Note
Go to the end to download the full example code.
Profiling your models¶
Do you feel like your model is too slow? Do you want to make it faster? Instead of guessing which part of the code is responsible for the slowdown, you should profile your code to learn how much time is spent in each function and where to focus any optimization efforts.
In this tutorial you’ll learn how to profile your model using the PyTorch profiler, how to read the output of the profiler, and how to add your own labels for new functions/steps in your model’s forward function.
from typing import Dict, List, Optional
import ase.build
import matplotlib.pyplot as plt
import numpy as np
import torch
from metatensor.torch import Labels, TensorBlock, TensorMap
from metatomic.torch import (
AtomisticModel,
ModelCapabilities,
ModelMetadata,
ModelOutput,
System,
)
from metatomic.torch.ase_calculator import MetatomicCalculator
When profiling your code, it is important to run the model on a representative system to ensure you are actually exercising the behavior of your model at the right scale. Here we’ll use a relatively large system with many atoms.
primitive = ase.build.bulk(name="C", crystalstructure="diamond", a=3.567)
atoms = ase.build.make_supercell(primitive, 10 * np.eye(3))
print(f"We have {len(atoms)} atoms in our system")
We have 2000 atoms in our system
We will use the same HarmonicModel
as in the previous tutorial as our machine learning potential.
Click to see the definition of HarmonicModel
class HarmonicModel(torch.nn.Module):
def __init__(self, force_constant: float, equilibrium_positions: torch.Tensor):
"""Create an ``HarmonicModel``.
:param force_constant: force constant, in ``energy unit / (length unit)^2``
:param equilibrium_positions: torch tensor with shape ``n x 3``, containing the
equilibrium positions of all atoms
"""
super().__init__()
assert force_constant > 0
self.force_constant = force_constant
self.equilibrium_positions = equilibrium_positions
def forward(
self,
systems: List[System],
outputs: Dict[str, ModelOutput],
selected_atoms: Optional[Labels],
) -> Dict[str, TensorMap]:
if list(outputs.keys()) != ["energy"]:
raise ValueError(
"this model can only compute 'energy', but `outputs` contains other "
f"keys: {', '.join(outputs.keys())}"
)
# we don't want to worry about selected_atoms yet
if selected_atoms is not None:
raise NotImplementedError("selected_atoms is not implemented")
if outputs["energy"].per_atom:
raise NotImplementedError("per atom energy is not implemented")
# compute the energy for each system by adding together the energy for each atom
energy = torch.zeros((len(systems), 1), dtype=systems[0].positions.dtype)
for i, system in enumerate(systems):
assert len(system) == self.equilibrium_positions.shape[0]
r0 = self.equilibrium_positions
energy[i] += torch.sum(self.force_constant * (system.positions - r0) ** 2)
# add metadata to the output
block = TensorBlock(
values=energy,
samples=Labels("system", torch.arange(len(systems)).reshape(-1, 1)),
components=[],
properties=Labels("energy", torch.tensor([[0]])),
)
return {
"energy": TensorMap(keys=Labels("_", torch.tensor([[0]])), blocks=[block])
}
model = HarmonicModel(
force_constant=3.14159265358979323846,
equilibrium_positions=torch.tensor(atoms.positions),
)
capabilities = ModelCapabilities(
outputs={
"energy": ModelOutput(quantity="energy", unit="eV", per_atom=False),
},
atomic_types=[6],
interaction_range=0.0,
length_unit="Angstrom",
supported_devices=["cpu"],
dtype="float32",
)
metadata = ModelMetadata()
wrapper = AtomisticModel(model.eval(), metadata, capabilities)
wrapper.export("exported-model.pt")
/home/runner/work/metatomic/metatomic/python/examples/4-profiling.py:128: DeprecationWarning: `export()` is deprecated, use `save()` instead
wrapper.export("exported-model.pt")
If you are trying to profile your own model, you can start here and create
MetatomicCalculator
with your own model.
Profiling energy calculation¶
We will start with an energy-only calculator, which can be enabled with
do_gradients_with_energy=False
.
atoms.calc = MetatomicCalculator("exported-model.pt", do_gradients_with_energy=False)
Before trying to profile the code, it is a good idea to run it a couple of times to allow torch to warmup internally.
for _ in range(10):
# force the model to re-run everytime, otherwise ASE caches calculation results
atoms.rattle(1e-6)
atoms.get_potential_energy()
Now we can run code using torch.profiler.profile()
to collect statistic on
how long each function takes to run.
atoms.rattle(1e-6)
with torch.profiler.profile() as energy_profiler:
atoms.get_potential_energy()
print(energy_profiler.key_averages().table(sort_by="self_cpu_time_total", row_limit=10))
------------------------------------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
------------------------------------------ ------------ ------------ ------------ ------------ ------------ ------------
MetatomicCalculator::prepare_inputs 21.33% 338.053us 31.11% 493.015us 493.015us 1
forward 11.57% 183.263us 20.00% 316.985us 316.985us 1
Model::forward 9.22% 146.134us 29.23% 463.119us 463.119us 1
AtomisticModel::check_atomic_types 6.68% 105.847us 11.37% 180.128us 180.128us 1
AtomisticModel::convert_units_output 5.25% 83.116us 5.37% 85.140us 85.140us 1
MetatomicCalculator::sum_energies 4.93% 78.056us 4.93% 78.056us 78.056us 1
AtomisticModel::convert_units_input 4.36% 69.139us 4.52% 71.664us 71.664us 1
MetatomicCalculator::compute_neighbors 3.97% 62.949us 7.49% 118.623us 118.623us 1
MetatomicCalculator::convert_outputs 3.92% 62.067us 5.07% 80.341us 80.341us 1
aten::copy_ 3.67% 58.179us 3.67% 58.179us 5.289us 11
------------------------------------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 1.585ms
There are a couple of interesting things to see here. First the total runtime of the
code is shown in the bottom; and then the most costly functions are visible on top,
one line per function. For each function, Self CPU
refers to the time spent in
this function excluding any called functions; and CPU total
refers to the time
spent in this function, including called functions.
For more options to record operations and display outputs, please refer to the official documentation for PyTorch profiler.
Here, Model::forward
indicates the time taken by your model’s forward()
.
Anything starting with aten::
comes from operations on torch tensors, typically
with the same function name as the corresponding torch functions (e.g.
aten::arange
is torch.arange()
). We can also see some internal functions
from metatomic, with the name staring with AtomisticModel::
for
AtomisticModel
; and MetatomicCalculator::
for
ase_calculator.MetatomicCalculator
.
If you want to see more details on the internal steps taken by your model, you
can add torch.profiler.record_function()
(https://pytorch.org/docs/stable/generated/torch.autograd.profiler.record_function.html)
inside your model code to give names to different steps in the calculation.
This is how we internally set names such as Model::forward
or
MetatomicCalculator::prepare_inputs
above.
Profiling forces calculation¶
Let us now do the same, but while also computing the forces for this system.
This mean we should now see some time spent in the backward()
function, on
top of everything else.
atoms.calc = MetatomicCalculator("exported-model.pt")
# warmup
for _ in range(10):
atoms.rattle(1e-6)
atoms.get_forces()
atoms.rattle(1e-6)
with torch.profiler.profile() as forces_profiler:
atoms.get_forces()
print(forces_profiler.key_averages().table(sort_by="self_cpu_time_total", row_limit=10))
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
MetatomicCalculator::prepare_inputs 12.91% 205.407us 19.52% 310.703us 310.703us 1
MetatomicCalculator::convert_outputs 9.36% 148.999us 12.46% 198.343us 198.343us 1
MetatomicCalculator::run_backward 8.79% 139.893us 22.92% 364.734us 364.734us 1
Model::forward 6.78% 107.923us 18.36% 292.199us 292.199us 1
forward 5.70% 90.789us 11.58% 184.276us 184.276us 1
aten::mm 5.06% 80.553us 5.11% 81.364us 20.341us 4
AtomisticModel::convert_units_output 4.77% 75.943us 4.88% 77.636us 77.636us 1
AtomisticModel::check_atomic_types 4.56% 72.636us 7.91% 125.936us 125.936us 1
MetatomicCalculator::sum_energies 4.37% 69.611us 4.37% 69.611us 69.611us 1
AtomisticModel::convert_units_input 3.71% 59.021us 3.84% 61.085us 61.085us 1
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 1.591ms
Let’s visualize this data in another way:
events = forces_profiler.key_averages()
events = sorted(events, key=lambda u: u.self_cpu_time_total, reverse=True)
total_cpu_time = sum(map(lambda u: u.self_cpu_time_total, events))
bottom = 0.0
for event in events:
self_time = event.self_cpu_time_total
name = event.key
if len(name) > 30:
name = name[:12] + "[...]" + name[-12:]
if self_time > 0.03 * total_cpu_time:
plt.bar(0, self_time, bottom=bottom, label=name)
bottom += self_time
else:
plt.bar(0, total_cpu_time - bottom, bottom=bottom, label="others")
break
plt.legend()
plt.xticks([])
plt.xlim(0, 1)
plt.ylabel("self time / µs")
plt.show()

Total running time of the script: (0 minutes 1.359 seconds)