def __init__(self,
dbname: str,
db_opts=None,
load_lut=True,
train=True,
db_root='../data'):
super(torch.utils.data.Dataset, self).__init__()
self.dbname = dbname.lower()
self.db_path = osp.join(db_root, self.dbname)
# Default arguments
opts = Edict()
opts.img_fmt = 'cam%d/%04d.png' # [cam_idx, fidx]
opts.lut_fmt = 'lt_(%d,%d,%d).hwd' # lookup table fmt [equi_h, w, d]
opts.gt_depth_fmt = 'omnidepth_gt_%d/%05d.tiff' # [equi_w, fidx]
opts.equirect_size, opts.num_invdepth = [160, 640], 192
opts.phi_deg, opts.phi2_deg = 45, -1.0
opts.min_depth = 0.5 # meter scale
opts.max_depth = 1 / EPS
opts.max_fov = 220.0 # maximum FOV of input fisheye images
opts.read_input_image = True # for evaluation, False if read only GT
opts.upsample_output = False # upsample network output
opts.start, opts.step, opts.end = 1, 1, 1000 # frame
opts.train_idx, opts.test_idx = [], []
opts.gt_phi = 0.0
opts.dtype = 'nogt'
opts.use_rgb = False
# first update opts using pre-defined config
# also load ocam parameters
opts, self.ocams = utils.dbhelper.loadDBConfigs(
self.dbname, self.db_path, opts)
# update opts from the argument
opts = argparse(opts, db_opts)
# set member variables
self.opts = opts
self.img_fmt, self.lut_fmt = opts.img_fmt, opts.lut_fmt
self.gt_depth_fmt = opts.gt_depth_fmt
self.frame_idx = list(
range(opts.start, opts.end + opts.step, opts.step))
self.train_idx, self.test_idx = opts.train_idx, opts.test_idx
self.gt_phi = opts.gt_phi
self.dtype = opts.dtype
self.use_rgb = opts.use_rgb
self.equirect_size = opts.equirect_size
self.min_depth, self.max_depth = opts.min_depth, opts.max_depth
self.max_theta = np.deg2rad(opts.max_fov) / 2.0
self.phi_deg, self.phi2_deg = opts.phi_deg, opts.phi2_deg
self.num_invdepth = opts.num_invdepth
self.read_input_image = opts.read_input_image
self.upsample_output = opts.upsample_output
self.data_size = len(self.frame_idx)
self.train_size = len(self.train_idx)
self.test_size = len(self.test_idx)
self.__initSweep(load_lut)
self.train = train
def loadSample(self, fidx: int, read_input_image=True, varargin=None):
opts = Edict()
opts.remove_gt_noise = True
opts.morph_win_size = 5
opts = argparse(opts, varargin)
imgs, raw_imgs = [], []
if read_input_image:
imgs, raw_imgs = self.loadImages(fidx, True, use_rgb=self.use_rgb)
gt, valid = [], []
if self.dtype == 'gt':
gt = self.loadGTInvdepthIndex(fidx, opts.remove_gt_noise,
opts.morph_win_size)
valid = np.logical_and(gt >= 0,
gt <= self.num_invdepth).astype(np.bool)
return imgs, gt, valid, raw_imgs
def get_config():
conf = Edict()
conf.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
conf.confidence_threshold = 0.5
conf.top_k = 5000
conf.nms_threshold = 0.4
conf.keep_top_k = 750
conf.trained_model = "detection/weights/mobilenet0.25_new.pth"
conf.network = "mobile0.25" # or resnet50
if (conf.device == torch.device("cpu")):
conf.cpu = True
print("use cpu!")
else:
conf.cpu = False
print("use gpu!")
return conf
def get_config():
conf = Edict()
conf.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# conf.device=torch.device("cpu")
conf.dataroot = "data"
conf.val_batch_size = 64
conf.train_batch_size = 64
conf.input_size = 112
conf.workers = 4
conf.num_nodes = 2
conf.local_rank = -1
# conf.back_bone="mobiv3"
conf.loss = "cross_entropy"
conf.back_bone = "efficient"
conf.learning_rate = 0.01
conf.start_step = 0
conf.save_model = conf.back_bone
conf.resume = conf.save_model
conf.pretrained = False
conf.num_classes = 7
# mixup
conf.mixup = 0.1
conf.step = 40 # decrease learning rate each time
conf.smooth_eps = 0.1
conf.mixup_warmup = 455
#optimizer_params
conf.min_lr = 2.5e-6
conf.max_lr = 0.225
conf.sched = "clr"
conf.decay = 2.5e-7
conf.momentum = 0.9
#epochs
conf.start_epoch = 0
conf.epochs = 100000
conf._dtype = torch.float16
conf.mode = "triangular2"
return conf
import torch.nn as nn
import torch.nn.functional as F
import torch
from backbone_block import basic_cnn
from utils_frequent.init_net import xavier_init
from easydict import EasyDict as Edict
"""
inputs:
输入的一个特征图
可以修改的地方、可以加上选项
bottomUp feature map 到 topDown feature map 的卷积形式
topDown 下采样的方式，目前是插值
"""
cfg = Edict()
cfg.featureMapNum = 3 #总共多少特征图输入， 经过几层
cfg.bottomUpChannelsNum = [58, 116, 232, 464
] #bootomUp每层的通道数， 第0个是输入特征图的featuremap， 后面几层是金字塔
cfg.topDownChannelsNum = 96
cfg.bottomUp2ChannelsNum = 96
cfg.outChannelsNum = [96, 96, 96] #FPN每层输出的channel 数目
class PAN(nn.Module):
def __init__(self, cfg):
super(PAN, self).__init__()
self.featureMapNum = cfg.featureMapNum
self.bottomUpChannelsNum = cfg.bottomUpChannelsNum
self.topDownChannelsNum = cfg.topDownChannelsNum
from net.shuffleNetV2 import ShuffleNetV2
from net.PAN_simplest import nanodet_PAN
from net.header import Head
from torch import nn
from easydict import EasyDict as Edict
cfgMe = Edict()
cfgMe.modelSize = '1.0x'
cfgMe.activation = 'LeakReLu'
class NanoNet(nn.Module):
def __init__(self, classNum, imgChannelNumber, regBoxNum):
super(NanoNet, self).__init__()
self.backbone = ShuffleNetV2(imgChannelNumber,
model_size='1.0x',
activation='LeakyReLU')
self.neck = nanodet_PAN( #featureMapNum=3,
#bottomUpChannelsNum =[116,232,464],
#topDownChannelsNum = 96,
#bottomUp2ChannelsNum = 96,
#outChannelsNum = 96
inputChanNum=[116, 232, 464],
chanNum=96)
self.clsNum = classNum
self.regBoxNum = regBoxNum
self.head = nn.ModuleList()
for i in range(3):
h = Head(
reg_max=self.
regBoxNum, # defalut =8个bbox,用于分布, general focal loss format
from easydict import EasyDict as Edict
import argparse
edict = Edict()
edict.data = Edict()
edict.data.LAPSFormat = "/home/aiw/mntData/data168/data/SR_LAPS1H/%s/%s/MSP1_PMSC_AIWSRPF_LAPS-1H-0p01_L88_CZ_%s00_00000-00000.nc"
edict.data.LAPSNPY = './dataset/InputData/%s.npy'
edict.data.GRAPESMESOFormat = "/home/aiw/mntData/data47/data/AIWSLL/GRAPESMESOORI/%s/%s/MSP2_PMSC_AIWSR%sF_GRAPESMESOORI-0P03-1H_L88_CZ_%s00_00000-01200.nc"
edict.data.GRAPESMESOVariable = '%s-component_of_wind_height_above_ground'
edict.data.GRAPESMESONPY = './dataset/TestData/%s.npy'
edict.checkpoint = Edict()
edict.checkpoint.cpkRoot = './checkpoint_%s'
def options():
parser = argparse.ArgumentParser(description='PyTorch Super Res Example')
parser.add_argument('--batchSize',
type=int,
default=32,
help='training batch size')
parser.add_argument('--nEpochs',
type=int,
default=40,
help='number of epochs to train for')
parser.add_argument('--lr',
type=float,
default=0.01,
help='Learning Rate. Default=0.001')
parser.add_argument('--seed',