AlexNet was proposed in 2012, one of the most influential neural networks. It got big success in ImageNet Dataset recognition than other models.
[Paper](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf): Krizhevsky A, Sutskever I, Hinton G E. ImageNet Classification with Deep ConvolutionalNeural Networks. *Advances In Neural Information Processing Systems*. 2012.
# [Model Architecture](#contents)
## [Model Architecture](#contents)
AlexNet composition consists of 5 convolutional layers and 3 fully connected layers. Multiple convolutional kernels can extract interesting features in images and get more accurate classification.
# [Dataset](#contents)
## [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
[论文](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-concumulational-neural-networks.pdf)： Krizhevsky A, Sutskever I, Hinton G E. ImageNet Classification with Deep ConvolutionalNeural Networks. *Advances In Neural Information Processing Systems*. 2012.
LeNet was proposed in 1998, a typical convolutional neural network. It was used for digit recognition and got big success.
LeNet was proposed in 1998, a typical convolutional neural network. It was used for digit recognition and got big success.
[Paper](https://ieeexplore.ieee.org/document/726791): Y.Lecun, L.Bottou, Y.Bengio, P.Haffner. Gradient-Based Learning Applied to Document Recognition. *Proceedings of the IEEE*. 1998.
# [Model Architecture](#contents)
## [Model Architecture](#contents)
LeNet is very simple, which contains 5 layers. The layer composition consists of 2 convolutional layers and 3 fully connected layers.
# [Dataset](#contents)
## [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
[论文](https://ieeexplore.ieee.org/document/726791)： Y.Lecun, L.Bottou, Y.Bengio, P.Haffner.Gradient-Based Learning Applied to Document Recognition.*Proceedings of the IEEE*.1998.
[论文](https://arxiv.org/abs/1409.1556): Simonyan K, zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition[J]. arXiv preprint arXiv:1409.1556, 2014.
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
# [WarpCTC Description](#contents)
## [WarpCTC Description](#contents)
This is an example of training WarpCTC with self-generated captcha image dataset in MindSpore.
# [Model Architecture](#content)
## [Model Architecture](#content)
WarpCTC is a two-layer stacked LSTM appending with one-layer FC neural network. See src/warpctc.py for details.
# [Dataset](#content)
## [Dataset](#content)
The dataset is self-generated using a third-party library called [captcha](https://github.com/lepture/captcha), which can randomly generate digits from 0 to 9 in image. In this network, we set the length of digits varying from 1 to 4.
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware（Ascend/GPU）
- Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. You will be able to have access to related resources once approved.
- Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. You will be able to have access to related resources once approved.
Run the script `scripts/run_process_data.sh` to generate a dataset. By default, the shell script will generate 10000 test images and 50000 train images separately.
```
```bash
$ cd scripts
$ sh run_process_data.sh
# after execution, you will find the dataset like the follows:
.
└─warpctc
@ -67,38 +66,41 @@ The dataset is self-generated using a third-party library called [captcha](https
- After the dataset is prepared, you may start running the training or the evaluation scripts as follows:
Parameters for both training and evaluation can be set in config.py.
```
"max_captcha_digits": 4, # max number of digits in each
```bash
"max_captcha_digits": 4, # max number of digits in each
"captcha_width": 160, # width of captcha images
"captcha_height": 64, # height of capthca images
"batch_size": 64, # batch size of input tensor
@ -158,45 +161,50 @@ Parameters for both training and evaluation can be set in config.py.
```
## [Dataset Preparation](#contents)
- You may refer to "Generate dataset" in [Quick Start](#quick-start) to automatically generate a dataset, or you may choose to generate a captcha dataset by yourself.
## [Training Process](#contents)
### [Training Process](#contents)
- Set options in `config.py`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/tutorial/training/zh-CN/master/use/data_preparation.html) for more information about dataset.
### [Training](#contents)
#### [Training](#contents)
- Run `run_standalone_train.sh` for non-distributed training of WarpCTC model, either on Ascend or on GPU.
You only look once (YOLO) is a state-of-the-art, real-time object detection system. YOLOv3 is extremely fast and accurate.
You only look once (YOLO) is a state-of-the-art, real-time object detection system. YOLOv3 is extremely fast and accurate.
Prior detection systems repurpose classifiers or localizers to perform detection. They apply the model to an image at multiple locations and scales. High scoring regions of the image are considered detections.
YOLOv3 use a totally different approach. It apply a single neural network to the full image. This network divides the image into regions and predicts bounding boxes and probabilities for each region. These bounding boxes are weighted by the predicted probabilities.
YOLOv3 use a totally different approach. It apply a single neural network to the full image. This network divides the image into regions and predicts bounding boxes and probabilities for each region. These bounding boxes are weighted by the predicted probabilities.
YOLOv3 uses a few tricks to improve training and increase performance, including: multi-scale predictions, a better backbone classifier, and more. The full details are in the paper!
@ -35,43 +34,39 @@ In order to reduce the size of the weight and improve the low-bit computing perf
[Paper](https://pjreddie.com/media/files/papers/YOLOv3.pdf): YOLOv3: An Incremental Improvement. Joseph Redmon, Ali Farhadi,
University of Washington
# [Model Architecture](#contents)
## [Model Architecture](#contents)
YOLOv3 use DarkNet53 for performing feature extraction, which is a hybrid approach between the network used in YOLOv2, Darknet-19, and that newfangled residual network stuff. DarkNet53 uses successive 3 × 3 and 1 × 1 convolutional layers and has some shortcut connections as well and is significantly larger. It has 53 convolutional layers.
## [Dataset](#contents)
# [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
- Note：Data will be processed in yolo_dataset.py, and unzip files before uses it.
- Note：Data will be processed in yolo_dataset.py, and unzip files before uses it.
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
After installing MindSpore via the official website, you can start training and evaluation in Ascend as follows:
After installing MindSpore via the official website, you can start training and evaluation in Ascend as follows:
```
# The yolov3_darknet53_noquant.ckpt in the follow script is got from yolov3-darknet53 training like paper.
```bash
# The yolov3_darknet53_noquant.ckpt in the follow script is got from yolov3-darknet53 training like paper.
# The parameter of resume_yolov3 is necessary.
# The parameter of training_shape define image shape for network, default is "".
# It means use 10 kinds of shape as input shape, or it can be set some kind of shape.
@ -103,17 +98,16 @@ python eval.py \
sh run_eval.sh dataset/coco2014/ checkpoint/yolov3_quant.ckpt 0
```
## [Script Description](#contents)
# [Script Description](#contents)
## [Script and Sample Code](#contents)
### [Script and Sample Code](#contents)
```
```bash
.
└─yolov3_darknet53_quant
└─yolov3_darknet53_quant
├─README.md
├─mindspore_hub_conf.md # config for mindspore hub
├─scripts
├─scripts
├─run_standalone_train.sh # launch standalone training(1p) in ascend
├─run_distribute_train.sh # launch distributed training(8p) in ascend
└─run_eval.sh # launch evaluating in ascend
@ -134,10 +128,9 @@ sh run_eval.sh dataset/coco2014/ checkpoint/yolov3_quant.ckpt 0
└─train.py # train net
```
### [Script Parameters](#contents)
## [Script Parameters](#contents)
```
```bash
Major parameters in train.py as follow.
optional arguments:
@ -194,21 +187,19 @@ optional arguments:
Resize rate for multi-scale training. Default: None
```
### [Training Process](#contents)
#### Training on Ascend
## [Training Process](#contents)
##### Distributed Training
### Training on Ascend
### Distributed Training
```
```bash
sh run_distribute_train.sh dataset/coco2014 yolov3_darknet53_noquant.ckpt rank_table_8p.json
```
The above shell script will run distribute training in the background. You can view the results through the file `train_parallel[X]/log.txt`. The loss value will be achieved as follows:
@ -250,7 +240,7 @@ sh run_eval.sh dataset/coco2014/ checkpoint/0-130_83330.ckpt 0
The above python command will run in the background. You can view the results through the file "log.txt". The mAP of the test dataset will be as follows:
```
```bash
# log.txt
=============coco eval reulst=========
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.310
Bayesian Graph Collaborative Filtering(BGCF) was proposed in 2020 by Sun J, Guo W, Zhang D et al. By naturally incorporating the
uncertainty in the user-item interaction graph shows excellent performance on Amazon recommendation dataset.This is an example of
@ -29,12 +31,12 @@ training of BGCF with Amazon-Beauty dataset in MindSpore. More importantly, this
[Paper](https://dl.acm.org/doi/pdf/10.1145/3394486.3403254): Sun J, Guo W, Zhang D, et al. A Framework for Recommending Accurate and Diverse Items Using Bayesian Graph Convolutional Neural Networks[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 2030-2039.
# [Model Architecture](#contents)
## [Model Architecture](#contents)
Specially, BGCF contains two main modules. The first is sampling, which produce sample graphs based in node copying. Another module
aggregate the neighbors sampling from nodes consisting of mean aggregator and attention aggregator.
# [Dataset](#contents)
## [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
@ -69,13 +71,13 @@ Note that you can run the scripts based on the dataset mentioned in original pap
sh run_process_data_ascend.sh [SRC_PATH]
```
# [Features](#contents)
## [Features](#contents)
## Mixed Precision
### Mixed Precision
To ultilize the strong computation power of Ascend chip, and accelerate the training process, the mixed training method is used. MindSpore is able to cope with FP32 inputs and FP16 operators. In BGCF example, the model is set to FP16 mode except for the loss calculation part.
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware (Ascend/GPU)
- Framework
@ -84,7 +86,7 @@ To ultilize the strong computation power of Ascend chip, and accelerate the trai
[论文](https://dl.acm.org/doi/pdf/10.1145/3394486.3403254): Sun J, Guo W, Zhang D, et al.A Framework for Recommending Accurate and Diverse Items Using Bayesian Graph Convolutional Neural Networks[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining.2020: 2030-2039.
Graph Attention Networks(GAT) was proposed in 2017 by Petar Veličković et al. By leveraging masked self-attentional layers to address shortcomings of prior graph based method, GAT achieved or matched state of the art performance on both transductive datasets like Cora and inductive dataset like PPI. This is an example of training GAT with Cora dataset in MindSpore.
[Paper](https://arxiv.org/abs/1710.10903): Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., & Bengio, Y. (2017). Graph attention networks. arXiv preprint arXiv:1710.10903.
# [Model architecture](#contents)
## [Model architecture](#contents)
Note that according to whether this attention layer is the output layer of the network or not, the node update function can be concatenate or average.
# [Dataset](#contents)
## [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
- Dataset size:
Statistics of dataset used are summerized as below:
Statistics of dataset used are summarized as below:
@ -46,9 +52,9 @@ Note that you can run the scripts based on the dataset mentioned in original pap
| # Test Nodes | 1000 | 1000 |
- Data Preparation
- Place the dataset to any path you want, the folder should include files as follows(we use Cora dataset as an example):
```
- Place the dataset to any path you want, the folder should include files as follows(we use Cora dataset as an example):
```bash
.
└─data
├─ind.cora.allx
@ -61,58 +67,60 @@ Note that you can run the scripts based on the dataset mentioned in original pap
└─ind.cora.y
```
- Generate dataset in mindrecord format for cora or citeseer.
```buildoutcfg
cd ./scripts
# SRC_PATH is the dataset file path you downloaded, DATASET_NAME is cora or citeseer
sh run_process_data_ascend.sh [SRC_PATH] [DATASET_NAME]
```
- Generate dataset in mindrecord format for cora or citeseer.
- Launch
```
#Generate dataset in mindrecord format for cora
./run_process_data_ascend.sh ./data cora
#Generate dataset in mindrecord format for citeseer
./run_process_data_ascend.sh ./data citeseer
```
```buildoutcfg
cd ./scripts
# SRC_PATH is the dataset file path you downloaded, DATASET_NAME is cora or citeseer
sh run_process_data_ascend.sh [SRC_PATH] [DATASET_NAME]
```
- Launch
```bash
#Generate dataset in mindrecord format for cora
./run_process_data_ascend.sh ./data cora
#Generate dataset in mindrecord format for citeseer
./run_process_data_ascend.sh ./data citeseer
```
# [Features](#contents)
## [Features](#contents)
## Mixed Precision
### Mixed Precision
To ultilize the strong computation power of Ascend chip, and accelerate the training process, the mixed training method is used. MindSpore is able to cope with FP32 inputs and FP16 operators. In GAT example, the model is set to FP16 mode except for the loss calculation part.
GAT model contains lots of dropout operations, if you want to disable dropout, set the attn_dropout and feature_dropout to 0 in src/config.py. Note that this operation will cause the accuracy drop to approximately 80%.
# [ModelZoo Homepage](#contents)
## [ModelZoo Homepage](#contents)
Please check the official [homepage](http://gitee.com/mindspore/mindspore/tree/master/model_zoo).
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
# [GCN Description](#contents)
## [GCN Description](#contents)
GCN(Graph Convolutional Networks) was proposed in 2016 and designed to do semi-supervised learning on graph-structured data. A scalable approach based on an efficient variant of convolutional neural networks which operate directly on graphs was presented. The model scales linearly in the number of graph edges and learns hidden layer representations that encode both local graph structure and features of nodes.
[Paper](https://arxiv.org/abs/1609.02907): Thomas N. Kipf, Max Welling. 2016. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR 2016.
## [Model Architecture](#contents)
# [Model Architecture](#contents)
GCN contains two graph convolution layers. Each layer takes nodes features and adjacency matrix as input, nodes' features are then updated by aggregating neighbours' features.
GCN contains two graph convolution layers. Each layer takes nodes features and adjacency matrix as input, nodes' features are then updated by aggregating neighbours' features.
## [Dataset](#contents)
# [Dataset](#contents)
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
| Dataset | Type | Nodes | Edges | Classes | Features | Label rate |
@ -35,29 +33,25 @@ Note that you can run the scripts based on the dataset mentioned in original pap
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Download the dataset Cora or Citeseer provided by /kimiyoung/planetoid from github.
- Place the dataset to any path you want, the folder should include files as follows(we use Cora dataset as an example):
```
```bash
.
└─data
├─ind.cora.allx
@ -71,30 +65,33 @@ Note that you can run the scripts based on the dataset mentioned in original pap
```
- Generate dataset in mindrecord format for cora or citeseer.
####Usage
### Usage
```buildoutcfg
cd ./scripts
# SRC_PATH is the dataset file path you downloaded, DATASET_NAME is cora or citeseer
sh run_process_data.sh [SRC_PATH] [DATASET_NAME]
```
####Launch
```
### Launch
```bash
#Generate dataset in mindrecord format for cora
sh run_process_data.sh ./data cora
#Generate dataset in mindrecord format for citeseer
sh run_process_data.sh ./data citeseer
```
# [Script Description](#contents)
### [Script Description](#contents)
### [Script and Sample Code](#contents)
## [Script and Sample Code](#contents)
```shell
.
└─gcn
└─gcn
├─README.md
├─scripts
├─scripts
| ├─run_process_data.sh # Generate dataset in mindrecord format
| └─run_train.sh # Launch training, now only Ascend backend is supported.
|
@ -106,12 +103,12 @@ sh run_process_data.sh ./data citeseer
|
└─train.py # Train net, evaluation is performed after every training epoch. After the verification result converges, the training stops, then testing is performed.
```
## [Script Parameters](#contents)
Parameters for training can be set in config.py.
```
```bash
"learning_rate": 0.01, # Learning rate
"epochs": 200, # Epoch sizes for training
"hidden1": 16, # Hidden size for the first graph convolution layer
@ -120,27 +117,26 @@ Parameters for training can be set in config.py.
"early_stopping": 10, # Tolerance for early stopping
```
## [Training, Evaluation, Test Process](#contents)
### [Training, Evaluation, Test Process](#contents)
#### Usage
```
```bash
# run train with cora or citeseer dataset, DATASET_NAME is cora or citeseer
sh run_train.sh [DATASET_NAME]
```
#### Launch
```bash
sh run_train.sh cora
```
#### Result
Training result will be stored in the scripts path, whose folder name begins with "train". You can find the result like the followings in log.
- Seed is set in train.py according to input argument --seed.
- Dropout operations.
Some seeds have already been set in train.py to avoid the randomness of weight initialization. If you want to disable dropout, please set the corresponding dropout_prob parameter to 0 in src/config.py.
# [ModelZoo Homepage](#contents)
## [ModelZoo Homepage](#contents)
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).
[论文](https://arxiv.org/abs/1609.02907): Thomas N. Kipf, Max Welling.2016.Semi-Supervised Classification with Graph Convolutional Networks.In ICLR 2016.
The number of Ascend accelerators can be automatically allocated based on the device_num set in hccl config file, You don not need to specify that.
## How to use
## how to use
For example, if we want to generate the launch command of the distributed training of Bert model on Ascend accelerators, we can run the following command in `/bert/` dir:
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
## [Wide&Deep Description](#contents)
# [Wide&Deep Description](#contents)
Wide&Deep model is a classical model in Recommendation and Click Prediction area. This is an implementation of Wide&Deep as described in the [Wide & Deep Learning for Recommender System](https://arxiv.org/pdf/1606.07792.pdf) paper.
# [Model Architecture](#contents)
Wide&Deep model jointly trained wide linear models and deep neural network, which combined the benefits of memorization and generalization for recommender systems.
## [Model Architecture](#contents)
Wide&Deep model jointly trained wide linear models and deep neural network, which combined the benefits of memorization and generalization for recommender systems.
# [Dataset](#contents)
## [Dataset](#contents)
- [1] A dataset used in Click Prediction
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware（Ascend or GPU）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
All executable scripts can be found in [here](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/recommend/wide_and_deep_multitable/script)
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
# [GhostNet Description](#contents)
## [GhostNet Description](#contents)
The GhostNet architecture is based on an Ghost module structure which generate more features from cheap operations. Based on a set of intrinsic feature maps, a series of cheap operations are applied to generate many ghost feature maps that could fully reveal information underlying intrinsic features.
[Paper](https://openaccess.thecvf.com/content_CVPR_2020/papers/Han_GhostNet_More_Features_From_Cheap_Operations_CVPR_2020_paper.pdf): Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, Chang Xu. GhostNet: More Features from Cheap Operations. CVPR 2020.
# [Model architecture](#contents)
## [Model architecture](#contents)
The overall network architecture of GhostNet is show below:
- Dataset size: 7049 colorful images in 1000 classes
- Train: 3680 images
- Test: 3369 images
- Train: 3680 images
- Test: 3369 images
- Data format: RGB images.
- Note: Data will be processed in src/dataset.py
- Note: Data will be processed in src/dataset.py
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware（Ascend/GPU)
- Prepare hardware environment with Ascend or GPU. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Prepare hardware environment with Ascend or GPU. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage)
# [GhostNet Description](#contents)
## [GhostNet Description](#contents)
The GhostNet architecture is based on an Ghost module structure which generate more features from cheap operations. Based on a set of intrinsic feature maps, a series of cheap operations are applied to generate many ghost feature maps that could fully reveal information underlying intrinsic features.
[Paper](https://openaccess.thecvf.com/content_CVPR_2020/papers/Han_GhostNet_More_Features_From_Cheap_Operations_CVPR_2020_paper.pdf): Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, Chang Xu. GhostNet: More Features from Cheap Operations. CVPR 2020.
# [Quantization Description](#contents)
## [Quantization Description](#contents)
Quantization refers to techniques for performing computations and storing tensors at lower bitwidths than floating point precision. For 8bit quantization, we quantize the weights into [-128,127] and the activations into [0,255]. We finetune the model a few epochs after post-quantization to achieve better performance.
# [Model architecture](#contents)
## [Model architecture](#contents)
The overall network architecture of GhostNet is show below:
- Dataset size: 7049 colorful images in 1000 classes
- Train: 3680 images
- Test: 3369 images
- Train: 3680 images
- Test: 3369 images
- Data format: RGB images.
- Note: Data will be processed in src/dataset.py
- Note: Data will be processed in src/dataset.py
# [Environment Requirements](#contents)
## [Environment Requirements](#contents)
- Hardware（Ascend/GPU）
- Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend, please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
The number of Ascend accelerators can be automatically allocated based on the device_num set in hccl config file, You don not need to specify that.
The number of Ascend accelerators can be automatically allocated based on the device_num set in hccl config file, You don not need to specify that.
## how to use
For example, if we want to generate the launch command of the distributed training of Bert model on Ascend accelerators, we can run the following command in `/bert/` dir:
- [Create MindDataset By MindRecord](#create-minddataset-by-mindrecord)
<!-- /TOC -->
## What does the example do
This example is used to read data from Caltech-UCSD Birds-200-2011 dataset and generate mindrecord. It just transfers the Caltech-UCSD Birds-200-2011 dataset to mindrecord without any data preprocessing. You can modify the example or follow the example to implement your own example.
1. run.sh: generate MindRecord entry script.
1. run.sh: generate MindRecord entry script.
- gen_mindrecord.py : read the Caltech-UCSD Birds-200-2011 data and transfer it to mindrecord.
2. run_read.py: create MindDataset by MindRecord entry script.
2. run_read.py: create MindDataset by MindRecord entry script.
- create_dataset.py: use MindDataset to read MindRecord to generate dataset.
## How to use the example to generate MindRecord
@ -32,25 +31,30 @@ Download Caltech-UCSD Birds-200-2011 dataset, transfer it to mindrecord, use Min
> **2) -> Download -> Segmentations**
2. Unzip the training data to dir example/nlp_to_mindrecord/Caltech-UCSD-Birds-200-2011/data.
```
```bash
tar -zxvf CUB_200_2011.tgz -C {your-mindspore}/example/cv_to_mindrecord/Caltech-UCSD-Birds-200-2011/data/
tar -zxvf segmentations.tgz -C {your-mindspore}/example/cv_to_mindrecord/Caltech-UCSD-Birds-200-2011/data/
```
- The unzip should like this:
```
```
The unzip should like this:
```bash
$ ls {your-mindspore}/example/cv_to_mindrecord/Caltech-UCSD-Birds-200-2011/data/
@ -72,10 +76,11 @@ Download Caltech-UCSD Birds-200-2011 dataset, transfer it to mindrecord, use Min
[INFO] MD(11253,python):2020-05-20-16:22:21.964.034 [mindspore/ccsrc/mindrecord/io/shard_index_generator.cc:549] ExecuteTransaction] Insert 11788 rows to index db.
[INFO] MD(11253,python):2020-05-20-16:22:21.978.087 [mindspore/ccsrc/mindrecord/io/shard_index_generator.cc:620] DatabaseWriter] Generate index db for shard: 0 successfully.
[INFO] ME(11253:139923799271232,MainProcess):2020-05-20-16:22:21.979.634 [mindspore/mindrecord/filewriter.py:313] The list of mindrecord files created are: ['output/CUB_200_2011.mindrecord'], and the list of index files are: ['output/CUB_200_2011.mindrecord.db']
@ -83,11 +88,13 @@ Download Caltech-UCSD Birds-200-2011 dataset, transfer it to mindrecord, use Min
### Create MindDataset By MindRecord
1. Run the run_read.sh script.
```bash
bash run_read.sh
```
2. Output like this:
```
[INFO] MD(12469,python):2020-05-20-16:26:38.308.797 [mindspore/ccsrc/dataset/util/task.cc:31] operator()] Op launched, OperatorId:0 Thread ID 139702598620928 Started.
[INFO] MD(12469,python):2020-05-20-16:26:38.322.433 [mindspore/ccsrc/mindrecord/io/shard_reader.cc:343] ReadAllRowsInShard] Get 11788 records from shard 0 index.
@ -123,6 +130,7 @@ Download Caltech-UCSD Birds-200-2011 dataset, transfer it to mindrecord, use Min
>> total rows: 11788
[INFO] MD(12469,python):2020-05-20-16:26:49.582.298 [mindspore/ccsrc/dataset/util/task.cc:128] Join] Watchdog Thread ID 139702607013632 Stopped.
```
- bbox : coordinate value of the bounding box in the picture.
- image: the image bytes which is from like "data/CUB_200_2011/images/001.Black_footed_Albatross/Black_Footed_Albatross_0001_796111.jpg".
- image_filename: the image name which is like "Black_Footed_Albatross_0001_796111.jpg"
Some files were not shown because too many files have changed in this diff
Show More
oacjiewen
4 years ago
committed by