Quantization-Sparsity Aware Training with NNCF, using PyTorch framework#
This Jupyter notebook can be launched after a local installation only.
This notebook is based on ImageNet training in PyTorch.
The goal of this notebook is to demonstrate how to use the Neural Network Compression Framework NNCF 8-bit quantization to optimize a PyTorch model for inference with OpenVINO Toolkit. The optimization process contains the following steps:
Transforming the original dense
FP32model to sparseINT8Using fine-tuning to improve the accuracy.
Exporting optimized and original models to OpenVINO IR
Measuring and comparing the performance of models.
For more advanced usage, refer to these examples.
This tutorial uses the ResNet-50 model with the ImageNet dataset. The dataset must be downloaded separately. To see ResNet models, visit PyTorch hub.
Table of contents:
Installation Instructions#
This is a self-contained example that relies solely on its own code.
We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.
%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu "openvino>=2024.0.0" "torch" "torchvision" "tqdm"
%pip install -q "nncf>=2.9.0"
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.
Imports and Settings#
On Windows, add the required C++ directories to the system PATH.
Import NNCF and all auxiliary packages from your Python code. Set a name for the model, and the image width and height that will be used for the network. Also define paths where PyTorch and OpenVINO IR versions of the models will be stored.
NOTE: All NNCF logging messages below ERROR level (INFO and WARNING) are disabled to simplify the tutorial. For production use, it is recommended to enable logging by removing
set_log_level(logging.ERROR).
import time
import warnings # To disable warnings on export model
from pathlib import Path
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.datasets as datasets
import torchvision.transforms as transforms
import torchvision.models as models
import openvino as ov
from torch.jit import TracerWarning
torch.manual_seed(0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using {device} device")
MODEL_DIR = Path("model")
OUTPUT_DIR = Path("output")
# DATA_DIR = Path("...") # Insert path to folder containing imagenet folder
# DATASET_DIR = DATA_DIR / "imagenet"
Using cpu device
# Fetch `notebook_utils` module
import zipfile
import requests
r = requests.get(
url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py",
)
open("notebook_utils.py", "w").write(r.text)
from notebook_utils import download_file, device_widget
DATA_DIR = Path("data")
def download_tiny_imagenet_200(
data_dir: Path,
url="http://cs231n.stanford.edu/tiny-imagenet-200.zip",
tarname="tiny-imagenet-200.zip",
):
archive_path = data_dir / tarname
download_file(url, directory=data_dir, filename=tarname)
zip_ref = zipfile.ZipFile(archive_path, "r")
zip_ref.extractall(path=data_dir)
zip_ref.close()
def prepare_tiny_imagenet_200(dataset_dir: Path):
# Format validation set the same way as train set is formatted.
val_data_dir = dataset_dir / "val"
val_annotations_file = val_data_dir / "val_annotations.txt"
with open(val_annotations_file, "r") as f:
val_annotation_data = map(lambda line: line.split("\t")[:2], f.readlines())
val_images_dir = val_data_dir / "images"
for image_filename, image_label in val_annotation_data:
from_image_filepath = val_images_dir / image_filename
to_image_dir = val_data_dir / image_label
if not to_image_dir.exists():
to_image_dir.mkdir()
to_image_filepath = to_image_dir / image_filename
from_image_filepath.rename(to_image_filepath)
val_annotations_file.unlink()
val_images_dir.rmdir()
DATASET_DIR = DATA_DIR / "tiny-imagenet-200"
if not DATASET_DIR.exists():
download_tiny_imagenet_200(DATA_DIR)
prepare_tiny_imagenet_200(DATASET_DIR)
print(f"Successfully downloaded and prepared dataset at: {DATASET_DIR}")
BASE_MODEL_NAME = "resnet18"
image_size = 64
OUTPUT_DIR.mkdir(exist_ok=True)
MODEL_DIR.mkdir(exist_ok=True)
DATA_DIR.mkdir(exist_ok=True)
# Paths where PyTorch and OpenVINO IR models will be stored.
fp32_pth_path = Path(MODEL_DIR / (BASE_MODEL_NAME + "_fp32")).with_suffix(".pth")
fp32_ir_path = fp32_pth_path.with_suffix(".xml")
int8_sparse_ir_path = Path(MODEL_DIR / (BASE_MODEL_NAME + "_int8_sparse")).with_suffix(".xml")
tiny-imagenet-200.zip: 0%| | 0.00/237M [00:00<?, ?B/s]
Successfully downloaded and prepared dataset at: data/tiny-imagenet-200
Train Function#
def train(train_loader, model, compression_ctrl, criterion, optimizer, epoch):
batch_time = AverageMeter("Time", ":3.3f")
losses = AverageMeter("Loss", ":2.3f")
top1 = AverageMeter("Acc@1", ":2.2f")
top5 = AverageMeter("Acc@5", ":2.2f")
progress = ProgressMeter(
len(train_loader),
[batch_time, losses, top1, top5],
prefix="Epoch:[{}]".format(epoch),
)
# Switch to train mode.
model.train()
end = time.time()
for i, (images, target) in enumerate(train_loader):
images = images.to(device)
target = target.to(device)
# Compute output.
output = model(images)
loss = criterion(output, target)
# Measure accuracy and record loss.
acc1, acc5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), images.size(0))
top1.update(acc1[0], images.size(0))
top5.update(acc5[0], images.size(0))
# Compute gradient and do opt step.
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Measure elapsed time.
batch_time.update(time.time() - end)
end = time.time()
print_frequency = 50
if i % print_frequency == 0:
progress.display(i)
compression_ctrl.scheduler.step()
Validate Function#
def validate(val_loader, model, criterion):
batch_time = AverageMeter("Time", ":3.3f")
losses = AverageMeter("Loss", ":2.3f")
top1 = AverageMeter("Acc@1", ":2.2f")
top5 = AverageMeter("Acc@5", ":2.2f")
progress = ProgressMeter(len(val_loader), [batch_time, losses, top1, top5], prefix="Test: ")
# Switch to evaluate mode.
model.eval()
with torch.no_grad():
end = time.time()
for i, (images, target) in enumerate(val_loader):
images = images.to(device)
target = target.to(device)
# Compute output.
output = model(images)
loss = criterion(output, target)
# Measure accuracy and record loss.
acc1, acc5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), images.size(0))
top1.update(acc1[0], images.size(0))
top5.update(acc5[0], images.size(0))
# Measure elapsed time.
batch_time.update(time.time() - end)
end = time.time()
print_frequency = 10
if i % print_frequency == 0:
progress.display(i)
print(" * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}".format(top1=top1, top5=top5))
return top1.avg
Helpers#
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self, name, fmt=":f"):
self.name = name
self.fmt = fmt
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def __str__(self):
fmtstr = "{name} {val" + self.fmt + "} ({avg" + self.fmt + "})"
return fmtstr.format(**self.__dict__)
class ProgressMeter(object):
def __init__(self, num_batches, meters, prefix=""):
self.batch_fmtstr = self._get_batch_fmtstr(num_batches)
self.meters = meters
self.prefix = prefix
def display(self, batch):
entries = [self.prefix + self.batch_fmtstr.format(batch)]
entries += [str(meter) for meter in self.meters]
print("\t".join(entries))
def _get_batch_fmtstr(self, num_batches):
num_digits = len(str(num_batches // 1))
fmt = "{:" + str(num_digits) + "d}"
return "[" + fmt + "/" + fmt.format(num_batches) + "]"
def accuracy(output, target, topk=(1,)):
"""Computes the accuracy over the k top predictions for the specified values of k"""
with torch.no_grad():
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
res.append(correct_k.mul_(100.0 / batch_size))
return res
Get a Pre-trained FP32 Model#
А pre-trained floating-point model is a prerequisite for quantization. It can be obtained by tuning from scratch with the code below.
num_classes = 1000
init_lr = 1e-4
batch_size = 128
epochs = 20
# model = models.resnet50(pretrained=True)
model = models.resnet18(pretrained=True)
model.fc = nn.Linear(in_features=512, out_features=200, bias=True)
model.to(device)
# Data loading code.
train_dir = DATASET_DIR / "train"
val_dir = DATASET_DIR / "val"
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
train_dir,
transforms.Compose(
[
transforms.Resize([image_size, image_size]),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]
),
)
val_dataset = datasets.ImageFolder(
val_dir,
transforms.Compose(
[
transforms.Resize([256, 256]),
transforms.CenterCrop([image_size, image_size]),
transforms.ToTensor(),
normalize,
]
),
)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1,
pin_memory=True,
sampler=None,
)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=1, pin_memory=True)
# Define loss function (criterion) and optimizer.
criterion = nn.CrossEntropyLoss().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=init_lr)
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead. warnings.warn( /opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or None for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing weights=ResNet18_Weights.IMAGENET1K_V1. You can also use weights=ResNet18_Weights.DEFAULT to get the most up-to-date weights. warnings.warn(msg)
Export the FP32 model to OpenVINO™ Intermediate Representation, to
benchmark it in comparison with the INT8 model.
dummy_input = torch.randn(1, 3, image_size, image_size).to(device)
ov_model = ov.convert_model(model, example_input=dummy_input, input=[1, 3, image_size, image_size])
ov.save_model(ov_model, fp32_ir_path, compress_to_fp16=False)
print(f"FP32 model was exported to {fp32_ir_path}.")
FP32 model was exported to model/resnet18_fp32.xml.
Create and Initialize Quantization and Sparsity Training#
NNCF enables compression-aware training by integrating into regular training pipelines. The framework is designed so that modifications to your original training code are minor.
from nncf import NNCFConfig
from nncf.torch import create_compressed_model, register_default_init_args
# load
nncf_config = NNCFConfig.from_json("config.json")
nncf_config = register_default_init_args(nncf_config, train_loader)
# Creating a compressed model
compression_ctrl, compressed_model = create_compressed_model(model, nncf_config)
compression_ctrl.scheduler.epoch_step()
INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino
INFO:nncf:Ignored adding weight sparsifier for operation: ResNet/NNCFConv2d[conv1]/conv2d_0
INFO:nncf:Collecting tensor statistics |█ | 8 / 79
INFO:nncf:Collecting tensor statistics |███ | 16 / 79
INFO:nncf:Collecting tensor statistics |████ | 24 / 79
INFO:nncf:Collecting tensor statistics |██████ | 32 / 79
INFO:nncf:Collecting tensor statistics |████████ | 40 / 79
INFO:nncf:Collecting tensor statistics |█████████ | 48 / 79
INFO:nncf:Collecting tensor statistics |███████████ | 56 / 79
INFO:nncf:Collecting tensor statistics |████████████ | 64 / 79
INFO:nncf:Collecting tensor statistics |██████████████ | 72 / 79
INFO:nncf:Collecting tensor statistics |████████████████| 79 / 79
INFO:nncf:Compiling and loading torch extension: quantized_functions_cpu...
INFO:nncf:Finished loading torch extension: quantized_functions_cpu
2024-12-10 04:05:55.241385: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0. 2024-12-10 04:05:55.265169: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
INFO:nncf:BatchNorm statistics adaptation |█ | 1 / 16
INFO:nncf:BatchNorm statistics adaptation |██ | 2 / 16
INFO:nncf:BatchNorm statistics adaptation |███ | 3 / 16
INFO:nncf:BatchNorm statistics adaptation |████ | 4 / 16
INFO:nncf:BatchNorm statistics adaptation |█████ | 5 / 16
INFO:nncf:BatchNorm statistics adaptation |██████ | 6 / 16
INFO:nncf:BatchNorm statistics adaptation |███████ | 7 / 16
INFO:nncf:BatchNorm statistics adaptation |████████ | 8 / 16
INFO:nncf:BatchNorm statistics adaptation |█████████ | 9 / 16
INFO:nncf:BatchNorm statistics adaptation |██████████ | 10 / 16
INFO:nncf:BatchNorm statistics adaptation |███████████ | 11 / 16
INFO:nncf:BatchNorm statistics adaptation |████████████ | 12 / 16
INFO:nncf:BatchNorm statistics adaptation |█████████████ | 13 / 16
INFO:nncf:BatchNorm statistics adaptation |██████████████ | 14 / 16
INFO:nncf:BatchNorm statistics adaptation |███████████████ | 15 / 16
INFO:nncf:BatchNorm statistics adaptation |████████████████| 16 / 16
Validate Compressed Model
Evaluate the new model on the validation set after initialization of quantization and sparsity.
acc1 = validate(val_loader, compressed_model, criterion)
print(f"Accuracy of initialized sparse INT8 model: {acc1:.3f}")
Test: [ 0/79] Time 0.393 (0.393) Loss 6.069 (6.069) Acc@1 0.00 (0.00) Acc@5 4.69 (4.69)
Test: [10/79] Time 0.154 (0.175) Loss 5.368 (5.689) Acc@1 0.78 (0.07) Acc@5 3.91 (2.41)
Test: [20/79] Time 0.149 (0.164) Loss 5.921 (5.653) Acc@1 0.00 (0.56) Acc@5 2.34 (3.16)
Test: [30/79] Time 0.152 (0.161) Loss 5.664 (5.670) Acc@1 0.00 (0.50) Acc@5 0.78 (2.90)
Test: [40/79] Time 0.164 (0.161) Loss 5.608 (5.632) Acc@1 1.56 (0.59) Acc@5 3.12 (3.09)
Test: [50/79] Time 0.130 (0.158) Loss 5.170 (5.618) Acc@1 0.00 (0.72) Acc@5 2.34 (3.32)
Test: [60/79] Time 0.131 (0.155) Loss 6.619 (5.634) Acc@1 0.00 (0.67) Acc@5 0.00 (3.00)
Test: [70/79] Time 0.152 (0.155) Loss 5.771 (5.653) Acc@1 0.00 (0.57) Acc@5 1.56 (2.77)
* Acc@1 0.570 Acc@5 2.770
Accuracy of initialized sparse INT8 model: 0.570
Fine-tune the Compressed Model#
At this step, a regular fine-tuning process is applied to further improve quantized model accuracy. Normally, several epochs of tuning are required with a small learning rate, the same that is usually used at the end of the training of the original model. No other changes in the training pipeline are required. Here is a simple example.
compression_lr = init_lr / 10
optimizer = torch.optim.Adam(compressed_model.parameters(), lr=compression_lr)
nr_epochs = 10
# Train for one epoch with NNCF.
print("Training")
for epoch in range(nr_epochs):
compression_ctrl.scheduler.epoch_step()
train(train_loader, compressed_model, compression_ctrl, criterion, optimizer, epoch=epoch)
# Evaluate on validation set after Quantization-Aware Training (QAT case).
print("Validating")
acc1_int8_sparse = validate(val_loader, compressed_model, criterion)
print(f"Accuracy of tuned INT8 sparse model: {acc1_int8_sparse:.3f}")
print(f"Accuracy drop of tuned INT8 sparse model over pre-trained FP32 model: {acc1 - acc1_int8_sparse:.3f}")
Training
Epoch:[0][ 0/782] Time 0.556 (0.556) Loss 5.673 (5.673) Acc@1 0.78 (0.78) Acc@5 3.12 (3.12)
Epoch:[0][ 50/782] Time 0.338 (0.348) Loss 5.634 (5.644) Acc@1 0.00 (0.74) Acc@5 3.12 (3.12)
Epoch:[0][100/782] Time 0.343 (0.346) Loss 5.572 (5.606) Acc@1 0.78 (0.77) Acc@5 2.34 (3.29)
Epoch:[0][150/782] Time 0.338 (0.355) Loss 5.531 (5.559) Acc@1 1.56 (0.90) Acc@5 3.12 (3.52)
Epoch:[0][200/782] Time 0.337 (0.352) Loss 5.276 (5.515) Acc@1 1.56 (1.07) Acc@5 7.03 (3.96)
Epoch:[0][250/782] Time 0.340 (0.350) Loss 5.361 (5.473) Acc@1 0.00 (1.21) Acc@5 6.25 (4.52)
Epoch:[0][300/782] Time 0.348 (0.348) Loss 5.242 (5.431) Acc@1 1.56 (1.41) Acc@5 7.81 (5.04)
Epoch:[0][350/782] Time 0.338 (0.352) Loss 5.092 (5.389) Acc@1 3.12 (1.65) Acc@5 11.72 (5.78)
Epoch:[0][400/782] Time 0.336 (0.351) Loss 5.052 (5.351) Acc@1 0.78 (1.85) Acc@5 12.50 (6.39)
Epoch:[0][450/782] Time 0.343 (0.350) Loss 5.033 (5.312) Acc@1 3.12 (2.14) Acc@5 12.50 (7.11)
Epoch:[0][500/782] Time 0.348 (0.352) Loss 4.859 (5.275) Acc@1 5.47 (2.41) Acc@5 13.28 (7.85)
Epoch:[0][550/782] Time 0.338 (0.351) Loss 4.697 (5.237) Acc@1 10.94 (2.75) Acc@5 23.44 (8.72)
Epoch:[0][600/782] Time 0.342 (0.351) Loss 4.616 (5.197) Acc@1 8.59 (3.15) Acc@5 25.78 (9.74)
Epoch:[0][650/782] Time 0.458 (0.351) Loss 4.610 (5.160) Acc@1 9.38 (3.55) Acc@5 23.44 (10.65)
Epoch:[0][700/782] Time 0.334 (0.352) Loss 4.633 (5.122) Acc@1 7.81 (4.00) Acc@5 23.44 (11.65)
Epoch:[0][750/782] Time 0.342 (0.351) Loss 4.462 (5.083) Acc@1 17.19 (4.53) Acc@5 34.38 (12.69)
Epoch:[1][ 0/782] Time 0.685 (0.685) Loss 4.324 (4.324) Acc@1 17.19 (17.19) Acc@5 33.59 (33.59)
Epoch:[1][ 50/782] Time 0.345 (0.379) Loss 4.327 (4.226) Acc@1 12.50 (16.76) Acc@5 33.59 (37.47)
Epoch:[1][100/782] Time 0.346 (0.361) Loss 4.195 (4.187) Acc@1 17.19 (17.98) Acc@5 37.50 (38.35)
Epoch:[1][150/782] Time 0.345 (0.355) Loss 4.044 (4.166) Acc@1 21.09 (18.47) Acc@5 39.84 (38.98)
Epoch:[1][200/782] Time 0.340 (0.352) Loss 4.096 (4.142) Acc@1 18.75 (18.82) Acc@5 39.06 (39.77)
Epoch:[1][250/782] Time 0.346 (0.356) Loss 4.091 (4.119) Acc@1 19.53 (19.14) Acc@5 42.19 (40.31)
Epoch:[1][300/782] Time 0.341 (0.354) Loss 4.201 (4.098) Acc@1 14.84 (19.40) Acc@5 34.38 (40.81)
Epoch:[1][350/782] Time 0.341 (0.352) Loss 3.818 (4.076) Acc@1 26.56 (19.74) Acc@5 45.31 (41.34)
Epoch:[1][400/782] Time 0.363 (0.355) Loss 4.093 (4.053) Acc@1 18.75 (20.20) Acc@5 36.72 (41.94)
Epoch:[1][450/782] Time 0.342 (0.354) Loss 3.788 (4.033) Acc@1 25.78 (20.55) Acc@5 44.53 (42.41)
Epoch:[1][500/782] Time 0.343 (0.353) Loss 3.821 (4.011) Acc@1 25.78 (20.93) Acc@5 50.78 (43.01)
Epoch:[1][550/782] Time 0.466 (0.355) Loss 3.625 (3.988) Acc@1 28.91 (21.30) Acc@5 50.00 (43.57)
Epoch:[1][600/782] Time 0.357 (0.354) Loss 3.691 (3.969) Acc@1 28.12 (21.55) Acc@5 46.09 (44.08)
Epoch:[1][650/782] Time 0.349 (0.354) Loss 3.736 (3.951) Acc@1 22.66 (21.75) Acc@5 47.66 (44.59)
Epoch:[1][700/782] Time 0.348 (0.354) Loss 3.740 (3.935) Acc@1 25.00 (21.97) Acc@5 44.53 (44.94)
Epoch:[1][750/782] Time 0.345 (0.359) Loss 3.619 (3.917) Acc@1 32.03 (22.25) Acc@5 52.34 (45.38)
Epoch:[2][ 0/782] Time 0.689 (0.689) Loss 3.427 (3.427) Acc@1 28.91 (28.91) Acc@5 58.59 (58.59)
Epoch:[2][ 50/782] Time 0.346 (0.356) Loss 3.394 (3.465) Acc@1 34.38 (29.43) Acc@5 58.59 (55.38)
Epoch:[2][100/782] Time 0.352 (0.366) Loss 3.294 (3.432) Acc@1 34.38 (30.09) Acc@5 60.16 (56.66)
Epoch:[2][150/782] Time 0.346 (0.360) Loss 3.359 (3.422) Acc@1 32.81 (30.34) Acc@5 61.72 (56.80)
Epoch:[2][200/782] Time 0.341 (0.357) Loss 3.217 (3.409) Acc@1 35.16 (30.47) Acc@5 64.06 (57.05)
Epoch:[2][250/782] Time 0.345 (0.356) Loss 3.366 (3.391) Acc@1 30.47 (30.75) Acc@5 57.03 (57.27)
Epoch:[2][300/782] Time 0.345 (0.359) Loss 3.484 (3.378) Acc@1 25.00 (30.86) Acc@5 51.56 (57.40)
Epoch:[2][350/782] Time 0.341 (0.358) Loss 3.327 (3.369) Acc@1 33.59 (30.97) Acc@5 61.72 (57.54)
Epoch:[2][400/782] Time 0.350 (0.356) Loss 3.425 (3.358) Acc@1 25.78 (31.11) Acc@5 58.59 (57.65)
Epoch:[2][450/782] Time 0.347 (0.358) Loss 3.440 (3.346) Acc@1 28.12 (31.40) Acc@5 59.38 (57.85)
Epoch:[2][500/782] Time 0.348 (0.358) Loss 3.122 (3.335) Acc@1 35.16 (31.58) Acc@5 66.41 (58.07)
Epoch:[2][550/782] Time 0.351 (0.357) Loss 3.368 (3.321) Acc@1 28.91 (31.76) Acc@5 55.47 (58.41)
Epoch:[2][600/782] Time 0.339 (0.358) Loss 3.077 (3.309) Acc@1 37.50 (31.89) Acc@5 64.06 (58.55)
Epoch:[2][650/782] Time 0.338 (0.357) Loss 3.177 (3.299) Acc@1 35.94 (32.09) Acc@5 64.84 (58.72)
Epoch:[2][700/782] Time 0.352 (0.357) Loss 3.139 (3.286) Acc@1 33.59 (32.24) Acc@5 60.94 (58.95)
Epoch:[2][750/782] Time 0.350 (0.356) Loss 3.238 (3.274) Acc@1 35.94 (32.48) Acc@5 57.81 (59.19)
Epoch:[3][ 0/782] Time 0.693 (0.693) Loss 3.069 (3.069) Acc@1 33.59 (33.59) Acc@5 64.84 (64.84)
Epoch:[3][ 50/782] Time 0.341 (0.355) Loss 2.916 (2.955) Acc@1 44.53 (38.13) Acc@5 64.06 (65.06)
Epoch:[3][100/782] Time 0.346 (0.353) Loss 3.027 (2.936) Acc@1 34.38 (38.15) Acc@5 63.28 (65.73)
Epoch:[3][150/782] Time 0.457 (0.360) Loss 2.753 (2.933) Acc@1 39.84 (38.11) Acc@5 70.31 (65.45)
Epoch:[3][200/782] Time 0.349 (0.356) Loss 3.030 (2.928) Acc@1 35.16 (38.22) Acc@5 59.38 (65.42)
Epoch:[3][250/782] Time 0.337 (0.354) Loss 2.841 (2.923) Acc@1 33.59 (38.19) Acc@5 67.19 (65.40)
Epoch:[3][300/782] Time 0.344 (0.353) Loss 2.888 (2.918) Acc@1 42.97 (38.22) Acc@5 71.88 (65.44)
Epoch:[3][350/782] Time 0.361 (0.357) Loss 2.760 (2.914) Acc@1 40.62 (38.24) Acc@5 69.53 (65.44)
Epoch:[3][400/782] Time 0.347 (0.355) Loss 3.104 (2.907) Acc@1 30.47 (38.21) Acc@5 59.38 (65.51)
Epoch:[3][450/782] Time 0.358 (0.355) Loss 2.911 (2.901) Acc@1 35.94 (38.32) Acc@5 62.50 (65.57)
Epoch:[3][500/782] Time 0.355 (0.357) Loss 2.736 (2.894) Acc@1 41.41 (38.37) Acc@5 64.84 (65.73)
Epoch:[3][550/782] Time 0.360 (0.357) Loss 3.151 (2.888) Acc@1 29.69 (38.40) Acc@5 60.16 (65.81)
Epoch:[3][600/782] Time 0.344 (0.356) Loss 3.021 (2.883) Acc@1 30.47 (38.44) Acc@5 59.38 (65.83)
Epoch:[3][650/782] Time 0.486 (0.357) Loss 2.929 (2.876) Acc@1 41.41 (38.55) Acc@5 66.41 (65.97)
Epoch:[3][700/782] Time 0.349 (0.357) Loss 2.975 (2.869) Acc@1 33.59 (38.60) Acc@5 62.50 (66.06)
Epoch:[3][750/782] Time 0.371 (0.356) Loss 2.790 (2.863) Acc@1 39.06 (38.73) Acc@5 64.84 (66.12)
Epoch:[4][ 0/782] Time 0.684 (0.684) Loss 2.629 (2.629) Acc@1 46.88 (46.88) Acc@5 67.97 (67.97)
Epoch:[4][ 50/782] Time 0.345 (0.383) Loss 2.676 (2.725) Acc@1 45.31 (40.36) Acc@5 67.19 (68.01)
Epoch:[4][100/782] Time 0.343 (0.366) Loss 2.824 (2.698) Acc@1 32.81 (41.15) Acc@5 66.41 (68.69)
Epoch:[4][150/782] Time 0.341 (0.361) Loss 2.700 (2.689) Acc@1 46.88 (41.41) Acc@5 62.50 (69.01)
Epoch:[4][200/782] Time 0.468 (0.360) Loss 2.516 (2.682) Acc@1 46.88 (41.59) Acc@5 75.00 (69.14)
Epoch:[4][250/782] Time 0.349 (0.361) Loss 2.395 (2.676) Acc@1 49.22 (41.80) Acc@5 73.44 (69.18)
Epoch:[4][300/782] Time 0.349 (0.359) Loss 2.625 (2.673) Acc@1 42.19 (41.85) Acc@5 65.62 (69.08)
Epoch:[4][350/782] Time 0.350 (0.357) Loss 2.616 (2.670) Acc@1 46.88 (41.88) Acc@5 71.88 (69.12)
Epoch:[4][400/782] Time 0.350 (0.360) Loss 2.459 (2.661) Acc@1 42.97 (42.00) Acc@5 72.66 (69.26)
Epoch:[4][450/782] Time 0.349 (0.359) Loss 2.520 (2.657) Acc@1 45.31 (42.02) Acc@5 75.00 (69.26)
Epoch:[4][500/782] Time 0.341 (0.358) Loss 2.788 (2.653) Acc@1 37.50 (42.08) Acc@5 64.84 (69.31)
Epoch:[4][550/782] Time 0.347 (0.359) Loss 2.466 (2.645) Acc@1 43.75 (42.25) Acc@5 68.75 (69.41)
Epoch:[4][600/782] Time 0.348 (0.358) Loss 2.392 (2.640) Acc@1 51.56 (42.30) Acc@5 73.44 (69.44)
Epoch:[4][650/782] Time 0.347 (0.358) Loss 2.593 (2.636) Acc@1 41.41 (42.33) Acc@5 71.09 (69.45)
Epoch:[4][700/782] Time 0.474 (0.357) Loss 2.537 (2.633) Acc@1 38.28 (42.34) Acc@5 73.44 (69.46)
Epoch:[4][750/782] Time 0.353 (0.358) Loss 2.407 (2.626) Acc@1 42.19 (42.42) Acc@5 76.56 (69.60)
Epoch:[5][ 0/782] Time 0.710 (0.710) Loss 2.314 (2.314) Acc@1 49.22 (49.22) Acc@5 73.44 (73.44)
Epoch:[5][ 50/782] Time 0.348 (0.355) Loss 2.585 (2.519) Acc@1 43.75 (43.64) Acc@5 69.53 (71.03)
Epoch:[5][100/782] Time 0.482 (0.378) Loss 2.277 (2.489) Acc@1 46.88 (44.35) Acc@5 76.56 (71.71)
Epoch:[5][150/782] Time 0.351 (0.380) Loss 2.283 (2.479) Acc@1 52.34 (44.65) Acc@5 75.78 (71.80)
Epoch:[5][200/782] Time 0.353 (0.372) Loss 2.444 (2.478) Acc@1 46.88 (44.71) Acc@5 69.53 (71.70)
Epoch:[5][250/782] Time 0.459 (0.372) Loss 2.566 (2.481) Acc@1 42.97 (44.73) Acc@5 69.53 (71.67)
Epoch:[5][300/782] Time 0.350 (0.369) Loss 2.404 (2.474) Acc@1 42.19 (44.81) Acc@5 77.34 (71.83)
Epoch:[5][350/782] Time 0.347 (0.367) Loss 2.306 (2.476) Acc@1 50.78 (44.73) Acc@5 77.34 (71.68)
Epoch:[5][400/782] Time 0.348 (0.364) Loss 2.418 (2.471) Acc@1 43.75 (44.84) Acc@5 72.66 (71.75)
Epoch:[5][450/782] Time 0.354 (0.366) Loss 2.359 (2.465) Acc@1 51.56 (44.92) Acc@5 74.22 (71.87)
Epoch:[5][500/782] Time 0.351 (0.364) Loss 2.418 (2.463) Acc@1 47.66 (44.95) Acc@5 75.00 (71.86)
Epoch:[5][550/782] Time 0.347 (0.363) Loss 2.405 (2.459) Acc@1 42.19 (45.00) Acc@5 71.09 (71.89)
Epoch:[5][600/782] Time 0.353 (0.364) Loss 2.330 (2.457) Acc@1 50.00 (45.04) Acc@5 76.56 (71.92)
Epoch:[5][650/782] Time 0.349 (0.363) Loss 2.273 (2.451) Acc@1 48.44 (45.10) Acc@5 72.66 (72.01)
Epoch:[5][700/782] Time 0.346 (0.362) Loss 2.231 (2.446) Acc@1 46.09 (45.19) Acc@5 72.66 (72.09)
Epoch:[5][750/782] Time 0.471 (0.362) Loss 2.482 (2.442) Acc@1 50.78 (45.26) Acc@5 67.19 (72.14)
Epoch:[6][ 0/782] Time 0.696 (0.696) Loss 2.563 (2.563) Acc@1 43.75 (43.75) Acc@5 64.06 (64.06)
Epoch:[6][ 50/782] Time 0.344 (0.355) Loss 2.414 (2.318) Acc@1 46.09 (47.76) Acc@5 70.31 (74.16)
Epoch:[6][100/782] Time 0.347 (0.352) Loss 2.413 (2.332) Acc@1 46.88 (47.07) Acc@5 71.88 (73.90)
Epoch:[6][150/782] Time 0.353 (0.359) Loss 2.217 (2.330) Acc@1 48.44 (47.10) Acc@5 75.78 (73.78)
Epoch:[6][200/782] Time 0.344 (0.356) Loss 2.341 (2.328) Acc@1 48.44 (47.20) Acc@5 73.44 (73.74)
Epoch:[6][250/782] Time 0.358 (0.355) Loss 2.578 (2.330) Acc@1 43.75 (47.19) Acc@5 67.19 (73.85)
Epoch:[6][300/782] Time 0.476 (0.358) Loss 2.454 (2.321) Acc@1 43.75 (47.48) Acc@5 71.88 (74.04)
Epoch:[6][350/782] Time 0.339 (0.358) Loss 2.336 (2.320) Acc@1 49.22 (47.46) Acc@5 75.00 (74.05)
Epoch:[6][400/782] Time 0.348 (0.356) Loss 2.060 (2.316) Acc@1 50.78 (47.57) Acc@5 81.25 (74.07)
Epoch:[6][450/782] Time 0.347 (0.356) Loss 2.363 (2.316) Acc@1 46.09 (47.42) Acc@5 71.88 (74.07)
Epoch:[6][500/782] Time 0.348 (0.358) Loss 2.333 (2.312) Acc@1 49.22 (47.43) Acc@5 70.31 (74.11)
Epoch:[6][550/782] Time 0.337 (0.357) Loss 2.198 (2.308) Acc@1 46.88 (47.51) Acc@5 75.00 (74.18)
Epoch:[6][600/782] Time 0.345 (0.356) Loss 2.199 (2.304) Acc@1 58.59 (47.62) Acc@5 77.34 (74.20)
Epoch:[6][650/782] Time 0.355 (0.357) Loss 2.126 (2.303) Acc@1 51.56 (47.62) Acc@5 80.47 (74.24)
Epoch:[6][700/782] Time 0.345 (0.357) Loss 2.313 (2.298) Acc@1 39.84 (47.71) Acc@5 71.88 (74.32)
Epoch:[6][750/782] Time 0.343 (0.356) Loss 2.078 (2.294) Acc@1 55.47 (47.77) Acc@5 78.12 (74.35)
Epoch:[7][ 0/782] Time 0.697 (0.697) Loss 2.202 (2.202) Acc@1 43.75 (43.75) Acc@5 75.78 (75.78)
Epoch:[7][ 50/782] Time 0.354 (0.384) Loss 2.119 (2.211) Acc@1 53.12 (48.94) Acc@5 76.56 (75.41)
Epoch:[7][100/782] Time 0.353 (0.366) Loss 2.285 (2.211) Acc@1 55.47 (49.30) Acc@5 71.09 (75.46)
Epoch:[7][150/782] Time 0.357 (0.361) Loss 1.987 (2.207) Acc@1 56.25 (49.26) Acc@5 81.25 (75.51)
Epoch:[7][200/782] Time 0.344 (0.364) Loss 2.240 (2.202) Acc@1 47.66 (49.49) Acc@5 75.00 (75.61)
Epoch:[7][250/782] Time 0.348 (0.361) Loss 2.206 (2.202) Acc@1 48.44 (49.41) Acc@5 77.34 (75.70)
Epoch:[7][300/782] Time 0.352 (0.359) Loss 2.387 (2.201) Acc@1 51.56 (49.46) Acc@5 69.53 (75.77)
Epoch:[7][350/782] Time 0.380 (0.358) Loss 2.073 (2.195) Acc@1 42.19 (49.53) Acc@5 81.25 (75.92)
Epoch:[7][400/782] Time 0.360 (0.361) Loss 1.702 (2.193) Acc@1 63.28 (49.61) Acc@5 84.38 (75.91)
Epoch:[7][450/782] Time 0.352 (0.359) Loss 2.209 (2.193) Acc@1 48.44 (49.64) Acc@5 76.56 (75.92)
Epoch:[7][500/782] Time 0.343 (0.359) Loss 2.164 (2.191) Acc@1 48.44 (49.61) Acc@5 76.56 (75.86)
Epoch:[7][550/782] Time 0.343 (0.360) Loss 2.102 (2.186) Acc@1 46.88 (49.73) Acc@5 78.91 (75.89)
Epoch:[7][600/782] Time 0.348 (0.359) Loss 2.209 (2.183) Acc@1 47.66 (49.76) Acc@5 71.88 (75.90)
Epoch:[7][650/782] Time 0.348 (0.358) Loss 2.071 (2.180) Acc@1 49.22 (49.84) Acc@5 75.78 (75.91)
Epoch:[7][700/782] Time 0.347 (0.359) Loss 2.158 (2.178) Acc@1 47.66 (49.87) Acc@5 75.00 (75.93)
Epoch:[7][750/782] Time 0.349 (0.359) Loss 2.076 (2.177) Acc@1 52.34 (49.89) Acc@5 76.56 (75.91)
Epoch:[8][ 0/782] Time 0.707 (0.707) Loss 1.827 (1.827) Acc@1 57.81 (57.81) Acc@5 82.81 (82.81)
Epoch:[8][ 50/782] Time 0.362 (0.360) Loss 2.204 (2.097) Acc@1 50.00 (51.07) Acc@5 77.34 (77.28)
Epoch:[8][100/782] Time 0.348 (0.368) Loss 2.199 (2.089) Acc@1 53.12 (51.28) Acc@5 73.44 (77.47)
Epoch:[8][150/782] Time 0.348 (0.362) Loss 2.295 (2.101) Acc@1 47.66 (51.02) Acc@5 74.22 (77.14)
Epoch:[8][200/782] Time 0.348 (0.359) Loss 2.163 (2.101) Acc@1 46.09 (50.88) Acc@5 77.34 (77.20)
Epoch:[8][250/782] Time 0.473 (0.366) Loss 2.244 (2.092) Acc@1 46.09 (51.15) Acc@5 71.88 (77.31)
Epoch:[8][300/782] Time 0.343 (0.370) Loss 2.068 (2.087) Acc@1 51.56 (51.28) Acc@5 76.56 (77.41)
Epoch:[8][350/782] Time 0.350 (0.367) Loss 1.878 (2.083) Acc@1 53.12 (51.31) Acc@5 81.25 (77.41)
Epoch:[8][400/782] Time 0.487 (0.366) Loss 2.356 (2.084) Acc@1 39.84 (51.37) Acc@5 72.66 (77.38)
Epoch:[8][450/782] Time 0.347 (0.366) Loss 1.727 (2.084) Acc@1 61.72 (51.35) Acc@5 82.81 (77.35)
Epoch:[8][500/782] Time 0.357 (0.364) Loss 2.142 (2.082) Acc@1 46.09 (51.33) Acc@5 78.12 (77.37)
Epoch:[8][550/782] Time 0.345 (0.363) Loss 2.170 (2.079) Acc@1 52.34 (51.39) Acc@5 74.22 (77.42)
Epoch:[8][600/782] Time 0.349 (0.365) Loss 2.189 (2.076) Acc@1 54.69 (51.51) Acc@5 74.22 (77.46)
Epoch:[8][650/782] Time 0.349 (0.364) Loss 2.114 (2.074) Acc@1 50.00 (51.54) Acc@5 79.69 (77.47)
Epoch:[8][700/782] Time 0.358 (0.363) Loss 2.255 (2.074) Acc@1 53.12 (51.53) Acc@5 73.44 (77.44)
Epoch:[8][750/782] Time 0.341 (0.364) Loss 2.060 (2.071) Acc@1 54.69 (51.57) Acc@5 76.56 (77.46)
Epoch:[9][ 0/782] Time 0.688 (0.688) Loss 1.831 (1.831) Acc@1 56.25 (56.25) Acc@5 84.38 (84.38)
Epoch:[9][ 50/782] Time 0.356 (0.358) Loss 2.054 (1.996) Acc@1 48.44 (53.31) Acc@5 81.25 (78.81)
Epoch:[9][100/782] Time 0.339 (0.353) Loss 1.864 (1.998) Acc@1 59.38 (53.34) Acc@5 82.03 (78.36)
Epoch:[9][150/782] Time 0.353 (0.359) Loss 2.027 (1.993) Acc@1 51.56 (53.24) Acc@5 80.47 (78.79)
Epoch:[9][200/782] Time 0.352 (0.357) Loss 1.873 (1.994) Acc@1 57.81 (53.24) Acc@5 81.25 (78.72)
Epoch:[9][250/782] Time 0.349 (0.355) Loss 2.171 (1.996) Acc@1 47.66 (53.16) Acc@5 75.00 (78.60)
Epoch:[9][300/782] Time 0.344 (0.359) Loss 2.138 (1.997) Acc@1 52.34 (53.12) Acc@5 75.78 (78.49)
Epoch:[9][350/782] Time 0.357 (0.357) Loss 2.202 (1.995) Acc@1 44.53 (53.02) Acc@5 75.00 (78.57)
Epoch:[9][400/782] Time 0.354 (0.357) Loss 1.884 (1.994) Acc@1 59.38 (52.95) Acc@5 81.25 (78.53)
Epoch:[9][450/782] Time 0.478 (0.357) Loss 2.046 (1.988) Acc@1 51.56 (53.07) Acc@5 75.78 (78.60)
Epoch:[9][500/782] Time 0.349 (0.358) Loss 2.284 (1.990) Acc@1 46.88 (53.00) Acc@5 72.66 (78.62)
Epoch:[9][550/782] Time 0.342 (0.357) Loss 1.614 (1.990) Acc@1 65.62 (53.05) Acc@5 82.81 (78.60)
Epoch:[9][600/782] Time 0.341 (0.356) Loss 1.783 (1.986) Acc@1 53.12 (53.10) Acc@5 85.16 (78.65)
Epoch:[9][650/782] Time 0.346 (0.358) Loss 1.669 (1.983) Acc@1 60.94 (53.14) Acc@5 82.81 (78.71)
Epoch:[9][700/782] Time 0.351 (0.357) Loss 2.272 (1.982) Acc@1 41.41 (53.14) Acc@5 75.78 (78.75)
Epoch:[9][750/782] Time 0.346 (0.357) Loss 1.714 (1.982) Acc@1 59.38 (53.12) Acc@5 80.47 (78.71)
Validating
Test: [ 0/79] Time 0.415 (0.415) Loss 4.184 (4.184) Acc@1 8.59 (8.59) Acc@5 31.25 (31.25)
Test: [10/79] Time 0.153 (0.173) Loss 5.948 (4.814) Acc@1 3.12 (7.67) Acc@5 6.25 (21.24)
Test: [20/79] Time 0.160 (0.162) Loss 6.329 (5.114) Acc@1 0.00 (4.95) Acc@5 3.91 (17.26)
Test: [30/79] Time 0.140 (0.157) Loss 5.530 (5.322) Acc@1 0.78 (4.11) Acc@5 17.19 (14.42)
Test: [40/79] Time 0.119 (0.153) Loss 5.589 (5.396) Acc@1 6.25 (4.04) Acc@5 8.59 (13.38)
Test: [50/79] Time 0.139 (0.151) Loss 4.862 (5.493) Acc@1 7.03 (3.80) Acc@5 23.44 (12.65)
Test: [60/79] Time 0.142 (0.150) Loss 5.924 (5.506) Acc@1 0.00 (3.92) Acc@5 6.25 (12.59)
Test: [70/79] Time 0.154 (0.148) Loss 4.818 (5.519) Acc@1 3.12 (3.71) Acc@5 10.16 (11.95)
* Acc@1 5.190 Acc@5 15.180
Accuracy of tuned INT8 sparse model: 5.190
Accuracy drop of tuned INT8 sparse model over pre-trained FP32 model: -4.620
Export INT8 Sparse Model to OpenVINO IR#
warnings.filterwarnings("ignore", category=TracerWarning)
warnings.filterwarnings("ignore", category=UserWarning)
# Export INT8 model to OpenVINO™ IR
ov_model = ov.convert_model(compressed_model, example_input=dummy_input, input=[1, 3, image_size, image_size])
ov.save_model(ov_model, int8_sparse_ir_path)
print(f"INT8 sparse model exported to {int8_sparse_ir_path}.")
WARNING:tensorflow:Please fix your imports. Module tensorflow.python.training.tracking.base has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.
INT8 sparse model exported to model/resnet18_int8_sparse.xml.
Benchmark Model Performance by Computing Inference Time#
Finally, measure the inference performance of the FP32 and INT8
models, using Benchmark
Tool
- inference performance measurement tool in OpenVINO. By default,
Benchmark Tool runs inference for 60 seconds in asynchronous mode on
CPU. It returns inference speed as latency (milliseconds per image) and
throughput (frames per second) values.
NOTE: This notebook runs
benchmark_appfor 15 seconds to give a quick indication of performance. For more accurate performance, it is recommended to runbenchmark_appin a terminal/command prompt after closing other applications. Runbenchmark_app -m model.xml -d CPUto benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Runbenchmark_app --helpto see an overview of all command-line options.
# Initialize OpenVINO runtime
core = ov.Core()
device = device_widget()
device
Dropdown(description='Device:', index=1, options=('CPU', 'AUTO'), value='AUTO')
def parse_benchmark_output(benchmark_output):
parsed_output = [line for line in benchmark_output if "FPS" in line]
print(*parsed_output, sep="\n")
print("Benchmark FP32 model (IR)")
benchmark_output = ! benchmark_app -m $fp32_ir_path -d $device.value -api async -t 15
parse_benchmark_output(benchmark_output)
print("Benchmark INT8 sparse model (IR)")
benchmark_output = ! benchmark_app -m $int8_ir_path -d $device.value -api async -t 15
parse_benchmark_output(benchmark_output)
Benchmark FP32 model (IR)
[ INFO ] Throughput: 2939.59 FPS
Benchmark INT8 sparse model (IR)
Show Device Information for reference.
import openvino.properties as props
core.get_property(device.value, props.device.full_name)
'AUTO'