PyTorch common framework to accelerate network implementation, training and validation.
This framework is inspired by works from MMLab, which modularize the data, network, loss, metric, etc. to make the framework to be flexible, easy to modify and to extend.
# install necessary libs
pip install -r requirements.txt
The framework contains six different subfolders:
- networks: all networks should be implemented under the networks folder with {NAME}_network.py filename.
- datasets: all datasets should be implemented under the datasets folder with {NAME}_dataset.py filename.
- losses: all losses should be implemented under the losses folder with {NAME}_loss.py filename.
- metrics: all metrics should be implemented under the metrics folder with {NAME}_metric.py filename.
- models: all models should be implemented under the models folder with {NAME}_model.py filename.
- utils: all util functions should be implemented under the utils folder with {NAME}_util.py filename.
The training and validation procedure can be defined in the specified .yaml file.
# training
CUDA_VISIBLE_DEVICES=gpu_ids python train.py --opt options/train.yaml
# validation/test
CUDA_VISIBLE_DEVICES=gpu_ids python test.py --opt options/test.yaml
In the .yaml file for training, you can define all the things related to training such as the experiment name, model, dataset, network, loss, optimizer, metrics and other hyper-parameters. Here is an example to train VGG16 for image classification:
# general setting
name: vgg_train
backend: dp # DataParallel
type: ClassifierModel
num_gpu: auto
# path to resume network
path:
resume_state: ~
# datasets
datasets:
train_dataset:
name: TrainDataset
type: ImageNet
data_root: ../data/train_data
val_dataset:
name: ValDataset
type: ImageNet
data_root: ../data/val_data
# setting for train dataset
batch_size: 8
# network setting
networks:
classifier:
type: VGG16
num_classes: 1000
# training setting
train:
total_iter: 10000
optims:
classifier:
type: Adam
lr: 1.0e-4
schedulers:
classifier:
type: none
losses:
ce_loss:
type: CrossEntropyLoss
# validation setting
val:
val_freq: 10000
# log setting
logger:
print_freq: 100
save_checkpoint_freq: 10000
In the .yaml file for validation, you can define all the things related to validation such as: model, dataset, metrics. Here is an example:
# general setting
name: test
backend: dp # DataParallel
type: ClassifierModel
num_gpu: auto
manual_seed: 1234
# path
path:
resume_state: experiments/train/models/final.pth
resume: false
# datasets
datasets:
val_dataset:
name: ValDataset
type: ImageNet
data_root: ../data/test_data
# network setting
networks:
classifier:
type: VGG
num_classes: 1000
# validation setting
val:
metrics:
accuracy:
type: calculate_accuracy
The core of the framework is the BaseModel in the base_model.py. The BaseModel controls the whole training/validation procedure from initialization over training/validation iteration to results saving.
- Initialization: In the model initialization, it will read the configuration in the .yaml file and construct the corresponding networks, datasets, losses, optimizers, metrics, etc.
- Training/Validation: In the training/validation procedure, you can refer the training process in the train.py and the validation process in the test.py.
- Results saving: The model will automatically save the state_dict for networks, optimizers and other hyperparameters during the training.
The configuration of the framework is down by Register in the registry.py. The Register has a object map (key-value pair). The key is the name of the object, the value is the class of the object. There are total 4 different registers for networks, datasets, losses and metrics. Here is an example to register a new network:
import torch
import torch.nn as nn
from utils.registry import NETWORK_REGISTRY
@NETWORK_REGISTRY.register()
class MyNet(nn.Module):
...