|
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172 |
- """
- # -*- coding: utf-8 -*-
- -----------------------------------------------------------------------------------
- # Author: Nguyen Mau Dung
- # DoC: 2020.08.17
- # email: nguyenmaudung93.kstn@gmail.com
- -----------------------------------------------------------------------------------
- # Description: The configurations of the project will be defined here
- """
-
- import os
- import argparse
-
- import torch
- from easydict import EasyDict as edict
-
-
- def parse_train_configs():
- parser = argparse.ArgumentParser(description='The Implementation using PyTorch')
- parser.add_argument('--seed', type=int, default=2020,
- help='re-produce the results with seed random')
- parser.add_argument('--saved_fn', type=str, default='fpn_resnet_18', metavar='FN',
- help='The name using for saving logs, models,...')
-
- parser.add_argument('--root_dir', type=str, default='../', metavar='PATH',
- help='The ROOT working directory')
- ####################################################################
- ############## Model configs ########################
- ####################################################################
- parser.add_argument('--arch', type=str, default='fpn_resnet_18', metavar='ARCH',
- help='The name of the model architecture')
- parser.add_argument('--model_load_dir', type=str, default=None, metavar='PATH',
- help='the path of the pretrained checkpoint')
-
- ####################################################################
- ############## Dataloader and Running configs #######
- ####################################################################
- parser.add_argument('--data_url', type=str, default='../dataset/apollo/training', metavar='PATH',
- help='the path of the dataset')
- parser.add_argument('--val_data_url', type=str, default='../dataset/apollo/val', metavar='PATH',
- help='the path of the dataset')
- parser.add_argument('--train_model_out', type=str, default='../checkpoints', metavar='PATH',
- help='the path of the model output')
- parser.add_argument('--train_out', type=str, default='../logs', metavar='PATH',
- help='the path of the logs output')
- parser.add_argument('--hflip_prob', type=float, default=0.5,
- help='The probability of horizontal flip')
- parser.add_argument('--no-val', action='store_true',
- help='If true, dont evaluate the model on the val set')
- parser.add_argument('--num_samples', type=int, default=None,
- help='Take a subset of the dataset to run and debug')
- parser.add_argument('--num_workers', type=int, default=4,
- help='Number of threads for loading data')
- parser.add_argument('--batch_size', type=int, default=8,
- help='mini-batch size (default: 16), this is the total'
- 'batch size of all GPUs on the current node when using'
- 'Data Parallel or Distributed Data Parallel')
- parser.add_argument('--print_freq', type=int, default=50, metavar='N',
- help='print frequency (default: 50)')
- parser.add_argument('--tensorboard_freq', type=int, default=50, metavar='N',
- help='frequency of saving tensorboard (default: 50)')
- parser.add_argument('--checkpoint_freq', type=int, default=2, metavar='N',
- help='frequency of saving checkpoints (default: 5)')
- parser.add_argument('--gpu_num_per_node', type=int, default=1,
- help='Number of GPU')
- ####################################################################
- ############## Training strategy ####################
- ####################################################################
-
- parser.add_argument('--start_epoch', type=int, default=1, metavar='N',
- help='the starting epoch')
- parser.add_argument('--num_epochs', type=int, default=300, metavar='N',
- help='number of total epochs to run')
- parser.add_argument('--lr_type', type=str, default='cosin',
- help='the type of learning rate scheduler (cosin or multi_step or one_cycle)')
- parser.add_argument('--lr', type=float, default=0.001, metavar='LR',
- help='initial learning rate')
- parser.add_argument('--minimum_lr', type=float, default=1e-7, metavar='MIN_LR',
- help='minimum learning rate during training')
- parser.add_argument('--momentum', type=float, default=0.949, metavar='M',
- help='momentum')
- parser.add_argument('-wd', '--weight_decay', type=float, default=0., metavar='WD',
- help='weight decay (default: 0.)')
- parser.add_argument('--optimizer_type', type=str, default='adam', metavar='OPTIMIZER',
- help='the type of optimizer, it can be sgd or adam')
- parser.add_argument('--steps', nargs='*', default=[150, 180],
- help='number of burn in step')
-
- ####################################################################
- ############## Loss weight ##########################
- ####################################################################
-
- ####################################################################
- ############## Distributed Data Parallel ############
- ####################################################################
- parser.add_argument('--world-size', default=-1, type=int, metavar='N',
- help='number of nodes for distributed training')
- parser.add_argument('--rank', default=-1, type=int, metavar='N',
- help='node rank for distributed training')
- parser.add_argument('--dist-url', default='tcp://127.0.0.1:29500', type=str,
- help='url used to set up distributed training')
- parser.add_argument('--dist-backend', default='nccl', type=str,
- help='distributed backend')
- parser.add_argument('--gpu_idx', default=0, type=int,
- help='GPU index to use.')
- parser.add_argument('--no_cuda', default= False,
- help='If true, cuda is not used.')
- parser.add_argument('--multiprocessing-distributed', action='store_true',
- help='Use multi-processing distributed training to launch '
- 'N processes per node, which has N GPUs. This is the '
- 'fastest way to use PyTorch for either single node or '
- 'multi node data parallel training')
- ####################################################################
- ############## Evaluation configurations ###################
- ####################################################################
- parser.add_argument('--evaluate', action='store_true',
- help='only evaluate the model, not training')
- parser.add_argument('--resume_path', type=str, default=None, metavar='PATH',
- help='the path of the resumed checkpoint')
- parser.add_argument('--K', type=int, default=50,
- help='the number of top K')
-
- configs = edict(vars(parser.parse_args()))
-
- ####################################################################
- ############## Hardware configurations #############################
- ####################################################################
- # configs.device = torch.device('cpu' if configs.no_cuda else 'cuda')
- configs.device = torch.device('cpu' if configs.no_cuda else 'cuda:{}'.format(configs.gpu_idx))
- configs.ngpus_per_node = torch.cuda.device_count()
-
- configs.pin_memory = True
- configs.input_size = (1216, 608)
- configs.hm_size = (304, 152)
- configs.down_ratio = 4
- configs.max_objects = 50
-
- configs.imagenet_pretrained = True
- configs.head_conv = 64
- configs.num_classes = 3
- configs.num_center_offset = 2
- configs.num_z = 1
- configs.num_dim = 3
- configs.num_direction = 2 # sin, cos
-
- configs.heads = {
- 'hm_cen': configs.num_classes,
- 'cen_offset': configs.num_center_offset,
- 'direction': configs.num_direction,
- 'z_coor': configs.num_z,
- 'dim': configs.num_dim
- }
-
- configs.num_input_features = 4
-
- ####################################################################
- ############## Dataset, logs, Checkpoints dir ######################
- ####################################################################
- configs.dataset = 'apollo' # or kitti
- configs.dataset_dir = configs.data_url
- # configs.checkpoints_dir = os.path.join(configs.train_model_out, configs.saved_fn)
- configs.checkpoints_dir = configs.train_model_out
- # configs.logs_dir = os.path.join(configs.train_out, configs.saved_fn)
- configs.logs_dir = configs.train_out
- configs.pretrained_path = configs.model_load_dir
-
- if not os.path.isdir(configs.checkpoints_dir):
- os.makedirs(configs.checkpoints_dir)
- if not os.path.isdir(configs.logs_dir):
- os.makedirs(configs.logs_dir)
-
- return configs
|
