[ad_1]
Two days in the past, I launched torch
, an R package deal that gives the native performance that is delivered to Python customers by PyTorch. In that submit, I assumed fundamental familiarity with TensorFlow/Keras. Consequently, I portrayed torch
in a method I figured could be useful to somebody who “grew up” with the Keras method of coaching a mannequin: Aiming to deal with variations, but not lose sight of the general course of.
This submit now modifications perspective. We code a easy neural community “from scratch”, making use of simply one among torch
’s constructing blocks: tensors. This community will probably be as “uncooked” (low-level) as might be. (For the much less math-inclined individuals amongst us, it could function a refresher of what’s truly occurring beneath all these comfort instruments they constructed for us. However the actual function is as an instance what might be finished with tensors alone.)
Subsequently, three posts will progressively present tips on how to scale back the trouble – noticeably proper from the beginning, enormously as soon as we end. On the finish of this mini-series, you’ll have seen how computerized differentiation works in torch
, tips on how to use module
s (layers, in keras
converse, and compositions thereof), and optimizers. By then, you’ll have a variety of the background fascinating when making use of torch
to real-world duties.
This submit would be the longest, since there’s a lot to study tensors: Learn how to create them; tips on how to manipulate their contents and/or modify their shapes; tips on how to convert them to R arrays, matrices or vectors; and naturally, given the omnipresent want for pace: tips on how to get all these operations executed on the GPU. As soon as we’ve cleared that agenda, we code the aforementioned little community, seeing all these elements in motion.
Tensors
Creation
Tensors could also be created by specifying particular person values. Right here we create two one-dimensional tensors (vectors), of varieties float
and bool
, respectively:
torch_tensor
1
2
[ CPUFloatType{2} ]
torch_tensor
1
0
[ CPUBoolType{2} ]
And listed here are two methods to create two-dimensional tensors (matrices). Word how within the second method, you should specify byrow = TRUE
within the name to matrix()
to get values organized in row-major order.
torch_tensor
1 2 0
3 0 0
4 5 6
[ CPUFloatType{3,3} ]
torch_tensor
1 2 3
4 5 6
7 8 9
[ CPULongType{3,3} ]
In increased dimensions particularly, it may be simpler to specify the kind of tensor abstractly, as in: “give me a tensor of <…> of form n1 x n2”, the place <…> may very well be “zeros”; or “ones”; or, say, “values drawn from a normal regular distribution”:
# a 3x3 tensor of standard-normally distributed values
t <- torch_randn(3, 3)
t
# a 4x2x2 (3d) tensor of zeroes
t <- torch_zeros(4, 2, 2)
t
torch_tensor
-2.1563 1.7085 0.5245
0.8955 -0.6854 0.2418
0.4193 -0.7742 -1.0399
[ CPUFloatType{3,3} ]
torch_tensor
(1,.,.) =
0 0
0 0
(2,.,.) =
0 0
0 0
(3,.,.) =
0 0
0 0
(4,.,.) =
0 0
0 0
[ CPUFloatType{4,2,2} ]
Many related capabilities exist, together with, e.g., torch_arange()
to create a tensor holding a sequence of evenly spaced values, torch_eye()
which returns an identification matrix, and torch_logspace()
which fills a specified vary with an inventory of values spaced logarithmically.
If no dtype
argument is specified, torch
will infer the information kind from the passed-in worth(s). For instance:
t <- torch_tensor(c(3, 5, 7))
t$dtype
t <- torch_tensor(1L)
t$dtype
torch_Float
torch_Long
However we are able to explicitly request a unique dtype
if we would like:
t <- torch_tensor(2, dtype = torch_double())
t$dtype
torch_Double
torch
tensors reside on a machine. By default, this would be the CPU:
torch_device(kind='cpu')
However we may additionally outline a tensor to reside on the GPU:
t <- torch_tensor(2, machine = "cuda")
t$machine
torch_device(kind='cuda', index=0)
We’ll speak extra about gadgets under.
There may be one other essential parameter to the tensor-creation capabilities: requires_grad
. Right here although, I have to ask in your endurance: This one will prominently determine within the follow-up submit.
Conversion to built-in R information varieties
To transform torch
tensors to R, use as_array()
:
t <- torch_tensor(matrix(1:9, ncol = 3, byrow = TRUE))
as_array(t)
[,1] [,2] [,3]
[1,] 1 2 3
[2,] 4 5 6
[3,] 7 8 9
Relying on whether or not the tensor is one-, two-, or three-dimensional, the ensuing R object will probably be a vector, a matrix, or an array:
[1] "numeric"
[1] "matrix" "array"
[1] "array"
For one-dimensional and two-dimensional tensors, additionally it is doable to make use of as.integer()
/ as.matrix()
. (One cause you may wish to do that is to have extra self-documenting code.)
If a tensor at the moment lives on the GPU, you should transfer it to the CPU first:
t <- torch_tensor(2, machine = "cuda")
as.integer(t$cpu())
[1] 2
Indexing and slicing tensors
Typically, we wish to retrieve not a whole tensor, however solely a few of the values it holds, and even only a single worth. In these circumstances, we discuss slicing and indexing, respectively.
In R, these operations are 1-based, which means that once we specify offsets, we assume for the very first ingredient in an array to reside at offset 1
. The identical habits was applied for torch
. Thus, a variety of the performance described on this part ought to really feel intuitive.
The way in which I’m organizing this part is the next. We’ll examine the intuitive elements first, the place by intuitive I imply: intuitive to the R person who has not but labored with Python’s NumPy. Then come issues which, to this person, might look extra stunning, however will transform fairly helpful.
Indexing and slicing: the R-like half
None of those needs to be overly stunning:
torch_tensor
1 2 3
4 5 6
[ CPUFloatType{2,3} ]
torch_tensor
1
[ CPUFloatType{} ]
torch_tensor
1
2
3
[ CPUFloatType{3} ]
torch_tensor
1
2
[ CPUFloatType{2} ]
Word how, simply as in R, singleton dimensions are dropped:
[1] 2 3
[1] 2
integer(0)
And similar to in R, you’ll be able to specify drop = FALSE
to maintain these dimensions:
t[1, 1:2, drop = FALSE]$measurement()
t[1, 1, drop = FALSE]$measurement()
[1] 1 2
[1] 1 1
Indexing and slicing: What to look out for
Whereas R makes use of unfavorable numbers to take away parts at specified positions, in torch
unfavorable values point out that we begin counting from the top of a tensor – with -1
pointing to its final ingredient:
torch_tensor
3
[ CPUFloatType{} ]
torch_tensor
2 3
5 6
[ CPUFloatType{2,2} ]
It is a characteristic you may know from NumPy. Identical with the next.
When the slicing expression m:n
is augmented by one other colon and a 3rd quantity – m:n:o
–, we are going to take each o
th merchandise from the vary specified by m
and n
:
t <- torch_tensor(1:10)
t[2:10:2]
torch_tensor
2
4
6
8
10
[ CPULongType{5} ]
Typically we don’t know what number of dimensions a tensor has, however we do know what to do with the ultimate dimension, or the primary one. To subsume all others, we are able to use ..
:
t <- torch_randint(-7, 7, measurement = c(2, 2, 2))
t
t[.., 1]
t[2, ..]
torch_tensor
(1,.,.) =
2 -2
-5 4
(2,.,.) =
0 4
-3 -1
[ CPUFloatType{2,2,2} ]
torch_tensor
2 -5
0 -3
[ CPUFloatType{2,2} ]
torch_tensor
0 4
-3 -1
[ CPUFloatType{2,2} ]
Now we transfer on to a subject that, in observe, is simply as indispensable as slicing: altering tensor shapes.
Reshaping tensors
Modifications in form can happen in two essentially other ways. Seeing how “reshape” actually means: preserve the values however modify their structure, we may both alter how they’re organized bodily, or preserve the bodily construction as-is and simply change the “mapping” (a semantic change, because it have been).
Within the first case, storage must be allotted for 2 tensors, supply and goal, and parts will probably be copied from the latter to the previous. Within the second, bodily there will probably be only a single tensor, referenced by two logical entities with distinct metadata.
Not surprisingly, for efficiency causes, the second operation is most well-liked.
Zero-copy reshaping
We begin with zero-copy strategies, as we’ll wish to use them at any time when we are able to.
A particular case usually seen in observe is including or eradicating a singleton dimension.
unsqueeze()
provides a dimension of measurement 1
at a place specified by dim
:
t1 <- torch_randint(low = 3, excessive = 7, measurement = c(3, 3, 3))
t1$measurement()
t2 <- t1$unsqueeze(dim = 1)
t2$measurement()
t3 <- t1$unsqueeze(dim = 2)
t3$measurement()
[1] 3 3 3
[1] 1 3 3 3
[1] 3 1 3 3
Conversely, squeeze()
removes singleton dimensions:
t4 <- t3$squeeze()
t4$measurement()
[1] 3 3 3
The identical may very well be completed with view()
. view()
, nonetheless, is far more normal, in that it means that you can reshape the information to any legitimate dimensionality. (Legitimate which means: The variety of parts stays the identical.)
Right here we have now a 3x2
tensor that’s reshaped to measurement 2x3
:
torch_tensor
1 2
3 4
5 6
[ CPUFloatType{3,2} ]
torch_tensor
1 2 3
4 5 6
[ CPUFloatType{2,3} ]
(Word how that is totally different from matrix transposition.)
As an alternative of going from two to a few dimensions, we are able to flatten the matrix to a vector.
t4 <- t1$view(c(-1, 6))
t4$measurement()
t4
[1] 1 6
torch_tensor
1 2 3 4 5 6
[ CPUFloatType{1,6} ]
In distinction to indexing operations, this doesn’t drop dimensions.
Like we mentioned above, operations like squeeze()
or view()
don’t make copies. Or, put otherwise: The output tensor shares storage with the enter tensor. We will in actual fact confirm this ourselves:
t1$storage()$data_ptr()
t2$storage()$data_ptr()
[1] "0x5648d02ac800"
[1] "0x5648d02ac800"
What’s totally different is the storage metadata torch
retains about each tensors. Right here, the related data is the stride:
A tensor’s stride()
methodology tracks, for each dimension, what number of parts need to be traversed to reach at its subsequent ingredient (row or column, in two dimensions). For t1
above, of form 3x2
, we have now to skip over 2 gadgets to reach on the subsequent row. To reach on the subsequent column although, in each row we simply need to skip a single entry:
[1] 2 1
For t2
, of form 3x2
, the space between column parts is identical, however the distance between rows is now 3:
[1] 3 1
Whereas zero-copy operations are optimum, there are circumstances the place they gained’t work.
With view()
, this could occur when a tensor was obtained by way of an operation – aside from view()
itself – that itself has already modified the stride. One instance could be transpose()
:
torch_tensor
1 2
3 4
5 6
[ CPUFloatType{3,2} ]
[1] 2 1
torch_tensor
1 3 5
2 4 6
[ CPUFloatType{2,3} ]
[1] 1 2
In torch
lingo, tensors – like t2
– that re-use current storage (and simply learn it otherwise), are mentioned to not be “contiguous”. One method to reshape them is to make use of contiguous()
on them earlier than. We’ll see this within the subsequent subsection.
Reshape with copy
Within the following snippet, making an attempt to reshape t2
utilizing view()
fails, because it already carries data indicating that the underlying information shouldn’t be learn in bodily order.
Error in (operate (self, measurement) :
view measurement shouldn't be suitable with enter tensor's measurement and stride (no less than one dimension spans throughout two contiguous subspaces).
Use .reshape(...) as an alternative. (view at ../aten/src/ATen/native/TensorShape.cpp:1364)
Nonetheless, if we first name contiguous()
on it, a new tensor is created, which can then be (just about) reshaped utilizing view()
.
t3 <- t2$contiguous()
t3$view(6)
torch_tensor
1
3
5
2
4
6
[ CPUFloatType{6} ]
Alternatively, we are able to use reshape()
. reshape()
defaults to view()
-like habits if doable; in any other case it can create a bodily copy.
t2$storage()$data_ptr()
t4 <- t2$reshape(6)
t4$storage()$data_ptr()
[1] "0x5648d49b4f40"
[1] "0x5648d2752980"
Operations on tensors
Unsurprisingly, torch
supplies a bunch of mathematical operations on tensors; we’ll see a few of them within the community code under, and also you’ll encounter tons extra while you proceed your torch
journey. Right here, we shortly check out the general tensor methodology semantics.
Tensor strategies usually return references to new objects. Right here, we add to t1
a clone of itself:
torch_tensor
2 4
6 8
10 12
[ CPUFloatType{3,2} ]
On this course of, t1
has not been modified:
torch_tensor
1 2
3 4
5 6
[ CPUFloatType{3,2} ]
Many tensor strategies have variants for mutating operations. These all carry a trailing underscore:
t1$add_(t1)
# now t1 has been modified
t1
torch_tensor
4 8
12 16
20 24
[ CPUFloatType{3,2} ]
torch_tensor
4 8
12 16
20 24
[ CPUFloatType{3,2} ]
Alternatively, you’ll be able to after all assign the brand new object to a brand new reference variable:
torch_tensor
8 16
24 32
40 48
[ CPUFloatType{3,2} ]
There may be one factor we have to focus on earlier than we wrap up our introduction to tensors: How can we have now all these operations executed on the GPU?
Operating on GPU
To verify in case your GPU(s) is/are seen to torch, run
cuda_is_available()
cuda_device_count()
[1] TRUE
[1] 1
Tensors could also be requested to reside on the GPU proper at creation:
machine <- torch_device("cuda")
t <- torch_ones(c(2, 2), machine = machine)
Alternatively, they are often moved between gadgets at any time:
torch_device(kind='cuda', index=0)
torch_device(kind='cpu')
That’s it for our dialogue on tensors — nearly. There may be one torch
characteristic that, though associated to tensor operations, deserves particular point out. It’s referred to as broadcasting, and “bilingual” (R + Python) customers will comprehend it from NumPy.
Broadcasting
We regularly need to carry out operations on tensors with shapes that don’t match precisely.
Unsurprisingly, we are able to add a scalar to a tensor:
t1 <- torch_randn(c(3,5))
t1 + 22
torch_tensor
23.1097 21.4425 22.7732 22.2973 21.4128
22.6936 21.8829 21.1463 21.6781 21.0827
22.5672 21.2210 21.2344 23.1154 20.5004
[ CPUFloatType{3,5} ]
The identical will work if we add tensor of measurement 1
:
Including tensors of various sizes usually gained’t work:
Error in (operate (self, different, alpha) :
The scale of tensor a (2) should match the scale of tensor b (5) at non-singleton dimension 1 (infer_size at ../aten/src/ATen/ExpandUtils.cpp:24)
Nonetheless, underneath sure circumstances, one or each tensors could also be just about expanded so each tensors line up. This habits is what is supposed by broadcasting. The way in which it really works in torch
isn’t just impressed by, however truly similar to that of NumPy.
The principles are:
-
We align array shapes, ranging from the precise.
Say we have now two tensors, one among measurement
8x1x6x1
, the opposite of measurement7x1x5
.Right here they’re, right-aligned:
# t1, form: 8 1 6 1
# t2, form: 7 1 5
-
Beginning to look from the precise, the sizes alongside aligned axes both need to match precisely, or one among them must be equal to
1
: during which case the latter is broadcast to the bigger one.Within the above instance, that is the case for the second-from-last dimension. This now provides
# t1, form: 8 1 6 1
# t2, form: 7 6 5
, with broadcasting occurring in t2
.
-
If on the left, one of many arrays has an extra axis (or a couple of), the opposite is just about expanded to have a measurement of
1
in that place, during which case broadcasting will occur as said in (2).That is the case with
t1
’s leftmost dimension. First, there’s a digital growth
# t1, form: 8 1 6 1
# t2, form: 1 7 1 5
after which, broadcasting occurs:
# t1, form: 8 1 6 1
# t2, form: 8 7 1 5
In response to these guidelines, our above instance
may very well be modified in varied ways in which would permit for including two tensors.
For instance, if t2
have been 1x5
, it will solely have to get broadcast to measurement 3x5
earlier than the addition operation:
torch_tensor
-1.0505 1.5811 1.1956 -0.0445 0.5373
0.0779 2.4273 2.1518 -0.6136 2.6295
0.1386 -0.6107 -1.2527 -1.3256 -0.1009
[ CPUFloatType{3,5} ]
If it have been of measurement 5
, a digital main dimension could be added, after which, the identical broadcasting would happen as within the earlier case.
torch_tensor
-1.4123 2.1392 -0.9891 1.1636 -1.4960
0.8147 1.0368 -2.6144 0.6075 -2.0776
-2.3502 1.4165 0.4651 -0.8816 -1.0685
[ CPUFloatType{3,5} ]
Here’s a extra advanced instance. Broadcasting how occurs each in t1
and in t2
:
torch_tensor
1.2274 1.1880 0.8531 1.8511 -0.0627
0.2639 0.2246 -0.1103 0.8877 -1.0262
-1.5951 -1.6344 -1.9693 -0.9713 -2.8852
[ CPUFloatType{3,5} ]
As a pleasant concluding instance, by way of broadcasting an outer product might be computed like so:
torch_tensor
0 0 0
10 20 30
20 40 60
30 60 90
[ CPUFloatType{4,3} ]
And now, we actually get to implementing that neural community!
A easy neural community utilizing torch
tensors
Our process, which we method in a low-level method at present however significantly simplify in upcoming installments, consists of regressing a single goal datum primarily based on three enter variables.
We straight use torch
to simulate some information.
Toy information
library(torch)
# enter dimensionality (variety of enter options)
d_in <- 3
# output dimensionality (variety of predicted options)
d_out <- 1
# variety of observations in coaching set
n <- 100
# create random information
# enter
x <- torch_randn(n, d_in)
# goal
y <- x[, 1, drop = FALSE] * 0.2 -
x[, 2, drop = FALSE] * 1.3 -
x[, 3, drop = FALSE] * 0.5 +
torch_randn(n, 1)
Subsequent, we have to initialize the community’s weights. We’ll have one hidden layer, with 32
models. The output layer’s measurement, being decided by the duty, is the same as 1
.
Initialize weights
# dimensionality of hidden layer
d_hidden <- 32
# weights connecting enter to hidden layer
w1 <- torch_randn(d_in, d_hidden)
# weights connecting hidden to output layer
w2 <- torch_randn(d_hidden, d_out)
# hidden layer bias
b1 <- torch_zeros(1, d_hidden)
# output layer bias
b2 <- torch_zeros(1, d_out)
Now for the coaching loop correct. The coaching loop right here actually is the community.
Coaching loop
In every iteration (“epoch”), the coaching loop does 4 issues:
-
runs by way of the community, computing predictions (ahead move)
-
compares these predictions to the bottom reality and quantify the loss
-
runs backwards by way of the community, computing the gradients that point out how the weights needs to be modified
-
updates the weights, making use of the requested studying price.
Right here is the template we’re going to fill:
for (t in 1:200) {
### -------- Ahead move --------
# right here we'll compute the prediction
### -------- compute loss --------
# right here we'll compute the sum of squared errors
### -------- Backpropagation --------
# right here we'll move by way of the community, calculating the required gradients
### -------- Replace weights --------
# right here we'll replace the weights, subtracting portion of the gradients
}
The ahead move effectuates two affine transformations, one every for the hidden and output layers. In-between, ReLU activation is utilized:
# compute pre-activations of hidden layers (dim: 100 x 32)
# torch_mm does matrix multiplication
h <- x$mm(w1) + b1
# apply activation operate (dim: 100 x 32)
# torch_clamp cuts off values under/above given thresholds
h_relu <- h$clamp(min = 0)
# compute output (dim: 100 x 1)
y_pred <- h_relu$mm(w2) + b2
Our loss right here is imply squared error:
Calculating gradients the guide method is a bit tedious, however it may be finished:
# gradient of loss w.r.t. prediction (dim: 100 x 1)
grad_y_pred <- 2 * (y_pred - y)
# gradient of loss w.r.t. w2 (dim: 32 x 1)
grad_w2 <- h_relu$t()$mm(grad_y_pred)
# gradient of loss w.r.t. hidden activation (dim: 100 x 32)
grad_h_relu <- grad_y_pred$mm(w2$t())
# gradient of loss w.r.t. hidden pre-activation (dim: 100 x 32)
grad_h <- grad_h_relu$clone()
grad_h[h < 0] <- 0
# gradient of loss w.r.t. b2 (form: ())
grad_b2 <- grad_y_pred$sum()
# gradient of loss w.r.t. w1 (dim: 3 x 32)
grad_w1 <- x$t()$mm(grad_h)
# gradient of loss w.r.t. b1 (form: (32, ))
grad_b1 <- grad_h$sum(dim = 1)
The ultimate step then makes use of the calculated gradients to replace the weights:
learning_rate <- 1e-4
w2 <- w2 - learning_rate * grad_w2
b2 <- b2 - learning_rate * grad_b2
w1 <- w1 - learning_rate * grad_w1
b1 <- b1 - learning_rate * grad_b1
Let’s use these snippets to fill within the gaps within the above template, and provides it a attempt!
Placing all of it collectively
library(torch)
### generate coaching information -----------------------------------------------------
# enter dimensionality (variety of enter options)
d_in <- 3
# output dimensionality (variety of predicted options)
d_out <- 1
# variety of observations in coaching set
n <- 100
# create random information
x <- torch_randn(n, d_in)
y <-
x[, 1, NULL] * 0.2 - x[, 2, NULL] * 1.3 - x[, 3, NULL] * 0.5 + torch_randn(n, 1)
### initialize weights ---------------------------------------------------------
# dimensionality of hidden layer
d_hidden <- 32
# weights connecting enter to hidden layer
w1 <- torch_randn(d_in, d_hidden)
# weights connecting hidden to output layer
w2 <- torch_randn(d_hidden, d_out)
# hidden layer bias
b1 <- torch_zeros(1, d_hidden)
# output layer bias
b2 <- torch_zeros(1, d_out)
### community parameters ---------------------------------------------------------
learning_rate <- 1e-4
### coaching loop --------------------------------------------------------------
for (t in 1:200) {
### -------- Ahead move --------
# compute pre-activations of hidden layers (dim: 100 x 32)
h <- x$mm(w1) + b1
# apply activation operate (dim: 100 x 32)
h_relu <- h$clamp(min = 0)
# compute output (dim: 100 x 1)
y_pred <- h_relu$mm(w2) + b2
### -------- compute loss --------
loss <- as.numeric((y_pred - y)$pow(2)$sum())
if (t %% 10 == 0)
cat("Epoch: ", t, " Loss: ", loss, "n")
### -------- Backpropagation --------
# gradient of loss w.r.t. prediction (dim: 100 x 1)
grad_y_pred <- 2 * (y_pred - y)
# gradient of loss w.r.t. w2 (dim: 32 x 1)
grad_w2 <- h_relu$t()$mm(grad_y_pred)
# gradient of loss w.r.t. hidden activation (dim: 100 x 32)
grad_h_relu <- grad_y_pred$mm(
w2$t())
# gradient of loss w.r.t. hidden pre-activation (dim: 100 x 32)
grad_h <- grad_h_relu$clone()
grad_h[h < 0] <- 0
# gradient of loss w.r.t. b2 (form: ())
grad_b2 <- grad_y_pred$sum()
# gradient of loss w.r.t. w1 (dim: 3 x 32)
grad_w1 <- x$t()$mm(grad_h)
# gradient of loss w.r.t. b1 (form: (32, ))
grad_b1 <- grad_h$sum(dim = 1)
### -------- Replace weights --------
w2 <- w2 - learning_rate * grad_w2
b2 <- b2 - learning_rate * grad_b2
w1 <- w1 - learning_rate * grad_w1
b1 <- b1 - learning_rate * grad_b1
}
Epoch: 10 Loss: 352.3585
Epoch: 20 Loss: 219.3624
Epoch: 30 Loss: 155.2307
Epoch: 40 Loss: 124.5716
Epoch: 50 Loss: 109.2687
Epoch: 60 Loss: 100.1543
Epoch: 70 Loss: 94.77817
Epoch: 80 Loss: 91.57003
Epoch: 90 Loss: 89.37974
Epoch: 100 Loss: 87.64617
Epoch: 110 Loss: 86.3077
Epoch: 120 Loss: 85.25118
Epoch: 130 Loss: 84.37959
Epoch: 140 Loss: 83.44133
Epoch: 150 Loss: 82.60386
Epoch: 160 Loss: 81.85324
Epoch: 170 Loss: 81.23454
Epoch: 180 Loss: 80.68679
Epoch: 190 Loss: 80.16555
Epoch: 200 Loss: 79.67953
This appears to be like prefer it labored fairly properly! It additionally ought to have fulfilled its function: Exhibiting what you’ll be able to obtain utilizing torch
tensors alone. In case you didn’t really feel like going by way of the backprop logic with an excessive amount of enthusiasm, don’t fear: Within the subsequent installment, this can get considerably much less cumbersome. See you then!
[ad_2]