Before FasterRcnn, the target detection networks rely on the region proposal algorithm to assume the location of targets, such as SPPnet and Fast R-CNN. Progress has reduced the running time of these detection networks, but it also reveals that the calculation of the region proposal is a bottleneck.
FasterRcnn proposed that convolution feature maps based on region detectors (such as Fast R-CNN) can also be used to generate region proposals. At the top of these convolution features, a Region Proposal Network (RPN) is constructed by adding some additional convolution layers (which share the convolution characteristics of the entire image with the detection network, thus making it possible to make regions almost costlessProposal), outputting both region bounds and objectness score for each location.Therefore, RPN is a full convolutional network (FCN), which can be trained end-to-end, generate high-quality region proposals, and then fed into Fast R-CNN for detection.
[Paper](https://arxiv.org/abs/1506.01497): Ren S , He K , Girshick R , et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39(6).
#Model Architecture
FasterRcnn is a two-stage target detection network,This network uses a region proposal network (RPN), which can share the convolution features of the whole image with the detection network, so that the calculation of region proposal is almost cost free. The whole network further combines RPN and FastRcnn into a network by sharing the convolution features.
- We use coco2017 as training dataset in this example by default, and you can also use your own datasets.
- We use COCO2017 as training dataset in this example by default, and you can also use your own datasets.
1. If coco dataset is used. **Select dataset to coco when run script.**
Install Cython and pycocotool, and you can also install mmcv to process data.
...
...
@@ -20,7 +58,7 @@ FasterRcnn is a two-stage target detection network,This network uses a region pr
pip install pycocotools
pip install mmcv
pip install mmcv==0.2.14
```
And change the COCO_ROOT and other settings you need in `config.py`. The directory structure is as follows:
...
...
@@ -45,58 +83,71 @@ FasterRcnn is a two-stage target detection network,This network uses a region pr
Each row is an image annotation which split by space, the first column is a relative path of image, the others are box and class infomations of the format [xmin,ymin,xmax,ymax,class]. We read image from an image path joined by the `IMAGE_DIR`(dataset directory) and the relative path in `ANNO_PATH`(the TXT file path), `IMAGE_DIR` and `ANNO_PATH` are setting in `config.py`.
# Quick Start
After installing MindSpore via the official website, you can start training and evaluation as follows:
```
# standalone training
sh run_standalone_train_ascend.sh [PRETRAINED_MODEL]
# distributed training
sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL]
# eval
sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH]
```
# Script Description
## Script and Sample Code
## Example structure
```shell
.
└─FasterRcnn
├─README.md
├─README.md // descriptions about fasterrcnn
├─scripts
├─run_download_process_data.sh
├─run_standalone_train.sh
├─run_train.sh
└─run_eval.sh
├─run_standalone_train_ascend.sh // shell script for standalone on ascend
├─run_distribute_train_ascend.sh // shell script for distributed on ascend
└─run_eval_ascend.sh // shell script for eval on ascend
├─src
├─FasterRcnn
├─__init__.py
├─anchor_generator.py
├─bbox_assign_sample.py
├─bbox_assign_sample_stage2.py
├─faster_rcnn_r50.py
├─fpn_neck.py
├─proposal_generator.py
├─rcnn.py
├─resnet50.py
├─roi_align.py
└─rpn.py
├─config.py
├─dataset.py
├─lr_schedule.py
├─network_define.py
└─util.py
├─eval.py
└─train.py
├─__init__.py // init file
├─anchor_generator.py // anchor generator
├─bbox_assign_sample.py // first stage sampler
├─bbox_assign_sample_stage2.py // second stage sampler
├─faster_rcnn_r50.py // fasterrcnn network
├─fpn_neck.py //feature pyramid network
├─proposal_generator.py // proposal generator
├─rcnn.py // rcnn network
├─resnet50.py // backbone network
├─roi_align.py // roi align network
└─rpn.py // region proposal network
├─config.py // total config
├─dataset.py // create dataset and process dataset
├─lr_schedule.py // learning ratio generator
├─network_define.py // network define for fasterrcnn
└─util.py // routine operation
├─eval.py //eval scripts
└─train.py // train scripts
```
## Running the example
### Train
## Training Process
#### Usage
### Usage
```
# distributed training
sh run_distribute_train.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL]
# standalone training
sh run_standalone_train.sh [PRETRAINED_MODEL]
# standalone training on ascend
sh run_standalone_train_ascend.sh [PRETRAINED_MODEL]
# distributed training on ascend
sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL]
```
> Rank_table.json which is specified by RANK_TABLE_FILE is needed when you are running a distribute task. You can generate it by using the [hccl_tools](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools).
> As for PRETRAINED_MODEL，it should be a ResNet50 checkpoint that trained over ImageNet2012. Ready-made pretrained_models are not available now. Stay tuned.
#### Result
### Result
Training result will be stored in the example path, whose folder name begins with "train" or "train_parallel". You can find checkpoint file together with result like the followings in loss.log.
-[Description of random situation](#description-of-random-situation)
-[others](#others)
-[ModelZoo Homepage](#modelzoo-homepage)
...
...
@@ -50,12 +49,13 @@ Inspired by BERT, GPT and other language models, MicroSoft addressed [MASS: Mask
[Paper](https://www.microsoft.com/en-us/research/uploads/prod/2019/06/MASS-paper-updated-002.pdf): Song, Kaitao, Xu Tan, Tao Qin, Jianfeng Lu and Tie-Yan Liu. “MASS: Masked Sequence to Sequence Pre-training for Language Generation.” ICML (2019).
# Model Architecture
# Model architecture
The overall network architecture of MASS is shown below, which is Transformer(Vaswani et al., 2017):
MASS is consisted of 6-layer encoder and 6-layer decoder with 1024 embedding/hidden size, and 4096 intermediate size between feed forward network which has two full connection layers.
The MASS network is implemented by Transformer, which has multi-encoder layers and multi-decoder layers.
For pre-training, we use the Adam optimizer and loss-scale to get the pre-trained model.
During fine-turning, we fine-tune this pre-trained model with different dataset according to different tasks.
During testing, we use the fine-turned model to predict the result, and adopt a beam search algorithm to
get the most possible prediction results.
# Dataset
...
...
@@ -465,86 +465,18 @@ For Inverse square root scheduler, config could be like:
More detail about LR scheduler could be found in `src/utils/lr_scheduler.py`.
# Model description
The MASS network is implemented by Transformer, which has multi-encoder layers and multi-decoder layers.
For pre-training, we use the Adam optimizer and loss-scale to get the pre-trained model.
During fine-turning, we fine-tune this pre-trained model with different dataset according to different tasks.
During testing, we use the fine-turned model to predict the result, and adopt a beam search algorithm to
get the most possible prediction results.
## Performance
### Results
#### Fine-Tuning on Text Summarization
The comparisons between MASS and two other pre-training methods in terms of ROUGE score on the text summarization task
| Dataset | Gigaword corpus, Cornell Movie Dialog corpus |
| batch_size | --- |
| outputs | Sentence and probability |
| Accuracy | ppl=23.52 for conversation response, RG-1=29.79 for text summarization. |
| Speed | ---- sentences/s |
| Total time | --/-- |
| Model for inference | ---Mb, --, [A link]() |
# Environment Requirements
## Platform
- 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 could get the resources for trial.
- 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.
Transformer was proposed in 2017 and designed to process sequential data. It is adopted mainly in the field of natural language processing(NLP), for tasks like machine translation or text summarization. Unlike traditional recurrent neural network(RNN) which processes data in order, Transformer adopts attention mechanism and improve the parallelism, therefore reduced training times and made training on larger datasets possible. Since Transformer model was introduced, it has used to tackle many problems in NLP and derives many network models, such as BERT(Bidirectional Encoder Representations from Transformers) and GPT(Generative Pre-trained Transformer).
Transformer was proposed in 2017 and designed to process sequential data. It is adopted mainly in the field of natural language processing(NLP), for tasks like machine translation or text summarization. Unlike traditional recurrent neural network(RNN) which processes data in order, Transformer adopts attention mechanism and improve the parallelism, therefore reduced training times and made training on larger datasets possible. Since Transformer model was introduced, it has been used to tackle many problems in NLP and derives many network models, such as BERT(Bidirectional Encoder Representations from Transformers) and GPT(Generative Pre-trained Transformer).
[Paper](https://arxiv.org/abs/1706.03762): Ashish Vaswani, Noam Shazeer, Niki Parmar, JakobUszkoreit, Llion Jones, Aidan N Gomez, Ł ukaszKaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS 2017, pages 5998–6008.
