mindspore/model_zoo/official/cv/deeplabv3

• Contents
• DeeplabV3 Description
• Model architecture
• Dataset
• Features
• • Mixed Precision(Ascend)
• Environment Requirements
• Script description
• • Script and sample code
  • Script Parameters
  • Training process
  • • Usage
    • Launch
    • Result
  • Eval process
  • • Usage
    • Launch
    • Result
• Model description
• • Performance
  • • Training Performance
    • • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• DeeplabV3 Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Script Description
  • Script and Sample Code
  • Training Process
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
    • Training Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

DeeplabV3 Description

DeepLabv3 is a semantic segmentation architecture that improves upon DeepLabv2 with several modifications.To handle the problem of segmenting objects at multiple scales, modules are designed which employ atrous convolution in cascade or in parallel to capture multi-scale context by adopting multiple atrous rates.

Paper Chen L C , Papandreou G , Schroff F , et al. Rethinking Atrous Convolution for Semantic Image Segmentation[J]. 2017.

Model architecture

The overall network architecture of DeepLabv3 is show below:

Link

Dataset

Dataset used: VOC2012

20 classes. The train/val data has 11,530 images containing 27,450 ROI annotated objects and 6,929 segmentations. And we need to remove color map from annotation.

Features

Mixed Precision(Ascend)

The mixed precision training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware.

For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’.

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend. If you want to try Ascend, please send the application form to ascend@huawei.com. Once approved, you can get the resources.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore tutorials
  • MindSpore API

Script description

Script and sample code

.
└─deeplabv3   
    ├──README.md   
	├──eval.py
	├──train.py
	├──scripts
	│  ├──run_distribute_train.sh  # launch distributed training with ascend platform(8p)
	│  ├──run_eval.sh				# launch evaluating with ascend platform
	│  ├──run_standalone_train.sh  # launch standalone training with ascend platform(1p)
	├──src
	│	├──config.py				# parameter configuration
	│	├──deeplabv3.py				# network definition
	│	├──ei_dataset.py			# data preprocessing for EI
	│	├──losses.py				# customized loss function
	│	├──md_dataset.py			# data preprocessing
	│	├──miou_precision.py		# miou metrics
	│	├──__init__.py
	│	│
	│	├──backbone
	│	│  ├──resnet_deeplab.py	# backbone network definition
	│	│  ├──__init__.py
	│	│
	│	└──utils
    │      ├──adapter.py				# adapter of dataset
    │      ├──custom_transforms.py    # random process dataset
    │      ├──file_io.py              # file operation module
    │      ├──__init__.py                    

Script Parameters

Major parameters in train.py and config.py are:   
	learning_rate                   Learning rate, default is 0.0014.
    weight_decay                	Weight decay, default is 5e-5.
    momentum                    	Momentum, default is 0.97.
    crop_size                       Image crop size [height, width] during training, default is 513.
    eval_scales                     The scales to resize images for evaluation, default is [0.5, 0.75, 1.0, 1.25, 1.5, 1.75].
	output_stride					The ratio of input to output spatial resolution, default is 16.
	ignore_label					Ignore label value,	default is 255.
	seg_num_classes					Number of semantic classes, including the background class. 
									foreground classes + 1 background class in the PASCAL VOC 2012 dataset, default is 21.
	fine_tune_batch_norm			Fine tune the batch norm parameters or not, default is False.
	atrous_rates					Atrous rates for atrous spatial pyramid pooling, default is None.
	decoder_output_stride			The ratio of input to output spatial resolution when employing decoder
									to refine segmentation results, default is None.
	image_pyramid					Input scales for multi-scale feature extraction, default is None.
	epoch_size						Epoch size, default is 6.
    batch_size                      Batch size of input dataset: N, default is 2.
	enable_save_ckpt				Enable save checkpoint, default is true.
	save_checkpoint_steps			Save checkpoint steps, default is 1000.
	save_checkpoint_num				Save checkpoint numbers, default is 1.

Training process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

	sh scripts/run_distribute_train.sh RANK_TABLE_FILE DATA_PATH (CKPT_PATH)

Notes:
RANK_TABLE_FILE can refer to Link , and the device_ip can be got as https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools.

Launch

# training example
  python:
      python train.py --dataset_url DATA_PATH 

  shell:
      sh scripts/run_standalone_train.sh DEVICE_ID DATA_PATH (CKPT_PATH)

Notes:
If you are running a fine-tuning or evaluation task, prepare the corresponding checkpoint file.

Result

Training result(8p) will be stored in the example path. Checkpoints will be stored at . /train_parallel0/ by default, and training log will be redirected to ./train_parallel0/log.txt like followings.

epoch: 1 step: 732, loss is 0.11594
Epoch time: 78748.379, per step time: 107.378
epoch: 2 step: 732, loss is 0.092868
Epoch time: 160917.911, per step time: 36.631

Eval process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

	sh scripts/run_eval.sh DEVICE_ID DATA_PATH PRETRAINED_CKPT_PATH

Launch

# eval example
  python:
      python eval.py --device_id DEVICE_ID --dataset_url DATA_DIR --checkpoint_url PATH_CHECKPOINT

  shell:
      sh scripts/run_eval.sh DEVICE_ID DATA_PATH PRETRAINED_CKPT_PATH

checkpoint can be produced in training process.

Result

Evaluation result will be stored in the example path, you can find result like the followings in eval.log.

mIoU = 0.65049

Model description

Performance

Training Performance

Parameters	DeeplabV3
Model Version	V1
Resource	Ascend 910, cpu:2.60GHz 56cores, memory:314G
Uploaded Date	08/24/2020(month/day/year)
MindSpore Version	0.6.0-beta
Dataset	voc2012/train
Batch_size	2
Optimizer	Momentum
Loss Function	SoftmaxCrossEntropy
Outputs	probability
Loss	0.98
Accuracy	mIoU:65%
Total time	5mins
Params (M)	94M
Checkpoint for Fine tuning	100M
Scripts	deeplabv3 script

Inference Performance

Parameters	DeeplabV3
Model Version	V1
Resource	Ascend 910, cpu:2.60GHz 56cores, memory:314G
Uploaded Date	08/24/2020 (month/day/year)
MindSpore Version	0.6.0-beta
Dataset	voc2012/val
Batch_size	2
Outputs	probability
Accuracy	mIoU:65%
Total time	10mins
Model for inference	97M (.GEIR file)

Description of Random Situation

We use random in custom_transforms.py for data preprocessing.

ModelZoo Homepage

Please check the official homepage.