Skip to content

Latest commit

 

History

History
90 lines (64 loc) · 4.22 KB

convert.md

File metadata and controls

90 lines (64 loc) · 4.22 KB

Convert pre-trained model from .weights to .npy

  • This page explains how to load pre-trained model (DarkNet-53 and YOLOv3) from .weights file and store the model in a .npy file.
  • Since both DarkNet-53 and YOLOv3 are full convolutional networks, so only the learned parameters of convolutional layers (with learned batch normalization parameters if applicable) are stored in the pre-trained model.
  • The converted pre-trained model is stored as dictionary in .npy file.

Format of .weight

  • The first four 32-bit integers are: major version, minor version, revision, images seen.

  • After the first four integers, learned parameters (weights, bias and batch normalization parameters) for each convolutional layers are stored as float32 following the network layout (parsed from .cfg file).

  • For each convolutional layer with/without batch normalization(BN), the parameters are saved with their corresponding length in the order of (check here for reference):

    No BN With BN
    Conv Bias (out_dim) BN beta (out_dim)
    Conv Weights (out_dim * in_dim * filter_height * filter_width) BN gamma (out_dim)
    - BN Moving Mean (out_dim)
    - BN Moving Variance (out_dim)
    - Conv Weights (out_dim * in_dim * filter_height * filter_width)

Format of .npy

  • The converted model is stored as dictionary in .npy file.

  • First load the file as weight_dict = np.load(NPY_PATH, encoding='latin1').item()

  • Then for each convolutional layer with name 'Layer_Name' (naming rules explained [here]), the parameters can be access as (if applicable):

    Parameters Access
    Conv Weights weight_dict[Layer_Name][weights]
    Conv Bias weight_dict[Layer_Name][weights]
    BN gamma weight_dict[Layer_Name][0]
    BN beta weight_dict[Layer_Name][1]
    BN Moving Mean weight_dict[Layer_Name][2]
    BN Moving Variance weight_dict[Layer_Name][3]

Naming rules for convolutional layers in .npy file

Layer names are different from the names in .cfg files.

DarkNet-53

DarkNet-53 contains 52 convolutional layers for feature extractions named as conv_1...conv_52, and one output layer for classification named as conv_fc_1.

YOLOv3

The first 52 layers of DarkNet-53 are used as feature extractor, so the layer names are the same as DarkNet-53. YOLOv3 predicts bounding boxes at 3 different scales. There are 7 convolutional layers at each scales. Also at the second and third scale, an additional convolution layers is used for reduce the dimensionality of the input feature map. Thus, for the first scale, layers are named as conv_1_1...conv_1_7, and For the second and third scale, layers are named as conv_2_0...conv_2_7 and conv_3_0...conv_3_7, where conv_2_0 and conv_3_0 are used for dimensionality reduction.

Convert models

  • First download the pre-trained .weights file of Darknet-53 from here (Section 'Pre-Trained Models', Darknet53 448x448 link) and YOLOv3 from here (Section 'Performance on the COCO Dataset', YOLOv3-416 link).

  • Go to experiment/convert/

    DarkNet-53 for image classification

    python convert_model.py \
	--model darknet \
	--weights_dir WEIGHTS_DIR \
	--save_dir SAVE_DIR

    YOLOv3 pre-trained

    python convert_model.py \
	--model yolo \
	--weights_dir WEIGHTS_DIR \
	--save_dir SAVE_DIR

    DarkNet-53 feature extractor for YOLOv3

    python convert_model.py \
	--model yolov3_feat \
	--weights_dir WEIGHTS_DIR \
	--save_dir SAVE_DIR

  • --weights_dir is the directory to put the pre-trained .weights file and --save_dir is the directory to put the converted .npy file.

  • Converted parameters are saved as darknet53_448.npy, yolov3.npy and yolov3_feat.npy for pre-trained DarkNet-53 classification, YOLOv3 and DarkNet-53 feature extractor.

Reference code

Author

Qian Ge