torchfurnace 0.0.6

A tool package for training model, pre-processing dataset and managing experiment record in pytorch AI tasks.

torchfurnace

torchfurnace is a tool package for training model, pre-processing dataset and managing experiment record in pytorch AI tasks.

Quick Start

Usage

pip install torchfurnace

Example

VGG16 for CIFAR10

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
from torchvision.datasets import CIFAR10
from torch.optim.lr_scheduler import MultiStepLR
from torchfurnace import Engine, Parser

# define training process of your model
class VGGNetEngine(Engine):
    @staticmethod
    def _on_forward(training, model, inp, target, optimizer=None) -> dict:
        ret = {'loss': object, 'preds': object}
        output = model(inp)
        loss = F.cross_entropy(output, target)
        if training:
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        ret['loss'] = loss
        ret['preds'] = output
        return ret

    @staticmethod
    def _get_lr_scheduler(optim) -> list:
        return [MultiStepLR(optim, milestones=[150, 250, 350], gamma=0.1)]

def main():
    # define experiment name
    parser = Parser('TVGG16')
    args = parser.parse_args()
    experiment_name = '_'.join([args.dataset, args.exp_suffix])

    # Data
    ts = transforms.Compose([transforms.ToTensor(), transforms.Normalize(
        (0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))])
    trainset = CIFAR10(root='data', train=True, download=True, transform=ts)
    testset = CIFAR10(root='data', train=False, download=True, transform=ts)

    # define model and optimizer
    net = torchvision.models.vgg16(pretrained=False, num_classes=10)
    net.avgpool = nn.AvgPool2d(kernel_size=1, stride=1)
    net.classifier = nn.Linear(512, 10)
    optimizer = torch.optim.Adam(net.parameters())

    # new engine instance
    eng = VGGNetEngine(parser).experiment_name(experiment_name)
    acc1 = eng.learning(net, optimizer, trainset, testset)
    print('Acc1:', acc1)

if __name__ == '__main__':
    import sys
    run_params='--dataset CIFAR10 -lr 0.1 -bs 128 -j 2 --epochs 400 --adjust_lr'
    sys.argv.extend(run_params.split())
    main()

Introduction

Why do this?

There are some deep learning frameworks to quickly build a training system in pytorch AI tasks, however, I found that most of them are complex framework which have higher cost for learning it and seriously invade original code , for instance, maybe modify your model class to adapt the framework.

So, torchfurnace is born for perform your pytorch AI task quickly, simply and without invasion viz you don't have to change too much defined code.

What features?

1. torchfurnace consists of two independent components including engine and tracer. engine is a core component of proposed framework, and tracer is a manager of experiment whose obligation include log writing, model saving and training visualization.

2. torchfurnace integrates some practical tools, such as processing raw dataset to LMDB for solving I/O bottleneck and computing the number of parameter size and dataset with std and mean.

Components

Engine

from torchfurnace import Engine

This is core component of the proposed framework. Engine is an abstract class which controls the whole workflow to finish AI tasks. So, you have to implement _on_forward by inheriting the Engine when you start. The job of _on_forward is to define which criterion to use and how to optimize your model.

For example (a minimal configuration):

class VGGNetEngine(Engine):
    @staticmethod
    def _on_forward(training, model, inp, target, optimizer=None) -> dict:
        ret = {'loss': object, 'preds': object}
        output = model(inp)
        loss = F.cross_entropy(output, target)
        if training:
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        ret['loss'] = loss
        ret['preds'] = output
        return ret

Above, you have done the preparatory work of framework, and then you define model, train dataset, val dataset and optimizer as usual. This highlights the advantages of our method which don't heavily invade original code. In other words, you don't have to modify your code too much. As follow, only 3 lines to finish learning. Of course, you can also assign a list type as model and optimizer, however, it's vital each element of models and optimizer must be matched.

parser = Parser('TVGG11')
eng = VGGNetEngine(parser).experiment_name(experiment_name)
eng.learning(model, optimizer, train_ds, val_ds)

If you want more customizable features, you can override them with the following listed functions.

overridden function

1. _get_lr_scheduler: This function support to define other scheduler in torch.optim.lr_scheduler. Default: StepLR with gamma=0.1 and step_size=30
2. _on_start_epoch: You can add some meters , such as AverageMeter using from torchfurnace.utils.meter import AverageMeter in this function when you want to record more information in the training processing.
3. _on_start_batch: define how to get input and target from your dataset and put them on GPU.
4. _add_on_end_batch_log: You can add some log output information just like other meters in _on_end_batch.
5. _add_on_end_batch_tb: You can add some tensorboard output information just like other meters in _on_end_batch.

Tracer

from torchfurnace import Tracer

The responsibility of this component is to manage log writing, model saving and training visualization in experiment.

Especially, this component can run independently, meanwhile, it is integrated in Engine as a nested object to manage experiment record.

For example:

from torchfurnace import Tracer
tracer = Tracer('my_network').attach('expn')

save experiment configuration

import torchfurnace.utils.tracer_component as tc
cfg = {"optimizer": optim, 'addtion': 'hhh'}
tracer.store(tc.Config({**cfg, **vars(Parser('my_network').parse_args())}))

cfg: some customize information

Parser('my_network').parse_args() : parser snapshot

save checkpoint

import torchfurnace.utils.tracer_component as tc
model = models.vgg11(pretrained=False)
optimizer = torch.optim.Adam(model.parameters())
tracer.store(tc.Model(
        f"{model.__class__.__name__}_extinfo.pth.tar",
        {
            'epoch': 99,
            'arch': str(model),
            'state_dict': model.state_dict(),
            'best_acc1': 0.66,
            'optimizer': optimizer.state_dict(),
        }, True))

save checkpoint to appropriate location automatically.

load checkpoint

import torchfurnace.utils.tracer_component as tc
model = models.vgg11(pretrained=False)
optimizer = torch.optim.Adam(model.parameters())
# 'pth' has fix format. It contains two parts,
#  which are experiment file name and best model file name,
#  skipping medial path , and they are separate by '/' .
pth = f"expn/{model.__class__.__name__}_Epk{99}_Acc{0.66}_extinfo_best.pth.tar"
ret = tracer.load(tc.Model(
        pth, {
            'model': model,
            'optim': optimizer
        }))
print(ret)
# ret has start_epoch and best_acc1

load checkpoint from appropriate location automatically.

Note: --resume is assigned just like pth

display TensorBoard

tracer.tb_switch(True)
tracer.tb_switch(False)

redirect output

tracer.debug_switch(True)
tracer.debug_switch(False)

Parser

from torchfurnace import Parser

This just is an ArgumentParser, but have defined some frequently-used options.

print default options

from torchfurnace import Parser
p = Parser()
args = p.parse_args()
print(args)

{
    "dataset":"",
    "batch_size":1,
    "workers":2,
    "lr":0.01,
    "weight_decay":0.0005,
    "momentum":0.9,
    "dropout":0.5,
    "start_epoch":0,
    "epochs":200,
    "gpu":"0",
    "exp_suffix":"",
    "extension":"",
    "resume":"",
    "evaluate":false,
    "deterministic":false,
    "adjust_lr":false,
    "print_freq":10,
    "logger_name":"log.txt",
    "work_dir":"",
    "clean_up":5,
    "debug":false,
    "p_bar":false,
    "no_tb":true,
    "nowtime_exp":true
}

Also, you can open from torchfurnace import Parser to get more information.

add option in Engine

parser = Parser('TVGG16')
parser.add_argument('--add',default='addtion',type=str)
eng = Engine(parser).experiment_name(experiment_name)

ImageFolderLMDB

from torchfurnace import ImageFolderLMDB

This component is a tool to accelerate data reading speed. Especially, your dataset folder should have a specific named regular just like the regular in from torchvision.datasets import ImageFolder

Like this:

demo_dataset/
├── train
│   ├── cat
│   │   └── 001.jpg
│   └── dog
│       └── 001.jpg
└── val
    ├── cat
    │   └── 001.jpg
    └── dog
        └── 001.jpg

make LMDB

ImageFolderLMDB.folder2lmdb('demo_dataset', name='train', num_workers=16)
ImageFolderLMDB.folder2lmdb('demo_dataset', name='val', num_workers=16)

use LMDB

from torch.utils.data import DataLoader
_set = ImageFolderLMDB('demo_dataset/train.lmdb')
loader = DataLoader(_set, batch_size=1, collate_fn=lambda x: x)

ImageLMDB

from torchfurnace import ImageLMDB

This component is a tool for converting .jpg/.png to binary data and storing to LMDB. Because of reading lots of small size image always costing expensive I/O resources, it will improve speed of data loading.

make LMDB

data_mapping = {
   'key': [f"{k}_{no:04}" for k in ['dog', 'cat'] for no in range(1, 4)],
   'pic_path': [f"demo_dataset/{k}{no:03}.jpg" for k in ['d', 'c'] for no in range(1, 4)]
}
ImageLMDB.store(data_mapping, 'demo_imgs_db', 4800000, num_workers=16)

data_mapping provide key and picture path.

read LMDB

import matplotlib.pyplot as plt
read_keys = {
    'key': [f"{k}_{no:04}" for k in ['dog', 'cat'] for no in range(1, 4)]
}

for key in data_mapping['key']:
    img = ImageLMDB.read('demo_imgs_db', key)
    plt.imshow(img)
    plt.axis('off')
    plt.show()

The data from LMDB format is PIL object, so you can do other image processing, such as random crop, random flip and normalize using torchvision.transforms.Compose

Compute mean and std

Compute the mean and std value of dataset using GPU.

import torchvision.transforms as transforms
from torchvision.datasets import CIFAR10
from torchfurnace.utils.function import get_mean_and_std

ts = transforms.Compose([transforms.ToTensor()])
trainset = CIFAR10(root='data', train=True, download=True, transform=ts)
mean, std = get_mean_and_std(trainset)
print(mean, std)

Model Summary

This tool comes from pytorch-summary.

import torchvision
from torchfurnace.utils.torch_summary import summary, summary_string
net = torchvision.models.vgg16()

# this function will output result on screen.  
summary(net,(3,224,224))

# this funcion will return a string of description.
desc = summary_string(net,(3,224,224))

Directory Architecture

TVGG16/
├── logs
│   └── CIFAR10
│       └── log.txt
├── models
│   └── CIFAR10
│       ├── architecture.txt
│       ├── checkpoint
│       │   └── best
│       └── run_config.json
└── tensorboard
    └── CIFAR10
        └── events.out.tfevents

TVGG16: the root of your work name

CIFAR10: one of experiment name

TVGG16/logs: all kinds of experiments output files

TVGG16/models: all kinds of experiments checkpoints, network architectures and run configurations

TVGG16/tensorboard: all kinds of experiments outputs of tensorboard

Testing & Example

In this section, you have to git clone https://github.com/tianyu-su/torchfurnace.git.

1. torchfurnace/tests/test_furnace.py A unit test for Engine.
2. torchfurnace/tests/test_tracer.py A unit test for Tracer.
3. torchfurnace/tests/test_img2lmdb.py A unit test for convert images to LMDB.
4. torchfurnace/tests/test_vgg16.py A compare experiment with pytorch-cifar to validate pipeline of the proposed framework.

More Usages

Following methods are based on default setup.

1. open debug mode: --debug this mode is easy to debug.
2. close tensorboard: --no_tb
3. open process bar: --p_bar
4. change total data directory: --work_dir your_path
5. remain best checkpoint Top k: --clean_up k
6. run again with same setup, but default options will override old data. --nowtime_exp
7. make special marks for model checkpoint or other aspects: --ext sp1
8. make special marks for experiment name: --exp_suffix

Other settings please look at ops.

TODO

• DistributedDataParallel