any implementation for tiny yolov4 version?

[farhantandia]

Fail to convert tiny version, any help for it? thanks

[freitaucher]

well, instead of modifying convert.py I tried to work out demo.py. There one should import:

from yolov4.tf import YOLOv4 # (this can be installed with pip)

and read in the weight of yolov4-tiny:

yolo = YOLOv4(tiny=True) yolo.classes = "model_data/coco_classes.txt" yolo.make_model() yolo.load_weights("model_data/yolov4-tiny.weights", weights_type="yolo")

So far so good, until I start with encoding:

from tools import generate_detections as gdet ... encoder = gdet.create_box_encoder("model_data/mars-small128.pb", batch_size=1) ... features = encoder(frame, boxes) ...

if one looks in the appropriate source: tools/generate_detections .py, there one finds

import tensorflow as tf import tensorflow.compat.v1 as tf tf.disable_v2_behavior()

and exactly this spoils in the main code everything which has to deal with tensorflow 2. In fact, it is needed only inside class ImageEncoder(object)

Would someone hint how to work out this encoder in tf2 (best), or how to limit the use of tf1 to the encoder module only, or at least explain - what is the purpose of encoder?

[freitaucher]

Ok, the purpose of encoder is now clear. It simplifies the input info for tracker. Anyway, the technical issue with tf2 remains..