- [Script and Sample Code](#script-and-sample-code)
- [Script Parameters](#script-parameters)
- [Training Process](#training-process)
- [Training](#training)
- [Running on Ascend](#running-on-ascend)
- [Distributed Training](#distributed-training)
- [Running on Ascend](#running-on-ascend-1)
- [Evaluation Process](#evaluation-process)
- [Running on Ascend](#running-on-ascend-2)
- [Model Description](#model-description)
- [Performance](#performance)
- [Accuracy](#accuracy)
- [DPN92 (Training)](#dpn92-training)
- [Efficiency](#efficiency)
- [DPN92](#dpn92)
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
# [Description](#contents)
Dual Path Network (DPN) is a convolution-based neural network for the task of image classification. It combines the advantage of both ResNeXt and DenseNet to get higher accuracy. More detail about this model can be found in:
This repository contains a Mindspore implementation of DPNs based upon cypw's original MXNet implementation (<https://github.com/cypw/DPNs>). The training and validating scripts are also included, and the validation results with cypw’s pretrained weights are shown in the Results section.
# [Model Architecture](#contents)
The overall network architecture of DPN is show below:
[Link](https://arxiv.org/pdf/1707.01629.pdf)
# [Dataset](#contents)
All the models in this repository are trained and validated on ImageNet-1K. The models can achieve the [results](#model-description) with the configurations of the dataset preprocessing as follow:
- For the training dataset:
- Range (min, max) of the respective size of the original size to be cropped is (0.08, 1.0)
- Range (min, max) of aspect ratio to be cropped is (0.75, 1.333)
- The size of input images is reshaped to (width = 224, height = 224)
- Probability of random horizontal flip is 50%
- In normalization, the mean is (255\*0.485, 255\*0.456, 255\*0.406) and the standard deviation is (255\*0.229, 255\*0.224, 255\*0.225)
- For the evaluation dataset:
- Input size of images is 224\*224 (Resize to 256\*256 then crops images at the center)
- In normalization, the mean is (255\*0.485, 255\*0.456, 255\*0.406) and the standard deviation is (255\*0.229, 255\*0.224, 255\*0.225)
The [mixed precision](https://www.mindspore.cn/tutorial/training/en/master/advanced_use/enable_mixed_precision.html) 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’.
The DPN models use ImageNet-1K dataset to train and validate in this repository. Download the dataset from [ImageNet.org](http://image-net.org/download). You can place them anywhere and tell the scripts where they are when running.
## [Running](#contents)
To train the DPNs, run the shell script `scripts/train_standalone.sh` with the format below:
```shell
sh scripts/train_standalone.sh [device_id] [dataset_dir] [ckpt_path_to_save] [eval_each_epoch] [pretrained_ckpt(optional)]
```
To validate the DPNs, run the shell script `scripts/eval.sh` with the format below:
```shell
sh scripts/eval.sh [device_id] [dataset_dir] [pretrained_ckpt]
```
# [Script Description](#contents)
## [Script and Sample Code](#contents)
The structure of the files in this repository is shown below.
If eval_each_epoch is 1, it will evaluate after each epoch and save the parameters with the max accuracy. But in this case, the time of one epoch will be longer.
Save the maximum accuracy checkpoint,the accuracy is 0.2629158669225848
...
```
The model checkpoint will be saved into `[ckpt_path_to_save]`.
### [Distributed Training](#contents)
#### Running on Ascend
Run `scripts/train_distributed.sh` to train the model distributed. The usage of the script is:
```text
sh scripts/train_distributed.sh [rank_table] [dataset_dir] [ckpt_path_to_save] [rank_size] [eval_each_epoch] [pretrained_ckpt(optional)]
```
For example, you can run the shell command below to launch the training procedure.
```shell
sh scripts/train_distributed.sh /home/rank_table.json /data/dataset/imagenet/ ../scripts 8 0 ../pretrain/dpn92.ckpt
```
The above shell script will run distribute training in the background. You can view the results through the file `train_parallel[X]/log.txt` as follows:
The above shell script will run evaluation in the background. You can view the results through the file `eval_log.txt`. The result will be achieved as follows:
The evaluation of model performance is divided into two parts: accuracy and efficiency. The part of accuracy shows the accuracy of the model in classifying images on ImageNet-1K dataset, and it can be evaluated by top-k measure. The part of efficiency reveals the time cost by model training on ImageNet-1K.
All results are validated at image size of 224x224. The dataset preprocessing and training configurations are shown in [Dataset](#dataset) section.
