have to run a stochastic gradient descent. I'm not sure if this is really inefficient or not. For example, for each value of α I want to perform 500 SGD iterations and be able to specify the number of randomly picked samples in each iteration. It would be nice to do this so I could see how the number of samples influences the results. I'm having trouble through with the mini-batching and I want to be able to easily plot the results. , Sticking with what you have now

## all of this is the same

download.file("https://raw.githubusercontent.com/dbouquin/IS_605/master/sgd_ex_data/ex3x.dat", "ex3x.dat", method="curl")
x <- read.table('ex3x.dat')
x <- scale(x)
download.file("https://raw.githubusercontent.com/dbouquin/IS_605/master/sgd_ex_data/ex3y.dat", "ex3y.dat", method="curl")
y <- read.table('ex3y.dat')
data3 <- cbind(x,y)
colnames(data3) <- c("area_sqft", "bedrooms","price")
x1 <- rep(1, length(data3$area_sqft))
x <- as.matrix(cbind(x1,x))
y <- as.matrix(y)
L <- length(y)
cost <- function(x,y,theta){
  gradient <- (1/L)* (t(x) %*% ((x%*%t(theta)) - y))
  return(t(gradient)) 
}

GD <- function(x, y, alpha){
  theta <- matrix(c(0,0,0), nrow=1)
  theta_r <- NULL
  for (i in 1:500) {
    theta <- theta - alpha*cost(x,y,theta)  
    theta_r <- rbind(theta_r,theta)    
  }
  return(theta_r)
}

myGoD <- function(x, y, alpha, n = nrow(x)) {
  idx <- sample(nrow(x), n)
  y <- y[idx, , drop = FALSE]
  x <- x[idx, , drop = FALSE]
  GD(x, y, alpha)
}

all.equal(GD(x, y, 0.001), myGoD(x, y, 0.001))
# [1] TRUE

set.seed(1)
head(myGoD(x, y, 0.001, n = 20), 2)
#          x1        V1       V2
# V1 147.5978  82.54083 29.26000
# V1 295.1282 165.00924 58.48424

set.seed(1)
head(myGoD(x, y, 0.001, n = 40), 2)
#          x1        V1        V2
# V1 290.6041  95.30257  59.66994
# V1 580.9537 190.49142 119.23446

alphas <- c(0.001,0.01,0.1,1.0)
ns <- c(47, 40, 30, 20, 10)

par(mfrow = n2mfrow(length(alphas)))
for(i in 1:length(alphas)) {

  # result <- myGoD(x, y, alphas[i]) ## original
  result <- myGoD(x, y, alphas[i], ns[i])

  # red = price 
  # blue = sq ft 
  # green = bedrooms
  plot(result[,1],ylim=c(min(result),max(result)),col="#CC6666",ylab="Value",lwd=0.35,
       xlab=paste("alpha=", alphas[i]),xaxt="n") #suppress auto x-axis title
  lines(result[,2],type="b",col="#0072B2",lwd=0.35)
  lines(result[,3],type="b",col="#66CC99",lwd=0.35)
}

GD <- function(x, y, alpha, n = nrow(x)){
  idx <- sample(nrow(x), n)
  y <- y[idx, , drop = FALSE]
  x <- x[idx, , drop = FALSE]
  theta <- matrix(c(0,0,0), nrow=1)
  theta_r <- NULL

  for (i in 1:500) {
    theta <- theta - alpha*cost(x,y,theta)  
    theta_r <- rbind(theta_r,theta)    
  }
  return(theta_r)
}

wish helps you The problem isn't the Optimizer, it's your loss. It should return the mean loss, not the sum. If you're doing an L2 regression, for instance, it should look like this:

l_value = tf.pow(tf.abs(ground_truth - predict), 2) # distance for each individual position of the output matrix of shape = (n_examples, example_data_size)
regression_loss = tf.reduce_sum(l_value, axis=1) # distance per example, shape = (n_examples, 1)
total_regression_loss = tf.reduce_mean(regression_loss) # mean distance of all examples, shape = (1)

I wish this help you You are using a ReLU activation which basically cuts off the activations below 0, and using a default random_normal initialisation which has the parameters keras.initializers.RandomUniform(minval=-0.05, maxval=0.05, seed=None) by default. As you can see, the initialisation values are very close to 0 and half of them (-.05 to 0) don't get activated at all. And the ones that do get activated (0 to 0.05) propagate the gradients very very slowly.
My guess is to change the initialisation to be around 0 and n (which is the operating range for ReLUs) and your model should converge quickly.

I think the issue was by ths following , https://scikit-learn.org/stable/modules/sgd.html
if you want to use Gradient Descent approach, you should consider using SDRClassifier in SKlearn because SKlearn gives two Approaches to using Linear Regression. The first is LinearRegression class and is using Ordinary Least Squares solver from scipy the other one is SDRClassifier class which is an Implementation of the Gradient Descent Algorithm. So to answer your Question if you are using SDRClassifier in SKlearn then you are using an Implementation of Gradient Descent Algorithm behind the Scene.