julio 7, 2021
~ 7 MIN
Pytorch Benchmarking
< Blog RSS
Pytorch - Benchmarking
En el post anterior aprendimos a utilizar el Profiler de Pytorch para poder evaluar nuestro código y encontrar posibles cuellos de botella y mejoras que nos ayuden a entrenar nuestras redes neuronales de manera rápida y eficiente. En este post vamos a ver otra herramienta útil que nos ofrece esta librerÃa para poder comparar diferentes implementaciones de una misma funcionalidad y poder saber asà cual es la mejor.
Para este ejemplo vamos a entrenar un calisficador de imágenes usando el dataset de imágenes de satélite EuroSAT.
Cargando datos
Existen muchas maneras de cargar nuestras imágenes para entrenar un modelo, y unas serán más eficientes que otras. Vamos a empezar definiendo un Dataset simple que cargará una imagen, la normalizará y la convertirá en tensor.
import os
from sklearn.model_selection import train_test_split
def setup(path='./data', test_size=0.2, random_state=42):
classes = sorted(os.listdir(path))
print("Generating images and labels ...")
images, encoded = [], []
for ix, label in enumerate(classes):
_images = os.listdir(f'{path}/{label}')
images += [f'{path}/{label}/{img}' for img in _images]
encoded += [ix]*len(_images)
print(f'Number of images: {len(images)}')
# train / val split
print("Generating train / val splits ...")
train_images, val_images, train_labels, val_labels = train_test_split(
images,
encoded,
stratify=encoded,
test_size=test_size,
random_state=random_state
)
print("Training samples: ", len(train_labels))
print("Validation samples: ", len(val_labels))
return classes, train_images, train_labels, val_images, val_labels
classes, train_images, train_labels, val_images, val_labels = setup('./data')
Generating images and labels ...
Number of images: 27000
Generating train / val splits ...
Training samples: 21600
Validation samples: 5400
import torch
import rasterio
import numpy as np
class Dataset(torch.utils.data.Dataset):
def __init__(self, images, labels):
self.images = images
self.labels = labels
def __len__(self):
return len(self.images)
def __getitem__(self, ix):
img = rasterio.open(self.images[ix]).read((4, 3, 2))
img = np.clip(img / 4000, 0, 1)
img = torch.tensor(img, dtype=torch.float)
label = torch.tensor(self.labels[ix], dtype=torch.long)
return img, label
ds = {
'train': Dataset(train_images, train_labels),
'val': Dataset(val_images, val_labels)
}
batch_size = 32
dl = {
'train': torch.utils.data.DataLoader(ds['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds['val'], batch_size=batch_size, shuffle=False)
}
imgs, labels = next(iter(dl['train']))
imgs.shape, labels.shape
(torch.Size([32, 3, 64, 64]), torch.Size([32]))
import matplotlib.pyplot as plt
fig = plt.figure(figsize=(10,8))
c, r = 6, 4
for j in range(r):
for i in range(c):
ix = j*c + i
ax = plt.subplot(r, c, ix + 1)
img, label = imgs[ix], labels[ix]
ax.imshow(img.permute(1,2,0))
ax.set_title(classes[label.item()])
ax.axis('off')
plt.tight_layout()
plt.show()
Python nos ofrece varias maneras para medir el tiempo que tarda una función en ser ejectuada, sin embargo la cosa puede complicarse cuando empezamos a trabajar con código en asÃncrono o GPUs. Vamos a empezar usando el módulo de Python timeit.
import timeit
def load_data(dl):
for batch_ix, (imgs, labels) in enumerate(dl):
shape = imgs.shape
if batch_ix + 1 >= 10:
break
t1 = timeit.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl['train']})
print(f'baseline: {t1.timeit(10) / 100 * 1e6:>5.1f} us')
baseline: 57950.7 us
Ahora, vamos a probar una implementación del Dataset diferente, usando la liberrÃa skimage en vez de rasterio para cargar las imágenes.
from skimage import io
class Dataset2(Dataset):
def __init__(self, images, labels):
super().__init__(images, labels)
def __getitem__(self, ix):
img = io.imread(self.images[ix])[...,(3,2,1)]
img = np.clip(img / 4000, 0, 1)
img = torch.tensor(img, dtype=torch.float).permute(2,0,1)
label = torch.tensor(self.labels[ix], dtype=torch.long)
return img, label
ds2 = {
'train': Dataset2(train_images, train_labels),
'val': Dataset2(val_images, val_labels)
}
dl2 = {
'train': torch.utils.data.DataLoader(ds2['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds2['val'], batch_size=batch_size, shuffle=False)
}
t2 = timeit.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl2['train']})
print(f'rasterio: {t1.timeit(10) / 100 * 1e6:>5.1f} us')
print(f'skimage: {t2.timeit(10) / 100 * 1e6:>5.1f} us')
rasterio: 60753.2 us
skimage: 15369.8 us
Como puedes ver, la segunda implementación es mucho más rápida 😬. Pytorch nos ofrece una solción de benchmarking muy similar en el paquete benchmark.
import torch.utils.benchmark as benchmark
t1 = benchmark.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl['train']})
t2 = benchmark.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl2['train']})
print(f'rasterio: {t1.timeit(10)}')
print(f'skimage: {t2.timeit(10)}')
rasterio: <torch.utils.benchmark.utils.common.Measurement object at 0x7faa64e13670>
load_data(dl)
setup: from __main__ import load_data
605.19 ms
1 measurement, 10 runs , 1 thread
skimage: <torch.utils.benchmark.utils.common.Measurement object at 0x7faa64c47250>
load_data(dl)
setup: from __main__ import load_data
151.76 ms
1 measurement, 10 runs , 1 thread
Vamos a probar otra implementación, ahora cambiando la manera en la que hacemos la normalización. En vez de usar numpy usaremos directamente Pytorch.
class Dataset3(Dataset):
def __init__(self, images, labels):
super().__init__(images, labels)
def __getitem__(self, ix):
img = io.imread(self.images[ix])[...,(3,2,1)]
#img = np.clip(img / 4000, 0, 1)
img = torch.tensor(img / 4000, dtype=torch.float).clip(0,1).permute(2,0,1)
label = torch.tensor(self.labels[ix], dtype=torch.long)
return img, label
ds3 = {
'train': Dataset3(train_images, train_labels),
'val': Dataset3(val_images, val_labels)
}
dl3 = {
'train': torch.utils.data.DataLoader(ds3['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds3['val'], batch_size=batch_size, shuffle=False)
}
t3 = benchmark.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl3['train']}
)
#print(f'rasterio: {t1.timeit(10)}')
print(f'skimage: {t2.timeit(10)}')
print(f'from_numpy: {t3.timeit(10)}')
skimage: <torch.utils.benchmark.utils.common.Measurement object at 0x7faa6465a4f0>
load_data(dl)
setup: from __main__ import load_data
149.59 ms
1 measurement, 10 runs , 1 thread
from_numpy: <torch.utils.benchmark.utils.common.Measurement object at 0x7faa6525b1c0>
load_data(dl)
setup: from __main__ import load_data
139.29 ms
1 measurement, 10 runs , 1 thread
No hemos ganado mucho, pero algo es algo ! Por último, podemos comparar diferentes implementaciones usando varios parámteros, como el batch size, de la siguiente manera.
results = []
batch_sizes = [1, 32, 64, 1024]
for batch_size in batch_sizes:
label = 'Load data'
sub_label = f'[{batch_size}]'
for num_threads in [1, 4]:
dl = {
'train': torch.utils.data.DataLoader(ds['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds['val'], batch_size=batch_size, shuffle=False)
}
results.append(benchmark.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl['train']},
num_threads=num_threads,
label=label,
sub_label=sub_label,
description='rasterio',
).blocked_autorange(min_run_time=1))
dl2 = {
'train': torch.utils.data.DataLoader(ds2['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds2['val'], batch_size=batch_size, shuffle=False)
}
results.append(benchmark.Timer(
stmt='load_data(dl)',
setup='from __main__ import load_data',
globals={'dl': dl2['train']},
num_threads=num_threads,
label=label,
sub_label=sub_label,
description='skimage',
).blocked_autorange(min_run_time=1))
compare = benchmark.Compare(results)
compare.print()
[------------ Load data ------------]
| rasterio | skimage
1 threads: --------------------------
[1] | 18.8 | 5.9
[32] | 608.1 | 149.0
[64] | 1219.1 | 298.0
[1024] | 18613.5 | 4952.0
4 threads: --------------------------
[1] | 18.8 | 5.9
[32] | 571.8 | 149.1
[64] | 1150.4 | 298.7
[1024] | 18793.4 | 4942.7
Times are in milliseconds (ms).
Entrenamiento
Y, por supuesto, puedes usar el Profiler tal y como vimos en el post anterior para evaluar todo el entrenamiento de un modelo con las diferentes implementaciones del Dataset y evaluar todas las operaciones en Tensorboard.
import torch.nn.functional as F
import torchvision
class Model(torch.nn.Module):
def __init__(self, n_outputs=10):
super().__init__()
self.model = torchvision.models.resnet18(pretrained=True)
self.model.fc = torch.nn.Linear(512, n_outputs)
def forward(self, x):
return self.model(x)
model = Model()
output = model(torch.randn(32, 3, 32, 32))
output.shape
torch.Size([32, 10])
from tqdm import tqdm
def step(model, batch, device):
x, y = batch
x, y = x.to(device), y.to(device)
y_hat = model(x)
loss = F.cross_entropy(y_hat, y)
acc = (torch.argmax(y_hat, axis=1) == y).sum().item() / y.size(0)
return loss, acc
def train(model, dl, optimizer, epochs=10, device="cpu", prof=None, end=0):
model.to(device)
hist = {'loss': [], 'acc': [], 'test_loss': [], 'test_acc': []}
for e in range(1, epochs+1):
# train
model.train()
l, a = [], []
bar = tqdm(dl['train'])
stop=False
for batch_idx, batch in enumerate(bar):
optimizer.zero_grad()
loss, acc = step(model, batch, device)
loss.backward()
optimizer.step()
l.append(loss.item())
a.append(acc)
bar.set_description(f"training... loss {np.mean(l):.4f} acc {np.mean(a):.4f}")
# profiling
if prof:
if batch_idx >= end:
stop = True
break
prof.step()
hist['loss'].append(np.mean(l))
hist['acc'].append(np.mean(a))
if stop:
break
# eval
model.eval()
l, a = [], []
bar = tqdm(dl['test'])
with torch.no_grad():
for batch in bar:
loss, acc = step(model, batch, device)
l.append(loss.item())
a.append(acc)
bar.set_description(f"testing... loss {np.mean(l):.4f} acc {np.mean(a):.4f}")
hist['test_loss'].append(np.mean(l))
hist['test_acc'].append(np.mean(a))
# log
log = f'Epoch {e}/{epochs}'
for k, v in hist.items():
log += f' {k} {v[-1]:.4f}'
print(log)
return hist
from torch.profiler import profile, record_function, ProfilerActivity
model = Model()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
batch_size = 2048
dl = {
'train': torch.utils.data.DataLoader(ds['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds['val'], batch_size=batch_size, shuffle=False)
}
with torch.profiler.profile(
schedule=torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=2),
on_trace_ready=torch.profiler.tensorboard_trace_handler('./log/resnet18_rasterio'),
record_shapes=True,
with_stack=True
) as prof:
hist = train(model, dl, optimizer, epochs=3, device="cuda", prof=prof, end=(1 + 1 + 3) * 2)
model = Model()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
batch_size = 2048
dl2 = {
'train': torch.utils.data.DataLoader(ds2['train'], batch_size=batch_size, shuffle=True),
'val': torch.utils.data.DataLoader(ds2['val'], batch_size=batch_size, shuffle=False)
}
with torch.profiler.profile(
schedule=torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=2),
on_trace_ready=torch.profiler.tensorboard_trace_handler('./log/resnet18_skimage'),
record_shapes=True,
with_stack=True
) as prof:
hist = train(model, dl2, optimizer, epochs=3, device="cuda", prof=prof, end=(1 + 1 + 3) * 2)
Resumen
En este post hemos aprendido a utilizar la utlidad de benchmarking que nos ofrece Pytorch. Con ella, podemos comparar diferentes implementaciones de una misma función para saber cual es mejor. Puedes usarlo para optimizar la carga de datos, el cálculo de métricas, diferentes librerÃas (como por ejemplo al usar redes convolucionales de timm vs torchivion, Pytorch vs Pytorch Lightning), etc.