update 2rd training

Author: Sharpiless

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Jie Tang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

augs.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python
+# -*- coding:utf-8 -*-
+# @Filename:    augs.py
+# @Project:     BP-Net
+# @Author:      jie
+# @Time:        2021/3/14 8:27 PM
+
+
+import numpy as np
+
+__all__ = [
+    'Compose',
+    'Norm',
+    'Jitter',
+    'Flip',
+]
+
+
+class Compose(object):
+    """
+    Sequential operations on input images, (i.e. rgb, lidar and depth).
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, rgb, disp, lidar, depth, K_cam):
+        for t in self.transforms:
+            rgb, disp, lidar, depth, K_cam = t(rgb, disp, lidar, depth, K_cam)
+        return rgb, disp, lidar, depth, K_cam
+
+
+class Norm(object):
+    """
+    normalize rgb image.
+    """
+
+    def __init__(self, mean, std):
+        self.mean = np.array(mean)
+        self.std = np.array(std)
+
+    def __call__(self, rgb, disp, lidar, depth, K_cam):
+        rgb = (rgb - self.mean) / self.std
+        disp = disp / 255
+        return rgb, disp, lidar, depth, K_cam
+
+class Jitter(object):
+    """
+    borrow from https://github.com/kujason/avod/blob/master/avod/datasets/kitti/kitti_aug.py
+    """
+
+    def __call__(self, rgb, disp, lidar, depth, K_cam):
+        pca = compute_pca(rgb)
+        rgb = add_pca_jitter(rgb, pca)
+        return rgb, disp, lidar, depth, K_cam
+
+
+
+class Flip(object):
+    """
+    random horizontal flip of images.
+    """
+
+    def __call__(self, rgb, disp, lidar, depth, CamK):
+        width = rgb.shape[1]
+        flip = bool(np.random.randint(2))
+        if flip:
+            rgb = rgb[:, ::-1, :]
+            lidar = lidar[:, ::-1, :]
+            depth = depth[:, ::-1, :]
+            CamK[0, 2] = (width - 1) - CamK[0, 2]
+        return rgb, disp, lidar, depth, CamK
+
+
+def compute_pca(image):
+    """
+    calculate PCA of image
+    """
+
+    reshaped_data = image.reshape(-1, 3)
+    reshaped_data = (reshaped_data / 255.0).astype(np.float32)
+    covariance = np.cov(reshaped_data.T)
+    e_vals, e_vecs = np.linalg.eigh(covariance)
+    pca = np.sqrt(e_vals) * e_vecs
+    return pca
+
+
+def add_pca_jitter(img_data, pca):
+    """
+    add a multiple of principle components with Gaussian noise
+    """
+    new_img_data = np.copy(img_data).astype(np.float32) / 255.0
+    magnitude = np.random.randn(3) * 0.1
+    noise = (pca * magnitude).sum(axis=1)
+
+    new_img_data = new_img_data + noise
+    np.clip(new_img_data, 0.0, 1.0, out=new_img_data)
+    new_img_data = (new_img_data * 255).astype(np.uint8)
+
+    return new_img_data

checkpoints/.gitignore

@@ -0,0 +1,4 @@
+# Ignore everything in this directory
+*
+# Except this file
+!.gitignore

configs/config.yaml

@@ -0,0 +1,28 @@
+defaults:
+  - hydra
+  - optim: AdamW
+  - sched/lr: NoiseOneCycleCosMo
+  - data: KITTI
+  - net: PMP
+  - loss: MSMSE
+  - metric: RMSE
+  - _self_
+#
+
+nepoch: 30
+test_epoch: 25
+test_iter: 1000
+lr: 0.001
+gpus:
+  - 0
+train_batch_size: 2
+test_batch_size: 2
+num_workers: 4
+manual_seed: 1
+vis_iter: 100
+start_epoch: 0
+gpu_id: ???
+name: ???
+device:
+  _target_: torch.device
+  device: cuda:${gpu_id}

configs/data/KITTI.yaml

@@ -0,0 +1,35 @@
+mean:
+  - 90.9950
+  - 96.2278
+  - 94.3213
+std:
+  - 79.2382
+  - 80.5267
+  - 82.1483
+train_sample_size: 85898
+test_sample_size: 1000
+height: 256
+width: 1216
+
+trainset:
+  _target_: datasets.KITTI
+  mode: 'train'
+  RandCrop: true
+  path: ${data.path}
+  height: ${data.height}
+  width: ${data.width}
+  mean: ${data.mean}
+  std: ${data.std}
+
+testset:
+  _target_: datasets.KITTI
+  mode: 'selval'
+  RandCrop: false
+  path: ${data.path}
+  height: ${data.height}
+  width: ${data.width}
+  mean: ${data.mean}
+  std: ${data.std}
+
+path: datas/kitti
+mul_factor: 1.0

configs/data/NYU.yaml

@@ -0,0 +1,38 @@
+mean:
+  - 117.
+  - 97.
+  - 91.
+std:
+  - 70.
+  - 71.
+  - 74.
+train_sample_size: 47584
+test_sample_size: 654
+height: 256
+width: 320
+
+
+trainset:
+  _target_: datasets.NYU
+  mode: 'train'
+  path: ${data.path}
+  num_sample: ${data.npoints}
+  mul_factor: ${data.mul_factor}
+  num_mask: ${data.num_mask}
+  scale_kcam: true
+
+
+testset:
+  _target_: datasets.NYU
+  mode: 'val'
+  path: ${data.path}
+  num_sample: ${data.npoints}
+  mul_factor: ${data.mul_factor}
+  num_mask: ${data.num_mask}
+  scale_kcam: false
+
+
+path: datas/nyu
+npoints: 500
+mul_factor: 10.0
+num_mask: 1

configs/hydra.yaml

@@ -0,0 +1,10 @@
+hydra:
+  run:
+    dir: outputs/${name}/${now:%Y-%m-%d_%H-%M-%S}
+  output_subdir: .
+  job:
+    chdir: false
+#  job_logging:
+#    handlers:
+#      file:
+#        filename: ${name}.log

configs/loss/MSMSE.yaml

@@ -0,0 +1 @@
+_target_: criteria.MSMSE

configs/metric/MetricALL.yaml

@@ -0,0 +1,2 @@
+_target_: criteria.MetricALL
+mul_factor: ${data.mul_factor}

configs/metric/RMSE.yaml

@@ -0,0 +1,2 @@
+_target_: criteria.RMSE
+mul_factor: ${data.mul_factor}

configs/net/PMP.yaml

@@ -0,0 +1,16 @@
+# Propagation Field Baseline Relu6 Clip EMA
+
+name: PMP
+
+model:
+  _target_: models.Pre_MF_Post
+
+
+ema:
+  _target_: models.EMA
+  decay: 0.9999
+
+
+clip:
+  _target_: utils.clip_grad_norm_
+  max_norm: 0.1

configs/optim/AdamW.yaml

@@ -0,0 +1,6 @@
+_target_: optimizers.AdamW
+lr: ${lr}
+weight_decay: 0.05
+betas:
+  - 0.9
+  - 0.999

configs/sched/lr/NoiseOneCycleCosMo.yaml

@@ -0,0 +1,29 @@
+#Noise One Cycle Cos With cycle_momentum
+policy:
+  _target_: schedulers.NoiseLR
+  lr_sched: OneCycleLR
+  anneal_strategy: cos
+  epochs: ${nepoch}
+  div_factor: 40.0
+  final_div_factor: 0.1
+  last_epoch: -1
+  max_lr: ${lr}
+  cycle_momentum: true
+  base_momentum: 0.85
+  max_momentum: 0.95
+  pct_start: 0.1
+  noise_pct: 0.1
+  steps_per_epoch: ${sched.lr.steps_per_epoch}
+
+iter: true
+
+steps_per_epoch:
+  _target_: utils.FloorDiv
+  _args_:
+    - _target_: utils.CeilDiv
+      _args_:
+        - ${data.train_sample_size}
+        - _target_: builtins.len
+          _args_:
+            - ${gpus}
+    - ${train_batch_size}