Custom model w/ AlexeyAB darknet : export fine, load_weights issue

[alexis-gruet-deel]

Hello,
I trained a custom model with the c/cuda repo. model is (320,320,3) - 3 classes - custom anchors. I'm able to export the model using the convert_darknet_weights.py ( after modified the INPUT_SHAPE and NUM_CLASSES). However when doing :

import tensorflow as tf

from tf2_yolov4.anchors import YOLOV4_ANCHORS
from tf2_yolov4.model import YOLOv4

model = YOLOv4(
    input_shape=(320, 320, 3),
    anchors=YOLOV4_ANCHORS,
    num_classes=3,
    training=False,
    yolo_max_boxes=100,
    yolo_iou_threshold=0.5,
    yolo_score_threshold=0.5,
)

model.load_weights("....path-to-converted-weights.h5...") 

This is what I get :

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-6-f6463dea2434> in <module>
----> 1 model.load_weights("./yolov4.h5")

5 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/tensor_shape.py in assert_is_compatible_with(self, other)
   1115     """
   1116     if not self.is_compatible_with(other):
-> 1117       raise ValueError("Shapes %s and %s are incompatible" % (self, other))
   1118 
   1119   def most_specific_compatible_shape(self, other):

ValueError: Shapes (3, 3, 128, 256) and (1024, 512, 3, 3) are incompatible

[RaphaelMeudec]

Hi @MI-LA01, could you point to the .weights file you obtained with the darknet training?

[alexis-gruet-deel]

Hello @RaphaelMeudec and thanks for your kind help. Pls find the google link

[RaphaelMeudec]

@MI-LA01 I just looked into it:

• output number of classes is hardcoded inside convert_darknet_weights (to 80), it makes the conversion fail on my computer
• after putting num_classes=3 in the conversion script, the conversion is running fine
• I have been able to load the .h5 file properly

On my side, I'll open a PR to make the number of classes configurable in the conversion script.
On your side, can you check that you have been converting the right .weights file? I don't understand why it didn't fail on your side.

[alexis-gruet-deel]

Hi @RaphaelMeudec
Some pointers :

• I confirm this is the good weights file
• As mentioned in the issue, I already changed the NUM_CLASSES in the conversion script to reflect my number of class. Thus NUM_CLASSES=3 and later in the script :

model = YOLOv4(
        input_shape=INPUT_SHAPE, num_classes=NUM_CLASSES, anchors=YOLOV4_ANCHORS

would be ok, reason why the conversion script works.

• I posted the outcome from the conversion script in the google drive, if you can look at it HERE?
• The screenshot below gives the output of the conversion script :

- The same colab used to perform the load :

[alexis-gruet-deel]

Found ; By installing the tf addons and specifically the mish() activation layer I was able to solve this issue. I was using a different implementation than the one supplied by tfa due to compat issues with the jetson/tx* environment.