Creating the net without yolov3.weights makes it untrainable

[altvali]

Hello,

I'm trying to train a network for different tasks, based on your project. The main issue I have is that if I create the same net with random weights instead of those from yolov3.weights (using this script: https://pastebin.com/RECciwrs) I get a net with the same number of params, the same layers, but on inference it outputs a matrix of NaNs. Since it outputs NaNs, I cannot train it, it will not converge to the desired weights.

How to repro:

1. Save the script from https://pastebin.com/RECciwrs as create_yolo_from_scratch.py
2. python yad2k.py cfg\yolo.cfg yolov3.weights data\yolo.h5
3. python create_yolo_from_scratch.py cfg/yolo.cfg yolo_random_weights.h5
4. For convenience, resize giraffe.jpg to 416x416 and rename as giraffe416.jpg
5. Run in terminal: python

   from keras.models import load_model
   import numpy as np
   import cv2
   yolo = load_model('yolo.h5')
   yolo_rw = load_model('yolo_random_weights.h5')
   path = 'images/test/giraffe416.jpg'
   image = cv2.imread(path)
   image = np.array(image, dtype='float32')
   image /= 255.
   image = np.expand_dims(image, axis=0)
   yolo.predict(image)
   # will output a matrix of numbers
   yolo_rw.predict(image)
   # will output a matrix of NaNs
   yolo_rw.set_weights(yolo.get_weights())
   yolo_rw.predict(image)
   # will output a matrix of numbers

Without using yolov3.weights, do you have any advice on how to seed the network so that it's trainable? I want to then use a different config file for a different number of classes in order to train. Thank you.

[altvali]

Nevermind, it was an issue with my Tensorflow installation.