Download pkg from https://apps.cloud.blackmagicdesign.com/davinci-resolve

1
2
3
4

git clone https://aur.archlinux.org/davinci-resolve.git
cd davinci-resolve
mv ~/Downloads/DaVinci_Resolve_19.1.4_Linux.zip ./
makepkg -si

Q&A

1

What is standard dense supervised learning? Mentioned in [[CenterNet]].

Standard dense supervised learning typically refers to a supervised learning setup where:

1. Standard supervised learning means:
   • You have input data X and corresponding ground truth labels Y.
   • The goal is to train a model $f_\theta(X)$ that maps inputs to outputs by minimizing a loss function (e.g., cross-entropy, MSE) between the predicted labels and ground truth.
   • The training dataset is fully labeled (i.e., each input has a corresponding label).
2. Dense refers to:
   • A per-pixel or per-element prediction task, where every element in the input gets a corresponding label.
   • Common in vision tasks like:
     • Semantic segmentation (each pixel is labeled with a class).
     • Depth estimation (each pixel has a depth value).
     • Optical flow (each pixel has a motion vector).
     • Surface normal estimation (each pixel has a 3D orientation vector).

In contrast to sparse supervision, where only a subset of the input (e.g., bounding boxes, keypoints) is labeled, dense supervision provides full annotations for every relevant part of the input.

Example
In semantic segmentation:

• Input: an RGB image (e.g., 512×512 pixels).
• Output: a label map of the same size (512×512), where each pixel has a class label like “road”, “car”, “sky”, etc.
• Model: often a Fully Convolutional Network (FCN) or encoder-decoder like U-Net or DeepLab.
• Loss: usually pixel-wise cross-entropy.

FCSGG Repository Summary

FCSGG (Fully Convolutional Scene Graph Generation) is a PyTorch implementation of the paper “Fully Convolutional Scene Graph Generation” published in CVPR 2021. The project focuses on scene graph generation, which is the task of detecting objects in an image and identifying the relationships between them.

Core Components:

1. Architecture:

   • Built on Detectron2, a popular object detection framework by Facebook
   • Uses a one-stage detector approach (CenterNet) as the meta-architecture
   • Supports various backbones including ResNet, HRNet (High-Resolution Network), Hourglass networks, and DLA

2. Key Features:

   • Fully convolutional approach to scene graph generation
   • Multiple backbone options with different feature pyramid networks (FPN, BiFPN, HRFPN)
   • Various head designs including multiscale heads and attention mechanisms
   • Support for different input resolutions and training strategies

3. Dataset:

   • Primarily designed for the Visual Genome dataset, a large-scale dataset for scene understanding
   • Includes custom data loaders and preprocessing for scene graph generation

4. Model Components:

   • Backbones: Various CNN architectures (ResNet, HRNet, Hourglass, DLA)
   • Necks: Feature pyramid networks and variants (FPN, BiFPN, HRFPN, Trident)
   • Heads: Detection and relationship prediction heads
   • Loss Functions: Custom losses for object detection and relationship prediction

5. Utilities:

   • Visualization tools for scene graphs
   • Evaluation metrics for scene graph generation
   • Training and inference scripts

Project Structure:

• fcsgg/: Main module containing model implementation

  • modeling/: Neural network architecture components
    • backbone/: Feature extraction networks
    • necks/: Feature pyramid networks
    • heads/: Detection and relationship prediction heads
    • meta_arch/: High-level model architecture (CenterNet)
  • data/: Dataset handling and preprocessing
  • evaluation/: Metrics and evaluation code
  • utils/: Helper functions and utilities
  • layers/: Custom neural network layers
  • structures/: Data structures for scene graphs

• configs/: Configuration files for different model variants

• tools/: Training, evaluation, and visualization scripts

• GraphViz/: Visualization tools for scene graphs

Key Innovations:

The project implements a fully convolutional approach to scene graph generation, which differs from traditional two-stage methods. Instead of first detecting objects and then predicting relationships, it uses a one-stage detector to simultaneously predict objects and their relationships in a fully convolutional manner.

Benchmarks:

The repository provides several pre-trained models with different backbones:

1. HRNetW32-1S
2. ResNet50-4S-FPN×2
3. HRNetW48-5S-FPN×2

These models achieve competitive performance on the Visual Genome dataset for scene graph generation tasks.

Usage:

The project provides tools for training, evaluation, and visualization of scene graphs. It requires the Visual Genome dataset and can be run using Docker or directly with PyTorch.

In summary, FCSGG is a comprehensive implementation of a state-of-the-art approach to scene graph generation using fully convolutional networks, offering various model architectures and training configurations.

How Detectron2 is Used in FCSGG

FCSGG is built on top of Detectron2, Facebook’s object detection framework, and leverages many of its components while extending it for scene graph generation. Here’s a detailed breakdown:

1. Core Architecture Integration

• Meta Architecture: FCSGG registers a custom meta architecture called “CenterNet” with Detectron2’s META_ARCH_REGISTRY. This extends Detectron2’s modular architecture system while maintaining compatibility.

• Backbone Networks: FCSGG uses Detectron2’s backbone networks (ResNet, etc.) directly and also implements custom backbones like HRNet while following Detectron2’s backbone interface.

• Feature Pyramid Networks (FPN): The repository uses Detectron2’s FPN implementation and extends it with custom variants like BiFPN and HRFPN.

2. Configuration System

• YAML Configuration: FCSGG adopts Detectron2’s YAML-based configuration system, extending it with custom configurations for scene graph generation through add_fcsgg_config().

• Command Line Arguments: The training script uses Detectron2’s default_argument_parser() to maintain the same command-line interface.

3. Data Handling

• Dataset Registration: Visual Genome dataset is registered with Detectron2’s DatasetCatalog and MetadataCatalog, making it available through Detectron2’s data loading pipeline.

• Custom Dataset Mapper: FCSGG implements a custom DatasetMapper class that extends Detectron2’s mapper to handle scene graph annotations.

• Data Loaders: The repository uses Detectron2’s build_detection_train_loader and build_detection_test_loader with custom mappers.

4. Training and Evaluation

• Trainer Class: FCSGG extends Detectron2’s DefaultTrainer class to customize the training loop, evaluation metrics, and data loading.

• Checkpointing: The repository uses Detectron2’s DetectionCheckpointer for model saving and loading.

• Distributed Training: FCSGG leverages Detectron2’s distributed training utilities through detectron2.utils.comm and the launch function.

• Custom Evaluators: The repository implements a custom VGEvaluator for scene graph evaluation while following Detectron2’s evaluator interface.

5. Visualization and Logging

• Event Storage: FCSGG uses Detectron2’s event storage system for logging metrics during training.

• Visualization Tools: The repository leverages Detectron2’s visualization utilities for debugging and result analysis.

6. Extensions for Scene Graph Generation

• Custom Heads: While using Detectron2’s architecture, FCSGG implements custom prediction heads for relationship detection.

• Scene Graph Structures: The repository defines custom data structures for scene graphs that integrate with Detectron2’s Instances class.

• Loss Functions: FCSGG implements specialized loss functions for scene graph generation while maintaining compatibility with Detectron2’s loss computation framework.

7. Installation and Dependencies

• Submodule Integration: Detectron2 is included as a Git submodule, ensuring version compatibility.

• Build Process: The installation process includes building Detectron2 from source to ensure proper integration.

In summary, FCSGG uses Detectron2 as its foundation, leveraging its modular architecture, data handling, training infrastructure, and configuration system while extending it with custom components for scene graph generation. This approach allows FCSGG to benefit from Detectron2’s robust implementation and optimizations while adding specialized functionality for relationship detection between objects.

Official repo:
https://github.com/liuhengyue/fcsgg
Our repo:
https://github.com/PSGBOT/KAF-Generation

My venv: fcsgg

Installation

Environment Preparation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

git clone git@github.com:liuhengyue/fcsgg.git
cd fcsgg
git submodule init
git submodule update
conda create --name fcsgg
conda create -n fcsgg python=3.10
conda install nvidia/label/cuda-11.8.0::cuda-toolkit -c nvidia/label/cuda-11.8.0
conda install cudatoolkit
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# for building detectron
conda install -c conda-forge gcc=11.2.0
conda install -c conda-forge gxx=11.2.0

conda env config vars set LD_LIBRARY_PATH="/home/cyl/miniconda3/envs/fcsgg/lib/"
conda env config vars set CPATH="/home/cyl/miniconda3/envs/fcsgg/include/"
conda env config vars set CUDA_HOME="/home/cyl/miniconda3/envs/fcsgg/"

conda deactivate
conda activate fcsgg

export CC=$CONDA_PREFIX/bin/gcc
export CXX=$CONDA_PREFIX/bin/g++

pip install -r requirements.txt
python -m pip install -e detectron2

Downloads

Datasets:

1
2
3
4
5

cd ~/Reconst
wget https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip -P ./Data/vg/
wget https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip -P ./Data/vg/
unzip -j ./Data/vg/images.zip -d ./Data/vg/VG_100K
unzip -j ./Data/vg/images2.zip -d ./Data/vg/VG_100K

Download the scene graphs and extract them to datasets/vg/VG-SGG-with-attri.h5.

Issues

1

1

AttributeError: module 'PIL.Image' has no attribute 'LINEAR'. Did you mean: 'BILINEAR'?

LINEAR-> BILINEAR: commit

2

在尝试训练的过程中报错：

1
2
3

  File "/home/cyl/Reconst/fcsgg/fcsgg/data/detection_utils.py", line 432, in generate_score_map
    masked_fmap = torch.max(masked_fmap, gaussian_mask * k)
RuntimeError: The size of tensor a (55) must match the size of tensor b (56) at non-singleton dimension 1

modify detection_utils.py: commit

Training

首先更改训练的配置文件./config/quick_schedules/Quick-FCSGG-HRNet-W32.yaml, (原文件使用预训练的参数)

1
2
3
4

MODEL:
  META_ARCHITECTURE: "CenterNet"
  HRNET:
    WEIGHTS: "output/FasterR-CNN-HR32-3x.pth"

更改为train from scratch

1
2
3
4

MODEL:
  META_ARCHITECTURE: "CenterNet"
  HRNET:
    WEIGHTS: ""  # Empty string to train from scratch

再运行：

1

python tools/train_net.py --num-gpus 1 --config-file configs/quick_schedules/Quick-FCSGG-HRNet-W32.yaml

成功训练✌

1
2
3
4
5
6

...
[04/23 10:21:01] d2.utils.events INFO:  eta: 0:05:37  iter: 1159  total_loss: 1.042  loss_cls: 0.7593  loss_box_wh: 0.08827  loss_center_reg: 0.02485  loss_raf: 0.1754    time: 0.4042  last_time: 0.4519  data_time: 0.0043  last_data_time: 0.0044   lr: 0.001  max_mem: 4141M
[04/23 10:21:09] d2.utils.events INFO:  eta: 0:05:29  iter: 1179  total_loss: 1.028  loss_cls: 0.7208  loss_box_wh: 0.09246  loss_center_reg: 0.02669  loss_raf: 0.1625    time: 0.4042  last_time: 0.4035  data_time: 0.0041  last_data_time: 0.0044   lr: 0.001  max_mem: 4141M
[04/23 10:21:17] d2.utils.events INFO:  eta: 0:05:21  iter: 1199  total_loss: 1.01  loss_cls: 0.671  loss_box_wh: 0.1038  loss_center_reg: 0.02432  loss_raf: 0.1635    time: 0.4042  last_time: 0.3943  data_time: 0.0042  last_data_time: 0.0043   lr: 0.001  max_mem: 4141M
[04/23 10:21:25] d2.utils.events INFO:  eta: 0:05:13  iter: 1219  total_loss: 0.9737  loss_cls: 0.6887  loss_box_wh: 0.0929  loss_center_reg: 0.02574  loss_raf: 0.1749    time: 0.4041  last_time: 0.4101  data_time: 0.0041  last_data_time: 0.0042   lr: 0.001  max_mem: 4141M
...

Explanation

See [[FCSGG Repo Explanation]]