Quick Start

Quick Start#

This chapter uses dvs-mnist as an example to illustrate the usage of the BIDL project.

Choose Configuration File#

The dvs-mnist dataset uses the clif3fc3dm_itout-b16x1-dvsmnist.py configuration file. Since dvs-mnist belongs to the classification of dvs data, create a dvs_mnist folder in the applications/classification/dvs directory. Based on its model structure, create a clif3fc3dm folder under the new dvs_mnist folder, and finally place the configuration file in this path.

Attention

Configuration files are Python source code files, and their filenames have a .py suffix.

Prepare Dataset#

Check the dataset path set in the configuration file chosen in Choose Configuration File, and then create the corresponding dataset directory in the project’s root directory.

data = dict(
   samples_per_gpu=16,
   workers_per_gpu=2,
   train=dict(
      pipeline=train_pipeline
   ),
   val=dict(
      pipeline=test_pipeline,
      test_mode=True
   ),
   test=dict(
      pipeline=test_pipeline,
      test_mode=True
   )
)

Model Training#

In the tools directory, execute the following command to start model training.

python train.py --config clif3fc3dm_itout-b16x1-dvsmnist

Attention

The parameter --config followed by clif3fc3dm_itout-b16x1-dvsmnist is the name of the configuration file selected in the Choose Configuration File section, without the .py.

After training starts, the configuration information involved in this training will be printed in sequence:

  • Configuration information from the configuration file

  • The path information for saving checkpoints.

    By default, checkpoint files are saved in the project root directory /work_dirs/clif3fc3dm_itout-b16x1-dvsmnist/, where a latest.pth file is generated for the last saved checkpoint of the epoch, and a best.pth file is created to save the checkpoint with the best accuracy on the validation set. The directory tree is as follows:

    clif3fc3dm_itout-b16x1-dvsmnist/
├── 20240829_171552.log #All logs in this training process, such as loss, accuracy
├── best.pth #Checkpoint of the model with the highest accuracy on the validation set
└── latest.pth #Checkpoint of the model from the last training round

  • The following contents are model training log information. Below is the log information output in the first two epochs:

2024-08-29 17:33:37,021 - cls - INFO - Epoch [1][100/452] lr:
2.060e-02 top-1: 22.7500 top-5: 62.5000 loss: 2.1440

2024-08-29 17:33:48,696 - cls - INFO - Epoch [1][200/452] lr:
4.040e-02 top-1: 73.8750 top-5: 96.8125 loss: 1.1264

2024-08-29 17:34:00,255 - cls - INFO - Epoch [1][300/452] lr:
6.020e-02 top-1: 89.1875 top-5: 98.6250 loss: 0.8457

2024-08-29 17:34:11,309 - cls - INFO - Epoch [1][400/452] lr:
8.000e-02 top-1: 90.6250 top-5: 99.1250 loss: 0.7871

2024-08-29 17:34:16,431 - cls - INFO - Saving checkpoint at 1 epochs

2024-08-29 17:34:21,611 - cls - INFO - Epoch(val) [1][20] top-1:
87.1275 top-5: 98.9238

2024-08-29 17:34:31,200 - cls - INFO - Epoch [2][100/452] lr:
9.980e-02 top-1: 94.3750 top-5: 99.6250 loss: 0.7186

2024-08-29 17:34:41,470 - cls - INFO - Epoch [2][200/452] lr:
9.980e-02 top-1: 94.6875 top-5: 99.6250 loss: 0.6894

2024-08-29 17:34:51,961 - cls - INFO - Epoch [2][300/452] lr:
9.980e-02 top-1: 97.3125 top-5: 99.8750 loss: 0.6128

2024-08-29 17:35:02,393 - cls - INFO - Epoch [2][400/452] lr:
9.980e-02 top-1: 97.3750 top-5: 99.9375 loss: 0.5926

2024-08-29 17:35:07,965 - cls - INFO - Saving checkpoint at 2 epochs

2024-08-29 17:35:14,549 - cls - INFO - Epoch(val) [2][20] top-1:
98.1374 top-5: 100.0000

After the training, copy the best weight file to the weight file path in the resource package: /weight_files/classification/clif3fc3dm/lif

Model Inference (GPU)#

In the tools directory, run the following command for model inference.

python test.py --config clif3fc3dm_itout-b16x1-dvsmnist --checkpoint latest.pth --use_lyngor 0 --use_legacy 0

The inference process is as follows:

100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2409/2409 [01:23<00:00, 28.85it/s]
acc top-1: 99.21 top-5: 100.00
gpu test speed = 576.6303 fps
gpu test speed (without pipeline) = 583.3422 fps

Since the checkpoint files’ save path is fixed by default during training, and the subdirectory is the same as the configuration file name, the parameter for --checkpoint only needs to specify the pth filename.

Model Inference (Lynxi brain-inspired systems)#

In the BIDL framework, a switch is used to toggle the backend execution environment. Specifically, when using test.py, configure as follows:

Configuration

Description

--use_lyngor

Used to specify whether to use Lyngor for compilation.

[Data Type] ENUM

[Value Range]

  • 1: Use Lyngor for compilation, executed with Lynxi brain-inspired systems;

  • 0: Do not use Lyngor for compilation (GPU execution).

--use_legacy

Used to specify whether to directly load historical compilation artifacts, skipping the compilation step.

[Data Type] ENUM

[Value Range]

  • 1: Load historical compilation artifacts, skip compilation step;

  • 0: Do not load historical compilation artifacts, do not skip compilation step.

Like this:

python test.py --config clif3fc3dm_itout-b16x1-dvsmnist --checkpoint latest.pth --use_lyngor 1 --use_legacy 0

Attention

You need to copy the trained files (located in work_dirs) and the validation dataset to Lynxi brain-inspiredcomputing devices before executing this script.

Contents