--- title: "Using autograd" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Using autograd} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", eval = identical(Sys.getenv("TORCH_TEST", unset = "0"), "1"), purl = FALSE ) ``` ```{r setup} library(torch) ``` So far, all we've been using from torch is _tensors_, but we've been performing all calculations ourselves -- the computing the predictions, the loss, the gradients (and thus, the necessary updates to the weights), and the new weight values. In this chapter, we'll make a significant change: Namely, we spare ourselves the cumbersome calculation of gradients, and have torch do it for us. Before we see that in action, let's get some more background. ## Automatic differentiation with autograd Torch uses a module called _autograd_ to record operations performed on tensors, and store what has to be done to obtain the respective gradients. These actions are stored as functions, and those functions are applied in order when the gradient of the output (normally, the loss) with respect to those tensors is calculated: starting from the output node and _propagating_ gradients _back_ through the network. This is a form of _reverse mode automatic differentiation_. As users, we can see a bit of this implementation. As a prerequisite for this "recording" to happen, tensors have to be created with `requires_grad = TRUE`. E.g. ```{r} x <- torch_ones(2,2, requires_grad = TRUE) ``` To be clear, this is a tensor _with respect to which_ gradients have to be calculated -- normally, a tensor representing a weight or a bias, not the input data ^[Unless we _want_ to change the data, as in adversarial example generation]. If we now perform some operation on that tensor, assigning the result to `y` ```{r} y <- x$mean() ``` we find that `y` now has a non-empty `grad_fn` that tells torch how to compute the gradient of `y` with respect to `x`: ```{r} y$grad_fn ``` Actual computation of gradients is triggered by calling `backward()` on the output tensor. ```{r} y$backward() ``` That executed, `x` now has a non-empty field `grad` that stores the gradient of `y` with respect to `x`: ```{r} x$grad ``` With a longer chain of computations, we can peek at how torch builds up a graph of backward operations. Here is a slightly more complex example. We call `retain_grad()` on `y` and `z` just for demonstration purposes; by default, intermediate gradients -- while of course they have to be computed -- aren't stored, in order to save memory. ```{r} x1 <- torch_ones(2,2, requires_grad = TRUE) x2 <- torch_tensor(1.1, requires_grad = TRUE) y <- x1 * (x2 + 2) y$retain_grad() z <- y$pow(2) * 3 z$retain_grad() out <- z$mean() ``` Starting from `out$grad_fn`, we can follow the graph all back to the leaf nodes: ```{r} # how to compute the gradient for mean, the last operation executed out$grad_fn # how to compute the gradient for the multiplication by 3 in z = y$pow(2) * 3 out$grad_fn$next_functions # how to compute the gradient for pow in z = y.pow(2) * 3 out$grad_fn$next_functions[[1]]$next_functions # how to compute the gradient for the multiplication in y = x * (x + 2) out$grad_fn$next_functions[[1]]$next_functions[[1]]$next_functions # how to compute the gradient for the two branches of y = x * (x + 2), # where the left branch is a leaf node (AccumulateGrad for x1) out$grad_fn$next_functions[[1]]$next_functions[[1]]$next_functions[[1]]$next_functions # here we arrive at the other leaf node (AccumulateGrad for x2) out$grad_fn$next_functions[[1]]$next_functions[[1]]$next_functions[[1]]$next_functions[[2]]$next_functions ``` After calling `out$backward()`, all tensors in the graph will have their respective gradients created. Without our calls to `retain_grad` above, `z$grad` and `y$grad` would be empty: ```{r} out$backward() z$grad y$grad x2$grad x1$grad ``` Thus acquainted with autograd, we're ready to modify our example. ## The simple network, now using autograd For a single new line calling `loss$backward()`, now a number of lines (that did manual backprop) are gone: ```{r} ### generate training data ----------------------------------------------------- # input dimensionality (number of input features) d_in <- 3 # output dimensionality (number of predicted features) d_out <- 1 # number of observations in training set n <- 100 # create random data x <- torch_randn(n, d_in) y <- x[,1]*0.2 - x[..,2]*1.3 - x[..,3]*0.5 + torch_randn(n) y <- y$unsqueeze(dim = 1) ### initialize weights --------------------------------------------------------- # dimensionality of hidden layer d_hidden <- 32 # weights connecting input to hidden layer w1 <- torch_randn(d_in, d_hidden, requires_grad = TRUE) # weights connecting hidden to output layer w2 <- torch_randn(d_hidden, d_out, requires_grad = TRUE) # hidden layer bias b1 <- torch_zeros(1, d_hidden, requires_grad = TRUE) # output layer bias b2 <- torch_zeros(1, d_out,requires_grad = TRUE) ### network parameters --------------------------------------------------------- learning_rate <- 1e-4 ### training loop -------------------------------------------------------------- for (t in 1:200) { ### -------- Forward pass -------- y_pred <- x$mm(w1)$add(b1)$clamp(min = 0)$mm(w2)$add(b2) ### -------- compute loss -------- loss <- (y_pred - y)$pow(2)$mean() if (t %% 10 == 0) cat(t, as_array(loss), "\n") ### -------- Backpropagation -------- # compute the gradient of loss with respect to all tensors with requires_grad = True. loss$backward() ### -------- Update weights -------- # Wrap in torch.no_grad() because this is a part we DON'T want to record for automatic gradient computation with_no_grad({ w1$sub_(learning_rate * w1$grad) w2$sub_(learning_rate * w2$grad) b1$sub_(learning_rate * b1$grad) b2$sub_(learning_rate * b2$grad) # Zero the gradients after every pass, because they'd accumulate otherwise w1$grad$zero_() w2$grad$zero_() b1$grad$zero_() b2$grad$zero_() }) } ``` We still manually compute the forward pass, and we still manually update the weights. In the last two chapters of this section, we'll see how these parts of the logic can be made more modular and reusable, as well.