PyTorch Mannequin Efficiency Evaluation and Optimization — Half 3 | by Chaim Rand | Aug, 2023

That is the third a part of a collection of posts on the subject of analyzing and optimizing PyTorch fashions utilizing PyTorch Profiler and TensorBoard. Our intention has been to spotlight the advantages of efficiency profiling and optimization of GPU-based coaching workloads and their potential impression on the velocity and price of coaching. Particularly, we want to exhibit the accessibility of profiling instruments corresponding to PyTorch Profiler and TensorBoard to all ML builders. You don’t want to be a CUDA skilled to be able to derive significant efficiency positive aspects from making use of the strategies we focus on in our posts.

In our first post we demonstrated how the completely different views of the PyTorch Profiler TensorBoard plugin can be utilized to determine efficiency points and reviewed just a few common strategies for accelerating coaching. Within the second post we confirmed how the TensorBoard plugin Hint View can be utilized to determine when tensors are being copied from the CPU to the GPU, and again. Such motion of information — which may trigger factors of synchronization and gradual the velocity of coaching significantly — is usually unintentional and might generally be simply prevented. The subject of this put up shall be conditions wherein we encounter factors of synchronization between the GPU and CPU which are not related to tensor copies. As within the case of tensor copies, these may cause stagnation in your coaching step and gradual the general time of your coaching significantly. We are going to exhibit the existence of such occurrences, how they are often recognized utilizing PyTorch Profiler and the PyTorch Profiler TensorBoard plugin Hint View, and the potential efficiency advantages of constructing your mannequin in a method that minimizes such synchronization occasions.

As in our earlier posts, we are going to outline a toy PyTorch mannequin after which iteratively profile its efficiency, determine bottlenecks, and try to repair them. We are going to run our experiments on an Amazon EC2 g5.2xlarge occasion (containing an NVIDIA A10G GPU and eight vCPUs) and utilizing the official AWS PyTorch 2.0 Docker image. Take into account that a few of the behaviors we describe could fluctuate between variations of PyTorch.

Within the following blocks we introduce a toy PyTorch mannequin that performs semantic segmentation on a 256×256 enter picture, i.e., it takes a 256×256 RGB picture and outputs a 256×256 map of “per-pixel” labels from a category of ten semantic classes.

import torch
import torch.nn as nn
import torch.nn.purposeful as F
import torch.optim
import torch.profiler
import torch.utils.knowledge
from torch import Tensor
class Internet(nn.Module):
def __init__(self, num_hidden=10, num_classes=10):
tremendous().__init__()
self.conv_in = nn.Conv2d(3, 10, 3, padding='similar')
hidden = []
for i in vary(num_hidden):
hidden.append(nn.Conv2d(10, 10, 3, padding='similar'))
hidden.append(nn.ReLU())
self.hidden = nn.Sequential(*hidden)
self.conv_out = nn.Conv2d(10, num_classes, 3, padding='similar')
def ahead(self, x):
x = F.relu(self.conv_in(x))
x = self.hidden(x)
x = self.conv_out(x)
return x

To coach our mannequin we are going to use the usual cross-entropy loss with just a few modifications:

1. We are going to assume that the goal labels embrace an ignore worth indicating pixels that we wish to exclude from the loss calculation.
2. We are going to assume that certainly one of semantic labels identifies sure pixels as belonging to the “background” of the picture. We outline our loss operate to deal with these as ignore labels.
3. We are going to replace our mannequin weights solely after we encounter batches with targets tensors that embrace no less than two distinctive values.

Whereas we now have chosen these modifications for the needs of our demonstration, these kind of operations usually are not unusual and might be discovered in lots of “customary” PyTorch fashions. Since we’re already “specialists” at efficiency profiling, we now have already gone forward and wrapped every of the operations in our loss operate with a torch.profiler.record_function context supervisor, (as described in our second post).

class MaskedLoss(nn.Module):
def __init__(self, ignore_val=-1, num_classes=10):
tremendous().__init__()
self.ignore_val = ignore_val
self.num_classes = num_classes
self.loss = torch.nn.CrossEntropyLoss()
def cross_entropy(self, pred: Tensor, goal: Tensor) -> Tensor:
# create a boolean masks of legitimate labels
with torch.profiler.record_function('create masks'):
masks = goal != self.ignore_val
# permute the logits in preparation for masking
with torch.profiler.record_function('permute'):
permuted_pred = torch.permute(pred, [0, 2, 3, 1])
# apply the boolean masks to the targets and logits
with torch.profiler.record_function('masks'):
masked_target = goal[mask]
masked_pred = permuted_pred[mask.unsqueeze(-1).expand(-1, -1, -1,
self.num_classes)]
masked_pred = masked_pred.reshape(-1, self.num_classes)
# calculate the cross-entropy loss
with torch.profiler.record_function('calc loss'):
loss = self.loss(masked_pred, masked_target)
return loss
def ignore_background(self, goal: Tensor) -> Tensor:
# uncover all indices the place goal label is "background"
with torch.profiler.record_function('non_zero'):
inds = torch.nonzero(goal == self.num_classes - 1, as_tuple=True)
# reset all "background" labels to the ignore index
with torch.profiler.record_function('index project'):
goal[inds] = self.ignore_val
return goal
def ahead(self, pred: Tensor, goal: Tensor) -> Tensor:
# ignore background labels
goal = self.ignore_background(goal)
# retrieve a listing of distinctive parts in goal
with torch.profiler.record_function('distinctive'):
distinctive = torch.distinctive(goal)
# examine if the variety of distinctive objects move the edge
with torch.profiler.record_function('numel'):
ignore_loss = torch.numel(distinctive) < 2
# calculate the cross-entropy loss
loss = self.cross_entropy(pred, goal)
# zero the loss within the case that the variety of distinctive parts
# is beneath the edge
if ignore_loss:
loss = 0. * loss
return loss

Our loss operate appears harmless sufficient, proper? Fallacious! As we are going to see beneath, the loss operate consists of a lot of operations that set off host-device synchronization occasions that gradual the velocity of coaching significantly — none of which contain copying tensors into or out of the GPU. As in our earlier put up, we problem you to attempt to determine three alternatives for efficiency optimization earlier than studying on.

For the needs of our demo, we use randomly generated photos and per-pixel label maps, as outlined beneath.

from torch.utils.knowledge import Dataset
# A dataset with random photos and label maps
class FakeDataset(Dataset):
def __init__(self, num_classes=10):
tremendous().__init__()
self.num_classes = num_classes
self.img_size = [256, 256]
def __len__(self):
return 1000000
def __getitem__(self, index):
rand_image = torch.randn([3]+self.img_size, dtype=torch.float32)
rand_label = torch.randint(low=-1, excessive=self.num_classes, 
dimension=self.img_size)
return rand_image, rand_label
train_set = FakeDataset()
train_loader = torch.utils.knowledge.DataLoader(train_set, batch_size=256, 
shuffle=True, num_workers=8, pin_memory=True)

Final, we outline our coaching step with the PyTorch Profiler configured to our need:

gadget = torch.gadget("cuda:0")
mannequin = Internet().cuda(gadget)
criterion = MaskedLoss().cuda(gadget)
optimizer = torch.optim.SGD(mannequin.parameters(), lr=0.001, momentum=0.9)
mannequin.practice()
# coaching loop wrapped with profiler object
with torch.profiler.profile(
schedule=torch.profiler.schedule(wait=1, warmup=4, lively=3, repeat=1),
on_trace_ready=torch.profiler.tensorboard_trace_handler('/tmp/prof'),
record_shapes=True,
profile_memory=True,
with_stack=True
) as prof:
for step, knowledge in enumerate(train_loader):
inputs = knowledge[0].to(gadget=gadget, non_blocking=True)
labels = knowledge[1].to(gadget=gadget, non_blocking=True)
if step >= (1 + 4 + 3) * 1:
break
outputs = mannequin(inputs)
loss = criterion(outputs, labels)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
prof.step()

In case you had been to naively run this coaching script, you’d in all probability see excessive GPU (~90%) utilization and never know that there was something unsuitable with it. It is just by means of profiling that we’re capable of determine the underlying efficiency bottlenecks and potential alternatives for coaching acceleration. So, with out additional ado, let’s see how our mannequin performs.

On this put up we are going to deal with the Hint View of the PyTorch Profiler TensorBoard plugin. Please see our previous posts for recommendations on easy methods to use a few of the different views supported by the plugin.

Within the picture beneath we present the Hint View of a single coaching step of our toy mannequin.

We will clearly see that our 1.3 second lengthy coaching step is utterly dominated by the torch.nonzero operator within the first line of our loss operate. All the opposite operations seem bunched collectively on both aspect of the massive cudaMemcpyAsyn occasion. What’s going on??!! Why would such a seemingly harmless operation trigger such an enormous eyesore?

Maybe we shouldn’t be so stunned, because the torch.nonzero documentation does embrace the next word: “When enter is on CUDA, torch.nonzero() causes host-device synchronization.” The necessity for synchronization arises from the truth that, opposite to different widespread PyTorch ops, the dimensions of the tensor that’s returned by torch.nonzero is not pre-determined. The CPU doesn’t know what number of non-zero parts there are within the enter tensor forward of time. It wants to attend for the sync occasion from the GPU to be able to carry out the suitable GPU reminiscence allocation and appropriately put together the next PyTorch ops.

Notice that the size of cudaMempyAsync just isn’t indicative of the complexity of the torch.nonzero op, however slightly displays the period of time that the CPU wants to attend for the GPU to complete the entire earlier kernels that the CPU launched. For instance, had been we to make an extra torch.nonzero name instantly after our first one, our second cudaMempyAsync occasion would seem considerably shorter than the primary because the CPU and GPU are already kind of “in sync”. (Take into account that this rationalization is coming from a non-CUDA skilled, so make of it what you’ll…)

Now that we perceive the supply of the bottleneck, the problem turns into discovering an alternate sequence of operations that performs the identical logic however that does not set off a host-device synchronization occasion. Within the case of our loss operate, we are able to simply accomplish this utilizing the torch.where operator as proven within the code block beneath:

def ignore_background(self, goal: Tensor) -> Tensor:
with torch.profiler.record_function('replace background'):
goal = torch.the place(goal==self.num_classes-1, 
-1*torch.ones_like(goal),goal)
return goal

Within the picture beneath we present the Hint View following this variation.

Whereas we now have succeeded in eradicating the cudaMempyAsync coming from the torch.nonzero op, it has been instantly changed with one coming from the torch.unique op, and our step time has not budged. Right here the PyTorch documentation is much less variety, however primarily based on our earlier expertise we are able to assume that, as soon as once more, we’re affected by a host-device synchronization occasion resulting from our use of tensors with undetermined dimension.

Changing the torch.unique operator with an equal various just isn’t at all times attainable. Nevertheless, in our case we don’t truly have to know the values of the distinctive labels, we have to know solely the quantity of distinctive labels. This may be calculated by making use of the torch.sort op on the flattened goal tensor and counting the variety of steps within the resultant step operate.

    def ahead(self, pred: Tensor, goal: Tensor) -> Tensor:
# ignore background labels
goal = self.ignore_background(goal)
# kind the checklist of labels
with torch.profiler.record_function('kind'):
sorted,_ = torch.kind(goal.flatten())
# indentify the steps of the resultant step operate
with torch.profiler.record_function('deriv'):
deriv = sorted[1:]-sorted[:-1]
# depend the variety of steps
with torch.profiler.record_function('count_nonzero'):
num_unique = torch.count_nonzero(deriv)+1
# calculate the cross-entropy loss
loss = self.cross_entropy(pred, goal)
# zero the loss within the case that the variety of distinctive parts
# is beneath the edge
with torch.profiler.record_function('the place'):
loss = torch.the place(num_unique<2, 0.*loss, loss)
return loss

Within the picture beneath we seize the Hint View following our second optimization:

As soon as once more, we now have solved one bottleneck solely to be confronted with a brand new one, this time coming from the boolean masks routine.

Boolean masking is a routine we generally use to be able to cut back the general variety of machine operations which are required. In our case, our intention was to cut back the quantity of computation by eradicating the “ignore” pixels and limiting the cross-entropy calculation to the pixels of curiosity. Clearly, this has backfired. As earlier than, making use of a boolean masks leads to a tensor of undetermined dimension, and the cudaMempyAsync that it triggers vastly overshadows any of the financial savings from excluding the “ignore” pixels.

In our case, fixing this concern is slightly easy because the PyTorch CrossEntropyLoss has a built-in choice for setting an ignore_index.

class MaskedLoss(nn.Module):
def __init__(self, ignore_val=-1, num_classes=10):
tremendous().__init__()
self.ignore_val = ignore_val
self.num_classes = num_classes
self.loss = torch.nn.CrossEntropyLoss(ignore_index=-1)
def cross_entropy(self, pred: Tensor, goal: Tensor) -> Tensor:
with torch.profiler.record_function('calc loss'):
loss = self.loss(pred, goal)
return loss

Within the picture beneath we present the resultant Hint View:

Holy cow!! Our step time has dropped all the best way down to five.4 milliseconds. That’s 240 (!!) occasions sooner than what we began with. By merely altering round just a few operate calls and with none modification to the loss operate logic, we had been capable of optimize the efficiency of the coaching step dramatically.

Necessary Notice: Within the toy instance we now have chosen, the steps that we took to cut back the quantity cudaMempyAsync occasions had a transparent impression on the coaching step time. Nevertheless, there could also be conditions the place the identical sorts of adjustments will hurt efficiency slightly than enhance it. For instance, within the case of boolean masking, if our masks is extraordinarily sparse and the unique tensors extraordinarily massive, the financial savings in computation from making use of the masks would possibly outweigh the worth of the host-device synchronization. Importantly, the impression of every optimization needs to be evaluated on a case-by-case foundation.

On this put up we now have targeted on efficiency points in coaching purposes which are brought on by host-device synchronization occasions. We noticed a number of examples of PyTorch operators that set off such occasions — the widespread property of all of them being that the dimension of the tensors that they output are depending on the enter. You may also encounter synchronization occasions from different operators, not lined on this put up. We demonstrated how efficiency analyzers corresponding to PyTorch Profiler and its related TensorBoard plugin can be utilized to determine these sorts of occasions.

Within the case of our toy instance, we had been capable of finding equal alternate options to the problematic operators that use mounted sized tensors and keep away from the necessity for synchronization occasions. These led to a major enchancment in coaching time. Nevertheless, in apply you would possibly discover it a lot tougher — even unattainable — to unravel these sorts of bottlenecks. Typically, overcoming them would possibly require redesigning elements of your mannequin.