Metric assortment is an important a part of each machine studying undertaking, enabling us to trace mannequin efficiency and monitor coaching progress. Ideally, Metrics needs to be collected and computed with out introducing any further overhead to the coaching course of. Nevertheless, similar to different parts of the coaching loop, inefficient metric computation can introduce pointless overhead, enhance training-step instances and inflate coaching prices.
This publish is the seventh in our sequence on efficiency profiling and optimization in PyTorch. The sequence has aimed to emphasise the crucial position of efficiency evaluation and Optimization in machine studying growth. Every publish has centered on completely different levels of the coaching pipeline, demonstrating sensible instruments and strategies for analyzing and boosting useful resource utilization and runtime effectivity.
On this installment, we deal with metric assortment. We’ll show how a naïve implementation of metric assortment can negatively influence runtime efficiency and discover instruments and strategies for its evaluation and optimization.
To implement our metric assortment, we are going to use TorchMetrics a preferred library designed to simplify and standardize metric computation in Pytorch. Our targets can be to:
- Display the runtime overhead attributable to a naïve implementation of metric assortment.
- Use PyTorch Profiler to pinpoint efficiency bottlenecks launched by metric computation.
- Display optimization strategies to cut back metric assortment overhead.
To facilitate our dialogue, we are going to outline a toy PyTorch mannequin and assess how metric assortment can influence its runtime efficiency. We’ll run our experiments on an NVIDIA A40 GPU, with a PyTorch 2.5.1 docker picture and TorchMetrics 1.6.1.
It’s necessary to notice that metric assortment habits can range vastly relying on the {hardware}, runtime setting, and mannequin structure. The code snippets offered on this publish are supposed for demonstrative functions solely. Please don’t interpret our point out of any device or approach as an endorsement for its use.
Toy Resnet Mannequin
Within the code block beneath we outline a easy picture classification mannequin with a ResNet-18 spine.
import time
import torch
import torchvision
system = "cuda"
mannequin = torchvision.fashions.resnet18().to(system)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mannequin.parameters())We outline an artificial dataset which we are going to use to coach our toy mannequin.
from torch.utils.knowledge import Dataset, DataLoader
# A dataset with random pictures and labels
class FakeDataset(Dataset):
def __len__(self):
return 100000000
def __getitem__(self, index):
rand_image = torch.randn([3, 224, 224], dtype=torch.float32)
label = torch.tensor(knowledge=index % 1000, dtype=torch.int64)
return rand_image, label
train_set = FakeDataset()
batch_size = 128
num_workers = 12
train_loader = DataLoader(
dataset=train_set,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=True
)We outline a set of normal metrics from TorchMetrics, together with a management flag to allow or disable metric calculation.
from torchmetrics import (
MeanMetric,
Accuracy,
Precision,
Recall,
F1Score,
)
# toggle to allow/disable metric assortment
capture_metrics = False
if capture_metrics:
metrics = {
"avg_loss": MeanMetric(),
"accuracy": Accuracy(activity="multiclass", num_classes=1000),
"precision": Precision(activity="multiclass", num_classes=1000),
"recall": Recall(activity="multiclass", num_classes=1000),
"f1_score": F1Score(activity="multiclass", num_classes=1000),
}
# Transfer all metrics to the system
metrics = {identify: metric.to(system) for identify, metric in metrics.objects()}Subsequent, we outline a PyTorch Profiler occasion, together with a management flag that permits us to allow or disable profiling. For an in depth tutorial on utilizing PyTorch Profiler, please check with the first publish on this sequence.
from torch import profiler
# toggle to allow/disable profiling
enable_profiler = True
if enable_profiler:
prof = profiler.profile(
schedule=profiler.schedule(wait=10, warmup=2, lively=3, repeat=1),
on_trace_ready=profiler.tensorboard_trace_handler("./logs/"),
profile_memory=True,
with_stack=True
)
prof.begin()Lastly, we outline a normal coaching step:
mannequin.practice()
t0 = time.perf_counter()
total_time = 0
depend = 0
for idx, (knowledge, goal) in enumerate(train_loader):
knowledge = knowledge.to(system, non_blocking=True)
goal = goal.to(system, non_blocking=True)
optimizer.zero_grad()
output = mannequin(knowledge)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
if capture_metrics:
# replace metrics
metrics["avg_loss"].replace(loss)
for identify, metric in metrics.objects():
if identify != "avg_loss":
metric.replace(output, goal)
if (idx + 1) % 100 == 0:
# compute metrics
metric_results = {
identify: metric.compute().merchandise()
for identify, metric in metrics.objects()
}
# print metrics
print(f"Step {idx + 1}: {metric_results}")
# reset metrics
for metric in metrics.values():
metric.reset()
elif (idx + 1) % 100 == 0:
# print final loss worth
print(f"Step {idx + 1}: Loss = {loss.merchandise():.4f}")
batch_time = time.perf_counter() - t0
t0 = time.perf_counter()
if idx > 10: # skip first steps
total_time += batch_time
depend += 1
if enable_profiler:
prof.step()
if idx > 200:
break
if enable_profiler:
prof.cease()
avg_time = total_time/depend
print(f'Common step time: {avg_time}')
print(f'Throughput: {batch_size/avg_time:.2f} pictures/sec')Metric Assortment Overhead
To measure the influence of metric assortment on coaching step time, we ran our coaching script each with and with out metric calculation. The outcomes are summarized within the following desk.
Our naïve metric assortment resulted in an almost 10% drop in runtime efficiency!! Whereas metric assortment is crucial for machine studying growth, it normally entails comparatively easy mathematical operations and hardly warrants such a major overhead. What’s going on?!!
Figuring out Efficiency Points with PyTorch Profiler
To raised perceive the supply of the efficiency degradation, we reran the coaching script with the PyTorch Profiler enabled. The resultant hint is proven beneath:
The hint reveals recurring “cudaStreamSynchronize” operations that coincide with noticeable drops in GPU utilization. Most of these “CPU-GPU sync” occasions had been mentioned intimately in half two of our sequence. In a typical coaching step, the CPU and GPU work in parallel: The CPU manages duties like knowledge transfers to the GPU and kernel loading, and the GPU executes the mannequin on the enter knowledge and updates its weights. Ideally, we want to reduce the factors of synchronization between the CPU and GPU with the intention to maximize efficiency. Right here, nonetheless, we will see that the metric assortment has triggered a sync occasion by performing a CPU to GPU knowledge copy. This requires the CPU to droop its processing till the GPU catches up which, in flip, causes the GPU to attend for the CPU to renew loading the next kernel operations. The underside line is that these synchronization factors result in inefficient utilization of each the CPU and GPU. Our metric assortment implmentation provides eight such synchronization occasions to every coaching step.
A better examination of the hint exhibits that the sync occasions are coming from the replace name of the MeanMetric TorchMetric. For the skilled profiling skilled, this can be ample to establish the basis trigger, however we are going to go a step additional and use the torch.profiler.record_function utility to establish the precise offending line of code.
Profiling with record_function
To pinpoint the precise supply of the sync occasion, we prolonged the MeanMetric class and overrode the replace methodology utilizing record_function context blocks. This strategy permits us to profile particular person operations throughout the methodology and establish efficiency bottlenecks.
class ProfileMeanMetric(MeanMetric):
def replace(self, worth, weight = 1.0):
# broadcast weight to worth form
with profiler.record_function("course of worth"):
if not isinstance(worth, torch.Tensor):
worth = torch.as_tensor(worth, dtype=self.dtype,
system=self.system)
with profiler.record_function("course of weight"):
if weight will not be None and never isinstance(weight, torch.Tensor):
weight = torch.as_tensor(weight, dtype=self.dtype,
system=self.system)
with profiler.record_function("broadcast weight"):
weight = torch.broadcast_to(weight, worth.form)
with profiler.record_function("cast_and_nan_check"):
worth, weight = self._cast_and_nan_check_input(worth, weight)
if worth.numel() == 0:
return
with profiler.record_function("replace worth"):
self.mean_value += (worth * weight).sum()
with profiler.record_function("replace weight"):
self.weight += weight.sum()We then up to date our avg_loss metric to make use of the newly created ProfileMeanMetric and reran the coaching script.
The up to date hint reveals that the sync occasion originates from the next line:
weight = torch.as_tensor(weight, dtype=self.dtype, system=self.system)This operation converts the default scalar worth weight=1.0 right into a PyTorch tensor and locations it on the GPU. The sync occasion happens as a result of this motion triggers a CPU-to-GPU knowledge copy, which requires the CPU to attend for the GPU to course of the copied worth.
Optimization 1: Specify Weight Worth
Now that we now have discovered the supply of the difficulty, we will overcome it simply by specifying a weight worth in our replace name. This prevents the runtime from changing the default scalar weight=1.0 right into a tensor on the GPU, avoiding the sync occasion:
# replace metrics
if capture_metric:
metrics["avg_loss"].replace(loss, weight=torch.ones_like(loss))Rerunning the script after making use of this variation reveals that we now have succeeded in eliminating the preliminary sync occasion… solely to have uncovered a brand new one, this time coming from the _cast_and_nan_check_input operate:
Profiling with record_function — Half 2
To discover our new sync occasion, we prolonged our customized metric with further profiling probes and reran our script.
class ProfileMeanMetric(MeanMetric):
def replace(self, worth, weight = 1.0):
# broadcast weight to worth form
with profiler.record_function("course of worth"):
if not isinstance(worth, torch.Tensor):
worth = torch.as_tensor(worth, dtype=self.dtype,
system=self.system)
with profiler.record_function("course of weight"):
if weight will not be None and never isinstance(weight, torch.Tensor):
weight = torch.as_tensor(weight, dtype=self.dtype,
system=self.system)
with profiler.record_function("broadcast weight"):
weight = torch.broadcast_to(weight, worth.form)
with profiler.record_function("cast_and_nan_check"):
worth, weight = self._cast_and_nan_check_input(worth, weight)
if worth.numel() == 0:
return
with profiler.record_function("replace worth"):
self.mean_value += (worth * weight).sum()
with profiler.record_function("replace weight"):
self.weight += weight.sum()
def _cast_and_nan_check_input(self, x, weight = None):
"""Convert enter ``x`` to a tensor and examine for Nans."""
with profiler.record_function("course of x"):
if not isinstance(x, torch.Tensor):
x = torch.as_tensor(x, dtype=self.dtype,
system=self.system)
with profiler.record_function("course of weight"):
if weight will not be None and never isinstance(weight, torch.Tensor):
weight = torch.as_tensor(weight, dtype=self.dtype,
system=self.system)
nans = torch.isnan(x)
if weight will not be None:
nans_weight = torch.isnan(weight)
else:
nans_weight = torch.zeros_like(nans).bool()
weight = torch.ones_like(x)
with profiler.record_function("any nans"):
anynans = nans.any() or nans_weight.any()
with profiler.record_function("course of nans"):
if anynans:
if self.nan_strategy == "error":
elevate RuntimeError("Encountered `nan` values in tensor")
if self.nan_strategy in ("ignore", "warn"):
if self.nan_strategy == "warn":
print("Encountered `nan` values in tensor."
" Will likely be eliminated.")
x = x[~(nans | nans_weight)]
weight = weight[~(nans | nans_weight)]
else:
if not isinstance(self.nan_strategy, float):
elevate ValueError(f"`nan_strategy` shall be float"
f" however you move {self.nan_strategy}")
x[nans | nans_weight] = self.nan_strategy
weight[nans | nans_weight] = self.nan_strategy
with profiler.record_function("return worth"):
retval = x.to(self.dtype), weight.to(self.dtype)
return retvalThe resultant hint is captured beneath:
The hint factors on to the offending line:
anynans = nans.any() or nans_weight.any()This operation checks for NaN values within the enter tensors, nevertheless it introduces a expensive CPU-GPU synchronization occasion as a result of the operation entails copying knowledge from the GPU to the CPU.
Upon a better inspection of the TorchMetric BaseAggregator class, we discover a number of choices for dealing with NAN worth updates, all of which move via the offending line of code. Nevertheless, for our use case — calculating the common loss metric — this examine is pointless and doesn’t justify the runtime efficiency penalty.
Optimization 2: Disable NAN Worth Checks
To remove the overhead, we suggest disabling the NaN worth checks by overriding the _cast_and_nan_check_input operate. As a substitute of a static override, we applied a dynamic answer that may be utilized flexibly to any descendants of the BaseAggregator class.
from torchmetrics.aggregation import BaseAggregator
def suppress_nan_check(MetricClass):
assert issubclass(MetricClass, BaseAggregator), MetricClass
class DisableNanCheck(MetricClass):
def _cast_and_nan_check_input(self, x, weight=None):
if not isinstance(x, torch.Tensor):
x = torch.as_tensor(x, dtype=self.dtype,
system=self.system)
if weight will not be None and never isinstance(weight, torch.Tensor):
weight = torch.as_tensor(weight, dtype=self.dtype,
system=self.system)
if weight is None:
weight = torch.ones_like(x)
return x.to(self.dtype), weight.to(self.dtype)
return DisableNanCheck
NoNanMeanMetric = suppress_nan_check(MeanMetric)
metrics["avg_loss"] = NoNanMeanMetric().to(system)Put up Optimization Outcomes: Success
After implementing the 2 optimizations — specifying the load worth and disabling the NaN checks—we discover the step time efficiency and the GPU utilization to match these of our baseline experiment. As well as, the resultant PyTorch Profiler hint exhibits that all the added “cudaStreamSynchronize” occasions that had been related to the metric assortment, have been eradicated. With a couple of small modifications, we now have diminished the price of coaching by ~10% with none modifications to the habits of the metric assortment.
Within the subsequent part we are going to discover a further Metric assortment optimization.
Instance 2: Optimizing Metric Machine Placement
Within the earlier part, the metric values resided on the GPU, making it logical to retailer and compute the metrics on the GPU. Nevertheless, in eventualities the place the values we want to combination reside on the CPU, it is perhaps preferable to retailer the metrics on the CPU to keep away from pointless system transfers.
Within the code block beneath, we modify our script to calculate the common step time utilizing a MeanMetric on the CPU. This alteration has no influence on the runtime efficiency of our coaching step:
avg_time = NoNanMeanMetric()
t0 = time.perf_counter()
for idx, (knowledge, goal) in enumerate(train_loader):
# transfer knowledge to system
knowledge = knowledge.to(system, non_blocking=True)
goal = goal.to(system, non_blocking=True)
optimizer.zero_grad()
output = mannequin(knowledge)
loss = criterion(output, goal)
loss.backward()
optimizer.step()
if capture_metrics:
metrics["avg_loss"].replace(loss)
for identify, metric in metrics.objects():
if identify != "avg_loss":
metric.replace(output, goal)
if (idx + 1) % 100 == 0:
# compute metrics
metric_results = {
identify: metric.compute().merchandise()
for identify, metric in metrics.objects()
}
# print metrics
print(f"Step {idx + 1}: {metric_results}")
# reset metrics
for metric in metrics.values():
metric.reset()
elif (idx + 1) % 100 == 0:
# print final loss worth
print(f"Step {idx + 1}: Loss = {loss.merchandise():.4f}")
batch_time = time.perf_counter() - t0
t0 = time.perf_counter()
if idx > 10: # skip first steps
avg_time.replace(batch_time)
if enable_profiler:
prof.step()
if idx > 200:
break
if enable_profiler:
prof.cease()
avg_time = avg_time.compute().merchandise()
print(f'Common step time: {avg_time}')
print(f'Throughput: {batch_size/avg_time:.2f} pictures/sec')The issue arises once we try to increase our script to help distributed coaching. To show the issue, we modified our mannequin definition to make use of DistributedDataParallel (DDP):
# toggle to allow/disable ddp
use_ddp = True
if use_ddp:
import os
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
os.environ["MASTER_ADDR"] = "127.0.0.1"
os.environ["MASTER_PORT"] = "29500"
dist.init_process_group("nccl", rank=0, world_size=1)
torch.cuda.set_device(0)
mannequin = DDP(torchvision.fashions.resnet18().to(system))
else:
mannequin = torchvision.fashions.resnet18().to(system)
# insert coaching loop
# append to finish of the script:
if use_ddp:
# destroy the method group
dist.destroy_process_group()The DDP modification leads to the next error:
RuntimeError: No backend kind related to system kind cpuBy default, metrics in distributed coaching are programmed to synchronize throughout all units in use. Nevertheless, the synchronization backend utilized by DDP doesn’t help metrics saved on the CPU.
One technique to remedy that is to disable the cross-device metric synchronization:
avg_time = NoNanMeanMetric(sync_on_compute=False)In our case, the place we’re measuring the common time, this answer is appropriate. Nevertheless, in some instances, the metric synchronization is crucial, and we now have could haven’t any alternative however to maneuver the metric onto the GPU:
avg_time = NoNanMeanMetric().to(system)Sadly, this case provides rise to a brand new CPU-GPU sync occasion coming from the replace operate.
This sync occasion ought to hardly come as a shock—in any case, we’re updating a GPU metric with a worth residing on the CPU, which ought to necessitate a reminiscence copy. Nevertheless, within the case of a scalar metric, this knowledge switch may be fully prevented with a easy optimization.
Optimization 3: Carry out Metric Updates with Tensors as an alternative of Scalars
The answer is easy: as an alternative of updating the metric with a float worth, we convert to a Tensor earlier than calling replace.
batch_time = torch.as_tensor(batch_time)
avg_time.replace(batch_time, torch.ones_like(batch_time))This minor change bypasses the problematic line of code, eliminates the sync occasion, and restores the step time to the baseline efficiency.
At first look, this outcome could seem shocking: We’d anticipate that updating a GPU metric with a CPU tensor ought to nonetheless require a reminiscence copy. Nevertheless, PyTorch optimizes operations on scalar tensors by utilizing a devoted kernel that performs the addition with out an express knowledge switch. This avoids the costly synchronization occasion that might in any other case happen.
Abstract
On this publish, we explored how a naïve strategy to TorchMetrics can introduce CPU-GPU synchronization occasions and considerably degrade PyTorch coaching efficiency. Utilizing PyTorch Profiler, we recognized the strains of code accountable for these sync occasions and utilized focused optimizations to remove them:
- Explicitly specify a weight tensor when calling the
MeanMetric.replaceoperate as an alternative of counting on the default worth. - Disable NaN checks within the base
Aggregatorclass or substitute them with a extra environment friendly different. - Fastidiously handle the system placement of every metric to attenuate pointless transfers.
- Disable cross-device metric synchronization when not required.
- When the metric resides on a GPU, convert floating-point scalars to tensors earlier than passing them to the
replaceoperate to keep away from implicit synchronization.
We now have created a devoted pull request on the TorchMetrics github web page overlaying a number of the optimizations mentioned on this publish. Please be at liberty to contribute your individual enhancements and optimizations!