Enable mixed precision

[DannyGJdeJong]

Is it possible to use TensorFlow's mixed precision for a performance boost on newer nvidia graphics cards?

from tensorflow.keras.mixed_precision import experimental as mixed_precision

policy = mixed_precision.Policy('mixed_float16')
mixed_precision.set_policy(policy)

yolo.make_model()

Right now it throws a TypeError when trying to do this because the outputs and inputs will not match in the existing model.

TypeError: Input 'y' of 'AddV2' Op has type float32 that does not match type float16 of argument 'x'.

Am I doing something wrong in enabling it this way or is it just not possible with the current implementation?

[hhk7734]

tensorflow-yolov4/py_src/yolov4/model/yolov4.py

Line 98 in be7b930

x = self.yolov3_head(x)

tensorflow-yolov4/py_src/yolov4/model/yolov4.py

Line 128 in be7b930

x = self.yolov3_head_tiny(x)

remove the code above.
I plan to modify the head part. (#47 ) But, I'm busy these days...
So, if it is difficult to wait for the update, you will have to write your own head.

[hhk7734]

On yolov4 v3.0.0, you can enable mixed precision.

yolo.predict or yolo.inference use yolo_diou_nms.

tensorflow-yolov4/py_src/yolov4/tf/__init__.py

Lines 134 to 136 in 211afdd

	pred_bboxes = self.yolo_diou_nms(
	candidates=candidates, beta_nms=self.config.yolo_0.beta_nms
	)

But, this allows only np.float32 type array. So if you want to enable mixed precision, modify the code shown below.

    if candidates is not np.float32:
        candidates = candidates.astype(np.float32)
    pred_bboxes = self.yolo_diou_nms( 
        candidates=candidates, beta_nms=self.config.yolo_0.beta_nms 
    )