Skip to content

GitLab

Commit 3aa48ea8 authored by Fan Yang's avatar Fan Yang Committed by A. Unique TensorFlower
Browse files

Internal change 

PiperOrigin-RevId: 428078415
parent f670e89c
Hide whitespace changes
Inline Side-by-side
Showing with 1126 additions and 0 deletions
official/projects/qat/vision/README.md 0 → 100644
View file @ 3aa48ea8
# Quantization Aware Training Project for Computer Vision Models
[TOC]
⚠️ Disclaimer: All datasets hyperlinked from this page are not owned or
distributed by Google. The dataset is made available by third parties.
Please review the terms and conditions made available by the third parties
before using the data.
## Overview
This project includes quantization aware training code for Computer Vision
models. These are examples to show how to apply the Model Optimization Toolkit's
[quantization aware training API](https://www.tensorflow.org/model_optimization/guide/quantization/training).
Note: Currently, we support a limited number of ML tasks & models (e.g., image
classification and semantic segmentation)
We will keep adding support for other ML tasks and models in the next releases.
## How to train a model
```
EXPERIMENT=xxx # Change this for your run, for example, 'mobilenet_imagenet_qat'
CONFIG_FILE=xxx # Change this for your run, for example, path of imagenet_mobilenetv2_qat_gpu.yaml
MODEL_DIR=xxx # Change this for your run, for example, /tmp/model_dir
$ python3 train.py \
--experiment=${EXPERIMENT} \
--config_file=${CONFIG_FILE} \
--model_dir=${MODEL_DIR} \
--mode=train_and_eval
```
## Model Accuracy
<figure align="center">
<img width=70% src=https://storage.googleapis.com/tf_model_garden/models/qat/images/readme-qat-classification-plot.png>
<figcaption>Comparison of Imagenet top-1 accuracy for the classification models</figcaption>
</figure>
Note: The Top-1 model accuracy is measured on the validation set of [ImageNet](https://www.image-net.org/).
### Pre-trained Models
|Model |Resolution|Top-1 Accuracy (FP32)|Top-1 Accuracy (Int8/PTQ)|Top-1 Accuracy (Int8/QAT)|Config |Download |
|----------------------|----------|---------------------|-------------------------|-------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------|
|MobileNetV2 |224x224 |72.782% |72.392% |72.792% |[config](https://github.com/tensorflow/models/blob/master/official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv2_qat_gpu.yaml) |[TFLite(Int8/QAT)](https://storage.googleapis.com/tf_model_garden/vision/mobilenet/v2_1.0_int8/mobilenet_v2_1.00_224_int8.tflite) |
|ResNet50 |224x224 |76.710% |76.420% |77.200% |[config](https://github.com/tensorflow/models/blob/master/official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu.yaml) |[TFLite(Int8/QAT)](https://storage.googleapis.com/tf_model_garden/vision/resnet50_imagenet/resnet_50_224_int8.tflite) |
|MobileNetV3.5 MultiAVG|224x224 |75.212% |74.122% |75.130% |[config](https://github.com/tensorflow/models/blob/master/official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv3.5_qat_gpu.yaml)|[TFLite(Int8/QAT)](https://storage.googleapis.com/tf_model_garden/vision/mobilenet/v3.5multiavg_1.0_int8/mobilenet_v3.5multiavg_1.00_224_int8.tflite)|
official/projects/qat/vision/configs/__init__.py 0 → 100644
View file @ 3aa48ea8
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Lint as: python3
"""Configs package definition."""
from official.projects.qat.vision.configs import image_classification
from official.projects.qat.vision.configs import semantic_segmentation
official/projects/qat/vision/configs/common.py 0 → 100644
View file @ 3aa48ea8
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Lint as: python3
"""Image classification configuration definition."""
import dataclasses
from typing import Optional
from official.modeling import hyperparams
@dataclasses.dataclass
class Quantization(hyperparams.Config):
"""Quantization parameters.
Attributes:
pretrained_original_checkpoint: A string indicate pretrained checkpoint
location.
change_num_bits: A `bool` indicates whether to manually allocate num_bits.
num_bits_weight: An `int` number of bits for weight. Default to 8.
num_bits_activation: An `int` number of bits for activation. Default to 8.
"""
pretrained_original_checkpoint: Optional[str] = None
change_num_bits: bool = False
num_bits_weight: int = 8
num_bits_activation: int = 8
official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv2_qat_gpu.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
filter_size_scale: 1.0
dropout_rate: 0.1
losses:
l2_weight_decay: 0.0000001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 512 # 64 * 8
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 512 # 64 * 8
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/mobilenetv2_gpu/22984194/ckpt-625500'
trainer:
# With below setting, the accuracy of QAT reaches to accuracy 0.7279 after 43 hours with 8 GPUS.
train_steps: 250200
validation_steps: 98
validation_interval: 2502
steps_per_loop: 2502
summary_interval: 2502
checkpoint_interval: 2502
optimizer_config:
learning_rate:
type: 'exponential'
exponential:
decay_rate: 0.9
decay_steps: 1251
initial_learning_rate: 0.0001
name: 'ExponentialDecay'
offset: 0
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 0
official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv2_qat_gpu_batch256.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
filter_size_scale: 1.0
dropout_rate: 0.0 # changed from 0.2 to 0.0
losses:
l2_weight_decay: 0.0000001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/mobilenetv2_gpu/22984194/ckpt-625500'
trainer:
# With below setting, the accuracy of QAT reaches Top1-accuracy 0.7251 at 420336 steps after
# 1 day 19 hours of training with 8GPUs, which is higher than the result of PTQ in MobileNetV2.
train_steps: 1000800 # 200 epochs
validation_steps: 196 # NUM_EXAMPLES (50000) // global_batch_size (256)
validation_interval: 5004 # 1 epoch
steps_per_loop: 5004 # NUM_EXAMPLES (1281167) // global_batch_size (256)
summary_interval: 5004 # 1 epoch
checkpoint_interval: 5004 # 1 epoch
max_to_keep: 200
optimizer_config:
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.0001
decay_steps: 1251 # steps_per_epoch // 4
decay_rate: 0.96
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 0
official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv2_qat_gpu_batch512.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
filter_size_scale: 1.0
dropout_rate: 0.0 # changed from 0.2 to 0.0
losses:
l2_weight_decay: 0.0000001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 512
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 512
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/mobilenetv2_gpu/22984194/ckpt-625500'
trainer:
# With below setting, the accuracy of QAT reaches Top1-accuracy 0.7266 at 312750 steps after
# 1 day 22 hours of training with 8GPUs, which is higher than the result of PTQ in MobileNetV2.
train_steps: 500400 # 200 epochs
validation_steps: 98 # NUM_EXAMPLES (50000) // global_batch_size (512)
validation_interval: 2502 # 1 epoch
steps_per_loop: 2502 # NUM_EXAMPLES (1281167) // global_batch_size (512)
summary_interval: 2502 # 1 epoch
checkpoint_interval: 2502 # 1 epoch
max_to_keep: 200
optimizer_config:
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.0002
decay_steps: 1251 # steps_per_epoch // 2
decay_rate: 0.96
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 0
official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv3.5_qat_gpu.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetMultiAVG'
filter_size_scale: 1.0
dropout_rate: 0.3
losses:
l2_weight_decay: 0.000001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 512
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 512
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/tf2_mhave_nobias_bn_aug05/28334857/ckpt-156000'
trainer:
# With below setting, the accuracy of QAT reaches to accuracy 0.7513 after 30 hours with 8 GPUS.
train_steps: 250200
validation_steps: 98
validation_interval: 2502
steps_per_loop: 2502
summary_interval: 2502
checkpoint_interval: 2502
optimizer_config:
learning_rate:
type: 'exponential'
exponential:
decay_rate: 0.9
decay_steps: 1251
initial_learning_rate: 0.0004
name: 'ExponentialDecay'
offset: 0
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 0
official/projects/qat/vision/configs/experiments/image_classification/imagenet_mobilenetv3large_qat_tpu.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'tpu'
mixed_precision_dtype: 'float32'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV3Large'
filter_size_scale: 1.0
dropout_rate: 0.3
losses:
l2_weight_decay: 1.0e-06 # 1/10 of original value.
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 4096
dtype: 'float32'
aug_rand_hflip: true
aug_type:
autoaug:
augmentation_name: v0
cutout_const: 100
translate_const: 250
type: autoaug
drop_remainder: true
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 4096
dtype: 'float32'
drop_remainder: false
aug_rand_hflip: true
quantization:
pretrained_original_checkpoint: 'gs://**/mobilenetv3_baseline_31/ckpt-156000'
trainer:
# With below setting, the accuracy of QAT reaches to accuracy 0.74.43 after ~2 hours with 4x4 DF.
train_steps: 62400
validation_steps: 13
validation_interval: 312
steps_per_loop: 312
summary_interval: 312
checkpoint_interval: 312
optimizer_config:
learning_rate:
cosine:
alpha: 0.0
decay_steps: 62400
initial_learning_rate: 0.0003 # 1/10 of original lr.
name: CosineDecay
offset: 0
type: cosine
optimizer:
adamw:
amsgrad: false
beta_1: 0.9
beta_2: 0.999
epsilon: 1.0e-07
gradient_clip_norm: 1.0
weight_decay_rate: 0.0
type: adamw
warmup:
type: 'linear'
linear:
warmup_steps: 0
official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'resnet'
resnet:
model_id: 50
losses:
l2_weight_decay: 0.0001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/resnet_classifier_gpu/ckpt-56160'
trainer:
# With below setting, the accuracy of QAT reaches to Top1-accuracy 0.7720 after 5 days of training
# with 8GPUs, which is higher than the non-quantized float32 version Resnet.
train_steps: 449280
validation_steps: 200
validation_interval: 5000
steps_per_loop: 5000
summary_interval: 5000
checkpoint_interval: 5000
optimizer_config:
optimizer:
type: 'sgd'
sgd:
momentum: 0.9
learning_rate:
type: 'stepwise'
stepwise:
boundaries: [150000, 300000, 400000]
values: [0.08, 0.008, 0.0008, 0.00008]
warmup:
type: 'linear'
linear:
warmup_steps: 40000
official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu_fast.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'resnet'
resnet:
model_id: 50
losses:
l2_weight_decay: 0.0001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/resnet_classifier_gpu/ckpt-56160'
trainer:
# With below setting, the accuracy of QAT reaches to the non-quantized float32 version after
# around 160k steps, which takes 1d 15h with 8 GPUS.
train_steps: 449280
validation_steps: 200
validation_interval: 5000
steps_per_loop: 5000
summary_interval: 5000
checkpoint_interval: 5000
optimizer_config:
optimizer:
type: 'sgd'
sgd:
momentum: 0.9
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.016
decay_steps: 25000
decay_rate: 0.5
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 1000
official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu_fast_4x4.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'resnet'
resnet:
model_id: 50
losses:
l2_weight_decay: 0.0001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/resnet_classifier_gpu/ckpt-56160'
change_num_bits: true
num_bits_weight: 4
num_bits_activation: 4
trainer:
# With below setting, the accuracy of QAT reaches Top1-accuracy 0.6822 at 205k steps with 8GPUs.
# TODO: Please change the configs when training is done.
train_steps: 449280
validation_steps: 200
validation_interval: 5000
steps_per_loop: 5000
summary_interval: 5000
checkpoint_interval: 5000
optimizer_config:
optimizer:
type: 'sgd'
sgd:
momentum: 0.9
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.016
decay_steps: 25000
decay_rate: 0.5
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 1000
official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu_fast_4x8.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'resnet'
resnet:
model_id: 50
losses:
l2_weight_decay: 0.0001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/resnet_classifier_gpu/ckpt-56160'
change_num_bits: true
num_bits_weight: 4
num_bits_activation: 8
trainer:
# With below setting, the accuracy of QAT reaches Top1-accuracy 0.7575 at 220k steps with 8GPUs.
# TODO: Please change the configs when training is done.
train_steps: 449280
validation_steps: 200
validation_interval: 5000
steps_per_loop: 5000
summary_interval: 5000
checkpoint_interval: 5000
optimizer_config:
optimizer:
type: 'sgd'
sgd:
momentum: 0.9
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.016
decay_steps: 25000
decay_rate: 0.5
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 1000
official/projects/qat/vision/configs/experiments/image_classification/imagenet_resnet50_qat_gpu_fast_6x6.yaml 0 → 100644
View file @ 3aa48ea8
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 1001
input_size: [224, 224, 3]
backbone:
type: 'resnet'
resnet:
model_id: 50
losses:
l2_weight_decay: 0.0001
one_hot: true
label_smoothing: 0.1
train_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/train*'
is_training: true
global_batch_size: 256
dtype: 'float32'
validation_data:
input_path: '/readahead/200M/placer/prod/home/distbelief/imagenet-tensorflow/imagenet-2012-tfrecord/valid*'
is_training: false
global_batch_size: 256
dtype: 'float32'
drop_remainder: false
quantization:
pretrained_original_checkpoint: 'gs://**/resnet_classifier_gpu/ckpt-56160'
change_num_bits: true
num_bits_weight: 6
num_bits_activation: 6
trainer:
# With below setting, the accuracy of QAT reaches Top1-accuracy 0.7607 at 190k steps with 8GPUs.
# TODO: Please change the configs when training is done.
train_steps: 449280
validation_steps: 200
validation_interval: 5000
steps_per_loop: 5000
summary_interval: 5000
checkpoint_interval: 5000
optimizer_config:
optimizer:
type: 'sgd'
sgd:
momentum: 0.9
learning_rate:
type: 'exponential'
exponential:
initial_learning_rate: 0.016
decay_steps: 25000
decay_rate: 0.5
staircase: true
warmup:
type: 'linear'
linear:
warmup_steps: 1000
official/projects/qat/vision/configs/experiments/retinanet/coco_spinenet49_mobile_qat_gpu.yaml 0 → 100644
View file @ 3aa48ea8
# --experiment_type=retinanet_spinenet_mobile_coco_qat
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
task:
losses:
l2_weight_decay: 3.0e-05
model:
anchor:
anchor_size: 3
aspect_ratios: [0.5, 1.0, 2.0]
num_scales: 3
backbone:
spinenet_mobile:
stochastic_depth_drop_rate: 0.2
model_id: '49'
se_ratio: 0.2
use_keras_upsampling_2d: true
type: 'spinenet_mobile'
decoder:
type: 'identity'
head:
num_convs: 4
num_filters: 48
use_separable_conv: true
input_size: [384, 384, 3]
max_level: 7
min_level: 3
norm_activation:
activation: 'swish'
norm_epsilon: 0.001
norm_momentum: 0.99
use_sync_bn: true
train_data:
dtype: 'float32'
global_batch_size: 128
is_training: true
parser:
aug_rand_hflip: true
aug_scale_max: 2.0
aug_scale_min: 0.5
validation_data:
dtype: 'float32'
global_batch_size: 8
is_training: false
quantization:
pretrained_original_checkpoint: 'gs://**/coco_spinenet49_mobile_tpu/ckpt-277200'
trainer:
checkpoint_interval: 924
optimizer_config:
learning_rate:
stepwise:
boundaries: [531300, 545160]
values: [0.0016, 0.00016, 0.000016]
type: 'stepwise'
warmup:
linear:
warmup_learning_rate: 0.0000335
warmup_steps: 4000
steps_per_loop: 924
train_steps: 554400
validation_interval: 924
validation_steps: 1250
summary_interval: 924
official/projects/qat/vision/configs/experiments/semantic_segmentation/deeplabv3_mobilenetv2_pascal_qat_gpu.yaml 0 → 100644
View file @ 3aa48ea8
# --experiment_type=mnv2_deeplabv3_pascal_qat
# Use 8 v100 GPUs for training and 4 v100 GPUs for eval.
# mIoU (unquantized fp32): 74.78
runtime:
distribution_strategy: 'mirrored'
mixed_precision_dtype: 'float32'
loss_scale: 'dynamic'
task:
model:
num_classes: 21
input_size: [512, 512, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
output_stride: 16
decoder:
aspp:
dilation_rates: []
level: 4
pool_kernel_size: null
output_tensor: true
type: 'aspp'
head:
feature_fusion: null
num_convs: 0
norm_activation:
activation: relu
norm_epsilon: 0.001
norm_momentum: 0.99
use_sync_bn: true
losses:
l2_weight_decay: 4.0e-07 # 1/100 of original value.
train_data:
output_size: [512, 512]
crop_size: [512, 512]
input_path: 'gs://**/pascal_voc_seg/train_aug*'
is_training: true
global_batch_size: 16
dtype: 'float32'
aug_rand_hflip: true
aug_scale_max: 2.0
aug_scale_min: 0.5
validation_data:
output_size: [512, 512]
input_path: 'gs://**/pascal_voc_seg/val*'
is_training: false
global_batch_size: 16
dtype: 'float32'
drop_remainder: false
resize_eval_groundtruth: false
groundtruth_padded_size: [512, 512]
quantization:
pretrained_original_checkpoint: 'gs://**/deeplabv3_mobilenetv2_pascal_coco_0.21/29808901/best_ckpt/best_ckpt-54'
init_checkpoint: null
trainer:
optimizer_config:
learning_rate:
polynomial:
decay_steps: 13240
initial_learning_rate: 0.00007 # 1/100 of original lr.
power: 0.9
type: polynomial
optimizer:
sgd:
momentum: 0.9
type: sgd
warmup:
linear:
name: linear
warmup_steps: 0 # No warmup
type: linear
best_checkpoint_eval_metric: 'mean_iou'
best_checkpoint_export_subdir: 'best_ckpt'
best_checkpoint_metric_comp: 'higher'
steps_per_loop: 662
summary_interval: 662
train_steps: 13240
validation_interval: 662
validation_steps: 90
checkpoint_interval: 662
official/projects/qat/vision/configs/experiments/semantic_segmentation/deeplabv3_mobilenetv2_pascal_qat_tpu.yaml 0 → 100644
View file @ 3aa48ea8
# --experiment_type=mnv2_deeplabv3_pascal_qat
# Use 4x2 DF for training and eval.
# mIoU (unquantized fp32): 74.69
runtime:
distribution_strategy: 'tpu'
mixed_precision_dtype: 'float32'
task:
model:
num_classes: 21
input_size: [512, 512, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
output_stride: 16
decoder:
aspp:
dilation_rates: []
level: 4
pool_kernel_size: null
output_tensor: true
type: 'aspp'
head:
feature_fusion: null
num_convs: 0
norm_activation:
activation: relu
norm_epsilon: 0.001
norm_momentum: 0.99
use_sync_bn: true
losses:
l2_weight_decay: 4.0e-07 # 1/100 of original value.
train_data:
output_size: [512, 512]
crop_size: [512, 512]
input_path: 'gs://**/pascal_voc_seg/train_aug*'
is_training: true
global_batch_size: 16
dtype: 'float32'
aug_rand_hflip: true
aug_scale_max: 2.0
aug_scale_min: 0.5
validation_data:
output_size: [512, 512]
input_path: 'gs://**/pascal_voc_seg/val*'
is_training: false
global_batch_size: 16
dtype: 'float32'
drop_remainder: false
resize_eval_groundtruth: false
groundtruth_padded_size: [512, 512]
quantization:
pretrained_original_checkpoint: 'gs://**/deeplabv3_mobilenetv2_pascal_coco_0.21/29808901/best_ckpt/best_ckpt-54'
init_checkpoint: null
trainer:
optimizer_config:
learning_rate:
polynomial:
decay_steps: 13240
initial_learning_rate: 0.00007 # 1/100 of original lr.
power: 0.9
type: polynomial
optimizer:
sgd:
momentum: 0.9
type: sgd
warmup:
linear:
name: linear
warmup_steps: 0 # No warmup
type: linear
best_checkpoint_eval_metric: 'mean_iou'
best_checkpoint_export_subdir: 'best_ckpt'
best_checkpoint_metric_comp: 'higher'
steps_per_loop: 662
summary_interval: 662
train_steps: 13240
validation_interval: 662
validation_steps: 90
checkpoint_interval: 662
official/projects/qat/vision/configs/experiments/semantic_segmentation/deeplabv3plus_mobilenetv2_cityscapes_qat_tpu.yaml 0 → 100644
View file @ 3aa48ea8
# --experiment_type=mnv2_deeplabv3plus_cityscapes_qat
# Use 4x2 DF for training and eval.
# mIoU (unquantized fp32): 73.84
runtime:
distribution_strategy: 'tpu'
mixed_precision_dtype: 'float32'
task:
model:
num_classes: 19
input_size: [1024, 2048, 3]
backbone:
type: 'mobilenet'
mobilenet:
model_id: 'MobileNetV2'
output_stride: 16
output_intermediate_endpoints: true
decoder:
aspp:
dilation_rates: []
level: 4
pool_kernel_size: [512, 1024]
output_tensor: true
type: 'aspp'
head:
feature_fusion: 'deeplabv3plus'
low_level: '2/depthwise'
low_level_num_filters: 48
level: 4
num_convs: 2
use_depthwise_convolution: true
norm_activation:
activation: relu
norm_epsilon: 0.001
norm_momentum: 0.99
use_sync_bn: true
losses:
l2_weight_decay: 4.0e-07 # 1/100 of original value.
train_data:
output_size: [1024, 2048]
crop_size: []
input_path: ''
tfds_name: 'cityscapes/semantic_segmentation'
tfds_split: 'train'
is_training: true
global_batch_size: 16
dtype: 'float32'
aug_rand_hflip: true
aug_scale_max: 2.0
aug_scale_min: 0.5
validation_data:
output_size: [1024, 2048]
input_path: ''
tfds_name: 'cityscapes/semantic_segmentation'
tfds_split: 'validation'
is_training: false
global_batch_size: 16
dtype: 'float32'
drop_remainder: false
resize_eval_groundtruth: true
quantization:
pretrained_original_checkpoint: 'gs://**/deeplabv3plus_mobilenetv2_cityscapes/29814723/best_ckpt/best_ckpt-408'
init_checkpoint: null
trainer:
optimizer_config:
learning_rate:
polynomial:
decay_steps: 20000
initial_learning_rate: 0.0001 # 1/100 of original lr.
power: 0.9
type: polynomial
optimizer:
sgd:
momentum: 0.9
type: sgd
warmup:
linear:
name: linear
warmup_learning_rate: 0
warmup_steps: 0 # No warmup
type: linear
steps_per_loop: 185
summary_interval: 185
train_steps: 20000
validation_interval: 185
validation_steps: 31
checkpoint_interval: 185
best_checkpoint_export_subdir: 'best_ckpt'
best_checkpoint_eval_metric: 'mean_iou'
best_checkpoint_metric_comp: 'higher'
official/projects/qat/vision/configs/image_classification.py 0 → 100644
View file @ 3aa48ea8
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Lint as: python3
"""Image classification configuration definition."""
import dataclasses
from typing import Optional
from official.core import config_definitions as cfg
from official.core import exp_factory
from official.projects.qat.vision.configs import common
from official.vision.beta.configs import image_classification
@dataclasses.dataclass
class ImageClassificationTask(image_classification.ImageClassificationTask):
quantization: Optional[common.Quantization] = None
@exp_factory.register_config_factory('resnet_imagenet_qat')
def image_classification_imagenet() -> cfg.ExperimentConfig:
"""Builds an image classification config for the resnet with QAT."""
config = image_classification.image_classification_imagenet()
task = ImageClassificationTask.from_args(
quantization=common.Quantization(), **config.task.as_dict())
config.task = task
return config
@exp_factory.register_config_factory('mobilenet_imagenet_qat')
def image_classification_imagenet_mobilenet() -> cfg.ExperimentConfig:
"""Builds an image classification config for the mobilenetV2 with QAT."""
config = image_classification.image_classification_imagenet_mobilenet()
task = ImageClassificationTask.from_args(
quantization=common.Quantization(), **config.task.as_dict())
config.task = task
return config
official/projects/qat/vision/configs/image_classification_test.py 0 → 100644
View file @ 3aa48ea8
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for image_classification."""
# pylint: disable=unused-import
from absl.testing import parameterized
import tensorflow as tf
from official.core import config_definitions as cfg
from official.core import exp_factory
from official.projects.qat.vision.configs import common
from official.projects.qat.vision.configs import image_classification as qat_exp_cfg
from official.vision import beta
from official.vision.beta.configs import image_classification as exp_cfg
class ImageClassificationConfigTest(tf.test.TestCase, parameterized.TestCase):
@parameterized.parameters(
('resnet_imagenet_qat',),
('mobilenet_imagenet_qat',),
)
def test_image_classification_configs(self, config_name):
config = exp_factory.get_exp_config(config_name)
self.assertIsInstance(config, cfg.ExperimentConfig)
self.assertIsInstance(config.task, qat_exp_cfg.ImageClassificationTask)
self.assertIsInstance(config.task.model,
exp_cfg.ImageClassificationModel)
self.assertIsInstance(config.task.quantization, common.Quantization)
self.assertIsInstance(config.task.train_data, exp_cfg.DataConfig)
config.task.train_data.is_training = None
with self.assertRaisesRegex(KeyError, 'Found inconsistncy between key'):
config.validate()
if __name__ == '__main__':
tf.test.main()
official/projects/qat/vision/configs/retinanet.py 0 → 100644
View file @ 3aa48ea8
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Lint as: python3
"""RetinaNet configuration definition."""
import dataclasses
from typing import Optional
from official.core import config_definitions as cfg
from official.core import exp_factory
from official.projects.qat.vision.configs import common
from official.vision.beta.configs import retinanet
from official.vision.beta.configs.google import backbones
@dataclasses.dataclass
class RetinaNetTask(retinanet.RetinaNetTask):
quantization: Optional[common.Quantization] = None
@exp_factory.register_config_factory('retinanet_spinenet_mobile_coco_qat')
def retinanet_spinenet_mobile_coco() -> cfg.ExperimentConfig:
"""Generates a config for COCO OD RetinaNet for mobile with QAT."""
config = retinanet.retinanet_spinenet_mobile_coco()
task = RetinaNetTask.from_args(
quantization=common.Quantization(), **config.task.as_dict())
task.model.backbone = backbones.Backbone(
type='spinenet_mobile',
spinenet_mobile=backbones.SpineNetMobile(
model_id='49',
stochastic_depth_drop_rate=0.2,
min_level=3,
max_level=7,
use_keras_upsampling_2d=True))
config.task = task
return config
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment