Bad training result after 90 epochs

[rginjapan]

Due to my GPU limitation, parameters are: sequence-size: 3 # must be >= 1 combinations: [ [ 0, 1 ],[0,2],[1,2] ] batch size:2 epochs:90

I trained the lidar with imu network, using all the training dataset in original config. I also plot the loss curve and test result in 2011-10-03: [27] for you to check.

What wrong with my implementation?

Looking forward to your reply, thanks ahead!

[rginjapan]

I used only one sequence to train and valid (I know it will overfit), through the loss curve, it already convergenced, however the trajectory is still very bad... Could you give me some instructions about how to train and evaluate the loss curve? Thanks ahead!!!

[rginjapan]

@ArashJavan sorry for disturbing, if you did not know the reason, could you please give me your training hyper parameters.

[ArashJavan]

@rginjapan your sequence is not correct for an RNN-based architecture. The combination should be:

combinations: [ [ 0, 1 ],[1, 2],[2,3] ]

Also, I would first try with only one Network for example only IMU to see if everything works fine. Then switch to more complex networks with both lidar and imu fused!

Another hint is about your grad-norm, if the above points did not help try to use gradient-clipping.

[rginjapan]

@ArashJavan Thanks for your reply, I think "sequences" means images numbers, for example, sequences:3 represents 3 continuous frames I0, I1 and I2. I remember you explain in other close issue, "combinations [0,2]" will increase the relative motion compared with "combination [0,1]" , which is good for training. Am I correct ?

"Another hint is about your grad-norm, if the above points did not help try to use gradient-clipping." Sorry, can you explain in more details, I did not understand.

[ArashJavan]

@rginjapan yes and no ;) for RNN-based network, it is better to have consecutive frame combination, e.g. [[0,1], [1,2], [2,3]].

Regarding gradient-clipping: see here, and here.

[rginjapan]

@ArashJavan Thanks for your quick reply. Right now I understand, I already used [0,1],[0,2],[1,2] to get good result in only lidar network, so then I thought it will work on lidar+imu network, but not... So, do you evaluate the difference between large sequence number like 5 in your config and samll number like 2? Since [0,1] and [4,5] is similar in training, right?

[ArashJavan]

@rginjapan plz look at the loss function here. As you can see in the loss function we already take care for global motion since the loss is calculated for local - e.g. [x_i, x_i+1] and global -e.g. [x0, x1],...,[x0, x_seq].

[rginjapan]

After I change the combinations and training 100 epochs, the result is still not good. What shoud I do, increasing the epochs numbers ? What is the ideal convergenced loss?? right now the loss is around 16 and does not decrease.

[ArashJavan]

@rginjapan Can you share your config-file and your meta-data settings (lr, wd and so on).

[rginjapan]

config.yaml

datasets: sequence-size: 3 # must be >= 1 combinations: [ [ 0, 1 ], [1,2], [2,3]]

kitti: root-path-sync: "datasets/KITTI/sync" root-path-unsync: "datasets/KITTI/extract" image-width: 720 image-height: 57 crop-factors: [0, 0] # [0, 4] # top, left fov-up: 3. fov-down: -25. max-depth: 80. min-depth: 1. inverse-depth: true train: 2011-10-03: [ 27, 42, 34 ] 2011-09-30: [ 16, 18, 20, 27, 28 ] test:

  2011-10-03: [ 27, 42, 34 ]
  2011-09-30: [ 16, 18, 20, 27, 28, 33, 34 ]
validation:
  #2011-09-26: [23, 39]
  2011-09-30: [ 33, 34 ]

  # channesl: x, y, z, remission, nx, ny ,nz, range)
mean-image: [-0.0014, 0.0043, -0.011, 0.2258, -0.0024, 0.0037, 0.3793, 0.1115]
std-image: [0.1269, 0.0951, 0.0108, 0.1758, 0.3436, 0.4445, 0.5664, 0.0884]

mean-imu: [-0.0685, 0.1672, 9.7967, -0., 0.0006, 0.0059]
std-imu: [0.8766, 0.9528, 0.3471, 0.0204, 0.0227, 0.1412]

DeepLIO Network

deeplio: dropout: 0.25 pretrained: true model-path: "/home/deeplio/outputs/train_deeplio_20210806_174151/cpkt_deeplio_best.tar" lidar-feat-net: name: "lidar-feat-pointseg" pretrained: true model-path: "/home/deeplio/outputs/train_deeplio_20210806_174151/cpkt_lidarpointsegfeat_best.tar" requires-grad: true imu-feat-net: name: "imu-feat-rnn" pretrained: true model-path: "/home/deeplio/outputs/train_deeplio_20210806_174151/cpkt_imufeatrnn0_best.tar" requires-grad: true odom-feat-net: name: "odom-feat-rnn" pretrained: true model-path: "/home/deeplio/outputs/train_deeplio_20210806_174151/cpkt_odomfeatrnn_best.tar" requires-grad: true fusion-net: name: "fusion-layer-soft" pretrained: true model-path: "/home/deeplio/outputs/train_deeplio_20210806_174151/cpkt_deepliofusionsoft_best.tar" requires-grad: true # only soft-fusion has trainable params

Lidar Feature Netowrks

lidar-feat-pointseg: # pointseg feature dropout: 0.1 classes: ['unknown', 'object'] bypass: "simple" fusion: add # [cat, sub, add] part: "encoder" # [encoder]

lidar-feat-flownet: dropout: 0. fusion: add # [cat, sub, add]

lidar-feat-resnet: dropout: 0.25 fusion: add # [cat, sub, add]

lidar-feat-simple-1: dropout: 0.25 fusion: add # [cat, sub, add] bypass: false

imu-feat-fc: # FC input-size: 6 # !fixed! do not chanage hidden-size: [128, 256, 512, 512, 256, 128] dropout: 0.

imu-feat-rnn: # RNN type: "lstm" input-size: 6 # !fixed! do not chanage hidden-size: 128 num-layers: 2 bidirectional: true dropout: 0.1

fusion-layer-soft: type: "soft"

Odometry Feature Netowrks

odometry feature network with fully connected layers

odom-feat-fc: size: [1024, 512, 256] dropout: 0.

odometry feature network with rnn-layers

odom-feat-rnn: type: "lstm" hidden-size: 1024 num-layers: 2 bidirectional: true dropout: 0.

Loss Configurations

losses: active: 'hwsloss' hwsloss: params: learn: true sx: 0. sq: -3. lwsloss: params: beta: 1125. loss-type: "local+global" # ["local", "global", "local+global"]

current-dataset: 'kitti' channels: [0, 1, 2, 4, 5, 6] # channesl: x, y, z, remission, nx, ny. nz range optimizer: 'adam'

hyper params in train.py

parser.add_argument('-j', '--workers', default=4, type=int, metavar='N', help='number of data loading workers (default: 4)') parser.add_argument('--epochs', default=100, type=int, metavar='N', help='number of total epochs to run') parser.add_argument('-b', '--batch-size', default=2, type=int, metavar='N', help='mini-batch size (default: 1), this is the total ' 'batch size of all GPUs on the current node when ' 'using Data Parallel or Distributed Data Parallel') parser.add_argument('--start-epoch', default=0, type=int, metavar='N', help='manual epoch number (useful on restarts)') parser.add_argument('--lr', '--learning-rate', default=1e-3, type=float, metavar='LR', help='initial learning rate', dest='lr') parser.add_argument('--lr-decay', '--learning-rate-decay', default=30, type=int, metavar='LR-DECAY-STEP', help='learning rate decay step', dest='lr_decay') parser.add_argument('--momentum', default=0.9, type=float, metavar='M', help='momentum') parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float, metavar='W', help='weight decay (default: 1e-4)', dest='weight_decay') parser.add_argument('-p', '--print-freq', default=10, type=int, metavar='N', help='print frequency (default: 10)') parser.add_argument('--resume', default=False, action='store_true', dest='resume', help='resume training') parser.add_argument('-e', '--evaluate', default='', type=str, metavar='PATH', help='evaluate model with given checkpoint on validation set (default: none)') parser.add_argument('-c', '--config', default="./config.yaml", help='Path to configuration file') parser.add_argument('-d', '--debug', default=False, help='debug logging', action='store_true', dest='debug') parser.add_argument('--device', default='cuda', type=str, metavar='DEVICE', help='Device to use [cpu, cuda].')

[rginjapan]

Could you please upload your trained model to let me try to get some good result, it is not to be as good as your paper's result.

P.S. I increased the epochs to 150, the loss can not be decreased, still around 16. And also the result is still bad... Looking forward to your guidance.

[huazai665]

Due to my GPU limitation, parameters are: sequence-size: 3 # must be >= 1 combinations: [ [ 0, 1 ],[0,2],[1,2] ] batch size:2 epochs:90

I trained the lidar with imu network, using all the training dataset in original config. I also plot the loss curve and test result in 2011-10-03: [27] for you to check.

What wrong with my implementation?

Looking forward to your reply, thanks ahead! hello, I have the same question. The result is bad and the loss is negative. Do you solve the problem?

[rginjapan]

@huazai665 I cannot reproduce the result in the paper in the end.