configs

File Structure and Naming

This folder contains training recipes and model readme files for each model. The folder structure and naming rule of model configurations are as follows.

    ├── configs
        ├── model_a                         // model name in lower case with _ seperator
        │   ├─ model_a_small_ascend.yaml    // training recipe denated as {model_name}_{specification}_{hardware}.yaml
        |   ├─ model_a_large_gpu.yaml
        │   ├─ README.md                    //readme file containing performance results and pretrained weight urls
        │   └─ README_CN.md                 //readme file in Chinese
        ├── model_b
        │   ├─ model_b_32_ascend.yaml
        |   ├─ model_l_16_ascend.yaml
        │   ├─ README.md
        │   └─ README_CN.md
        ├── README.md //this file

Note: Our training recipes are verified on specific hardware, and the suffix hardware (ascend or gpu) in the file name of training recipes indicates different hardware. Since Mindspore operators have different precision and performance on different hardware, different training recipes are required under different hardware. However, if you want to train on another hardware (e.g. GPU) using the training recipe under specific hardware (e.g. Ascend), you only need to make minor or no adjustments to the hyperparameters, because the training recipe has a certain degree of generalization across different hardware.

Model Readme Writing Guideline

The model readme file in each sub-folder provides the introduction, reproduced results, and running guideline for each model.

Please follow the outline structure and table format shown in densenet/README.md when contributing your models :)

Table Format

model name	params(M)	cards	batch size	resolution	jit level	graph compile	ms/step	img/s	acc@top1	acc@top5	recipe	weight
densenet121	8.06	8	32	224x224	O2	300s	47,34	5446.81	75.67	92.77	yaml	weights

Illustration:

• model name: model name in lower case with _ seperator.
• top-1 and top-5: Accuracy reported on the validatoin set of ImageNet-1K. Keep 2 digits after the decimal point.
• params(M): # of model parameters in millions (10^6). Keep 2 digits after the decimal point
• batch size: Training batch size
• cards: # of cards
• ms/step: Time used on training per step in ms
• jit level: Jit level of mindspore context, which contains 3 levels: O0/O1/O2
• recipe: Training recipe/configuration linked to a yaml config file.
• weight: url of the pretrained model weights

Model Checkpoint Format

The checkpoint (i.e., model weight) name should follow this format: {model_name}_{specification}-{sha256sum}.ckpt, e.g., poolformer_s12-5be5c4e4.ckpt.

You can run the following command and take the first 8 characters of the computing result as the sha256sum value in the checkpoint name.

sha256sum your_model.ckpt

Training Script Format

For consistency, it is recommended to provide distributed training commands based on msrun --bind_core=True --worker_num {num_devices} python train.py, instead of using shell script such as distrubuted_train.sh.

# standalone training on single NPU device
python train.py --config configs/densenet/densenet_121_gpu.yaml --data_dir /path/to/dataset --distribute False

# distributed training on NPU divices
msrun --bind_core=True --worker_num 8 python train.py --config configs/densenet/densenet_121_ascend.yaml --data_dir /path/to/imagenet

URL and Hyperlink Format

Please use absolute path in the hyperlink or url for linking the target resource in the readme file and table.