Skip to content

Latest commit

 

History

History
218 lines (200 loc) · 5.85 KB

README.md

File metadata and controls

218 lines (200 loc) · 5.85 KB
Raw

BenchMark

Based on https://github.com/victoresque/pytorch-template

How to train a model?

python train.py -c path_to_config.json

If 'n_gpu' in config.json is more than 1 the model will be wrapped with torch.nn.DataParallel()

How to train using DDP?

python train_ddp.py --resume path_to_model.pth

In addition to using train_ddp.py, dist_backend & dist_url options should be defined in config.json, other wise default values defined in train_ddp.py will be used.

How to test a model?

python test.py --resume path_to_model.pth

How to fuse nn.BatchNorm2d with QuantConv2d in a model?

fuser.py can be used to do so.

python fuser.py -r /path/to/pth/file
  • --bs Controls the batch size.

How to extract parameters from model?

python parameters_extractor.py --model path_to_model.pth

To fuse BatchNorm2d into the preceding Conv2d, add -f option to the previous command.

How to seprate the weights?

In this template the weights are saved together with config & other objects in the pth file, it can be seprated using the following command.

python checkpoint_separator.py --model path_to_model.pth

The resulting weights file will be saved next to the model.

How to create config.json?

Config.json is a json file that descrips the experiment to run, sevral examples are avalibe under config_json.

Keys dictionary

name

Name of the experment, a timestamp will be appended to it using '_%d_%m_%H%M%S' as format. This can be modified in parse_config.json

n_gpu

Total number of gpus to use. If set to -1 it will use all the available gpus

dist_backend

Distributed backend to use. For more info check relative PyTorch docs.

dist_url

URL specifying how to initialize the process group. For more info check relative PyTorch docs.

arch

The architecure can be defined using this key. The value should be a dictionary with two parameters type & args. Types can be {VGG_net,QuantVGG_pure,vgg16,alexnet}.
VGG_net: Local implementation of VGG.
QuantVGG_pure: Quantized VGG using xilinx-brevitas.
vgg16 & alexnet: Are the respective models taken from Pytorch.

Example of QuantVGG_pure

"arch": {
        "type": "QuantVGG_pure",
        "args": {
            "VGG_type":"D", 
            "batch_norm":false, 
            "bit_width":16, 
            "num_classes":1000,
            "pretrained_model":"Path to pretrained model or use pytorch to initialize with pytorch's version of the model"
        }
    }

Example of VGG_net

"arch": {
        "type": "VGG_net",
        "args": {
            "in_channels":3, 
            "num_classes":1000,
            "VGG_type":"D", 
            "batch_norm":true
        }
    }

Example of vgg16

"arch": {
        "type": "vgg16",
        "args": {
            "pretrained":true, 
            "progress":true
        }
    }

data_loader

Example of CIFAR_data_loader

    "data_loader": {
        "type": "CIFAR_data_loader",
        "args":{
            "data_dir": "/Path/to/Data_set",
            "batch_size": 512,
            "download": true,
            "shuffle": true,
            "validation_split": 0.1,
            "num_workers": 5,
            "flavor": 100,
            "training": true
        }
    }

Example of ImageNet_data_loader

    "data_loader": {
        "type": "ImageNet_data_loader",
        "args":{
            "data_dir": "/Path/to/Data_set",
            "batch_size": 512,
            "shuffle": true,
            "num_workers": 5,
            "pin_memory":true,
            "training": true
        }
    }

test_data_loader

Example of CIFER test_data_loader

    "data_loader": {
        "type": "CIFAR_data_loader",
        "args":{
            "data_dir": "/Path/to/Data_set",
            "batch_size": 512,
            "download": true,
            "shuffle": true,
            "validation_split": 0.0,
            "num_workers": 5,
            "flavor": 100,
            "training": false
        }
    }

loss

to do

metrics

Metrics used to evaluate the model, currently defined metrics are accuracy & top_k_acc. Default value for k in top_k_acc is 5. New metrics can be defined under model/metric.py.

lr_scheduler

Example of StepLR

"lr_scheduler": {
    "type": "StepLR",
    "args": {
        "step_size": 30,
        "gamma": 0.1,
        "verbose": true
    }
}

Example of MultiStepLR

"lr_scheduler": {
    "type": "MultiStepLR",
    "args": {
        "milestones": [60,120,160],
        "gamma": 0.2,
        "verbose": true
    }
}

trainer

Example of trainer config

"trainer": {
    "epochs": 200,
    "save_dir": "/Where/to/save/train_result",
    "save_period": 200,
    "verbosity": 2,                 // 0: quiet, 1: per epoch, 2: full
    "monitor": "max val_accuracy", 
    "early_stop": -1,
    "tensorboard": true
}

extract

If set the model parameters will be extracted during testing

"extract": true,

extractor

Configrations to be used in the resulting config.h file

"extractor": {
    "PE": 1,
    "SIMD": 1,
    "DATAWIDTH": 64,
    "SEQUENCE_LENGTH": 120000,
    "CLASS_LABEL_BITS": 1,
    "MUL_BITS": 16,
    "MUL_INT_BITS": 8,
    "ACC_BITS": 16,
    "ACC_INT_BITS": 8,
    "IA_BITS": 8,
    "IA_INT_BITS": 4
}

To Do

-Generic model initialization from PyTorch\

Notes

  • If you get NCCL errors rerun with export NCCL_DEBUG=WARN
  • If you get RuntimeError: CUDA error: all CUDA-capable devices are busy or unavailable while training with DDP try setting pin_memory to false in config.json