m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

modify resnet50-quant readme

Browse Source
pull/7603/head
yuchaojie 5 years ago
parent b2fb6c22e2
commit b37debbef5
2 changed files with 74 additions and 68 deletions
Show Stats Download Patch File Download Diff File

138
model_zoo/official/cv/resnet50_quant/Readme.md → model_zoo/official/cv/resnet50_quant/README.md
Unescape View File

@ -1,6 +1,6 @@
# Contents # Contents
- [resnet50 Description](#resnet50-description) - [ResNet50 Description](#resnet50-description)
- [Model Architecture](#model-architecture) - [Model Architecture](#model-architecture)
- [Dataset](#dataset) - [Dataset](#dataset)
- [Features](#features) - [Features](#features)
@ -18,15 +18,15 @@
- [Description of Random Situation](#description-of-random-situation) - [Description of Random Situation](#description-of-random-situation)
- [ModelZoo Homepage](#modelzoo-homepage) - [ModelZoo Homepage](#modelzoo-homepage)
# [resnet50 Description](#contents) # [ResNet50 Description](#contents)
ResNet-50 is a convolutional neural network that is 50 layers deep, which can classify ImageNet image to 1000 object categories with 76% accuracy. ResNet-50 is a convolutional neural network that is 50 layers deep, which can classify ImageNet image to 1000 object categories with 76% accuracy.
[Paper](https://arxiv.org/abs/1512.03385) Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun."Deep Residual Learning for Image Recognition." He, Kaiming , et al. "Deep Residual Learning for Image Recognition." IEEE Conference on Computer Vision & Pattern Recognition IEEE Computer Society, 2016. [Paper](https://arxiv.org/abs/1512.03385): Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun."Deep Residual Learning for Image Recognition." He, Kaiming , et al. "Deep Residual Learning for Image Recognition." IEEE Conference on Computer Vision & Pattern Recognition IEEE Computer Society, 2016.
This is the quantitative network of Resnet50. This is the quantitative network of ResNet50.
# [Model architecture](#contents) # [Model Architecture](#contents)
The overall network architecture of Resnet50 is show below: The overall network architecture of Resnet50 is show below:
@ -34,13 +34,20 @@ The overall network architecture of Resnet50 is show below:
# [Dataset](#contents) # [Dataset](#contents)
Dataset used: [imagenet](http://www.image-net.org/) Dataset used: [ImageNet2012](http://www.image-net.org/)
- Dataset size: ~125G, 1.2W colorful images in 1000 classes - Dataset size 224*224 colorful images in 1000 classes
- Train: 120G, 1.2W images - Train：1,281,167 images
- Test: 5G, 50000 images - Test： 50,000 images
- Data format: RGB images. - Data format：jpeg
- Note: Data will be processed in src/dataset.py - Note：Data will be processed in dataset.py
- Download the dataset, the directory structure is as follows:
```
└─dataset
├─ilsvrc # train dataset
└─validation_preprocess # evaluate dataset
```
# [Features](#contents) # [Features](#contents)
@ -70,9 +77,10 @@ For FP16 operators, if the input data type is FP32, the backend of MindSpore wil
├── Readme.md # descriptions about Resnet50-Quant ├── Readme.md # descriptions about Resnet50-Quant
├── scripts ├── scripts
│ ├──run_train.sh # shell script for train on Ascend │ ├──run_train.sh # shell script for train on Ascend
│ ├──run_infer.sh # shell script for evaluation on Ascend │ ├──run_infer.sh # shell script for evaluation on Ascend
├── model ├── models
│ ├──resnet_quant.py # define the network model of resnet50-quant │ ├──resnet_quant.py # define the network model of resnet50-quant
│ ├──resnet_quant_manual.py # define the manually quantized network model of resnet50-quant
├── src ├── src
│ ├──config.py # parameter configuration │ ├──config.py # parameter configuration
│ ├──dataset.py # creating dataset │ ├──dataset.py # creating dataset
@ -91,24 +99,24 @@ Parameters for both training and evaluation can be set in config.py
- config for Resnet50-quant, ImageNet2012 dataset - config for Resnet50-quant, ImageNet2012 dataset
```python ```python
'class_num': 10 # the number of classes in the dataset 'class_num': 10 # the number of classes in the dataset
'batch_size': 32 # training batch size 'batch_size': 32 # training batch size
'loss_scale': 1024 # the initial loss_scale value 'loss_scale': 1024 # the initial loss_scale value
'momentum': 0.9 # momentum 'momentum': 0.9 # momentum
'weight_decay': 1e-4 # weight decay value 'weight_decay': 1e-4 # weight decay value
'epoch_size': 120 # total training epochs 'epoch_size': 120 # total training epochs
'pretrained_epoch_size': 90 # pretraining epochs of resnet50, which is unquantative network of resnet50_quant 'pretrained_epoch_size': 90 # pretraining epochs of resnet50, which is unquantative network of resnet50_quant
'data_load_mode': 'mindata' # the style of loading data into device 'data_load_mode': 'mindata' # the style of loading data into device
'save_checkpoint':True # whether save checkpoint file after training finish 'save_checkpoint':True # whether save checkpoint file after training finish
'save_checkpoint_epochs': 1 # the step from which start to save checkpoint file. 'save_checkpoint_epochs': 1 # the step from which start to save checkpoint file.
'keep_checkpoint_max': 50 # only keep the last keep_checkpoint_max checkpoint 'keep_checkpoint_max': 50 # only keep the last keep_checkpoint_max checkpoint
'save_checkpoint_path': './' # the absolute full path to save the checkpoint file 'save_checkpoint_path': './' # the absolute full path to save the checkpoint file
"warmup_epochs": 0 # number of warmup epochs "warmup_epochs": 0 # number of warmup epochs
'lr_decay_mode': "cosine" #learning rate decay mode, including steps, steps_decay, cosine or liner 'lr_decay_mode': "cosine" # learning rate decay mode, including steps, steps_decay, cosine or liner
'use_label_smooth': True #whether use label smooth 'use_label_smooth': True # whether use label smooth
'label_smooth_factor': 0.1 #label smooth factor 'label_smooth_factor': 0.1 # label smooth factor
'lr_init': 0 # initial learning rate 'lr_init': 0 # initial learning rate
'lr_max': 0.005 # the max learning rate 'lr_max': 0.005 # the max learning rate
``` ```
## [Training process](#contents) ## [Training process](#contents)
@ -120,7 +128,7 @@ Parameters for both training and evaluation can be set in config.py
``` ```
# training example # training example
Ascend: bash run_train.sh Ascend ~/hccl_4p_0123_x.x.x.x.json ~/imagenet/train/ Ascend: bash run_train.sh Ascend ~/hccl.json ~/imagenet/train/ ~/pretrained_ckeckpoint
``` ```
### Result ### Result
@ -165,41 +173,39 @@ result: {'acc': 0.76576314102564111}
## [Performance](#contents) ## [Performance](#contents)
### Training Performance ### Evaluation Performance
| Parameters | Resnet50 | | Parameters | Ascend |
| -------------------------- | ---------------------------------------------------------- | | -------------------------- | ----------------------------------------------------------- |
| Model Version | V1 | | Model Version | ResNet50 V1.5 |
| Resource | Ascend 910, cpu:2.60GHz 56cores, memory:314G | | Resource | Ascend 910, CPU 2.60GHz, 56cores, Memory 314G |
| uploaded Date | 06/06/2020 | | uploaded Date | 06/06/2020 (month/day/year) |
| MindSpore Version | 0.3.0 | | MindSpore Version | 0.3.0-alpha |
| Dataset | ImageNet | | Dataset | ImageNet |
| Training Parameters | src/config.py | | Training Parameters | epoch=30(with pretrained) or 120, steps per epoch=5004, batch_size=32 |
| Optimizer | Momentum | | Optimizer | Momentum |
| Loss Function | SoftmaxCrossEntropy | | Loss Function | Softmax Cross Entropy |
| outputs | ckpt file | | outputs | probability |
| Loss | 1.8 | | Loss | 1.8 |
| Accuracy | | | Speed | 8pcs: 407 ms/step |
| Total time | 16h | | Total time | 8pcs: 17 hours(30 epochs with pretrained) |
| Params (M) | batch_size=32, epoch=30 | | Parameters (M) | 25.5 |
| Checkpoint for Fine tuning | | | Checkpoint for Fine tuning | 197M (.ckpt file) |
| Model for inference | | | Scripts | [resnet50-quant script](https://gitee.com/mindspore/mindspore/tree/r1.0/model_zoo/official/cv/resnet50_quant) |
#### Evaluation Performance #### Inference Performance
| Parameters | Resnet50 | | Parameters | Ascend |
| -------------------------- | ----------------------------- | | ------------------- | --------------------------- |
| Model Version | V1 | | Model Version | ResNet50 V1.5 |
| Resource | Ascend 910 | | Resource | Ascend 910, CPU 2.60GHz, 56cores, Memory 314G |
| uploaded Date | 06/06/2020 | | Uploaded Date | 06/06/2020 (month/day/year) |
| MindSpore Version | 0.3.0 | | MindSpore Version | 0.3.0-alpha |
| Dataset | ImageNet, 1.2W | | Dataset | ImageNet |
| batch_size | 130(8P) | | batch_size | 32 |
| outputs | probability | | outputs | probability |
| Accuracy | ACC1[76.57%] ACC5[92.90%] | | Accuracy | ACC1[76.57%] ACC5[92.90%] |
| Speed | 5ms/step | | Model for inference | 197M (.ckpt file) |
| Total time | 5min |
| Model for inference | |
# [Description of Random Situation](#contents) # [Description of Random Situation](#contents)

4
model_zoo/official/nlp/transformer/README.md
Unescape View File

@ -233,7 +233,7 @@ Parameters for learning rate:
### Training Performance ### Training Performance
| Parameters | Transformer | | Parameters | Ascend |
| -------------------------- | -------------------------------------------------------------- | | -------------------------- | -------------------------------------------------------------- |
| Resource | Ascend 910 | | Resource | Ascend 910 |
| uploaded Date | 06/09/2020 (month/day/year) | | uploaded Date | 06/09/2020 (month/day/year) |
@ -252,7 +252,7 @@ Parameters for learning rate:
### Evaluation Performance ### Evaluation Performance
| Parameters | GoogleNet | | Parameters | Ascend |
| ------------------- | --------------------------- | | ------------------- | --------------------------- |
| Resource | Ascend 910 | | Resource | Ascend 910 |
| Uploaded Date | 06/09/2020 (month/day/year) | | Uploaded Date | 06/09/2020 (month/day/year) |

Write Preview
Loading…
Cancel
Save