guigzzz / keras-yolo-v2 Goto Github PK

I am attempting to train a custom model with TinyYOLOv2, however when I replace the batch size and label count on the trainnet = TinyYOLOv2(13 * 32, batch_size, num_classes, is_learning_phase=True) call I get the following error when the model is being created:

  File "/test/yolo2/tiny_yolo_v2.py", line 24, in __init__
    self.m = self.buildModel()
  File "/test/yolo2/tiny_yolo_v2.py", line 50, in buildModel
    )(model)
  File "/test/venv/lib/python3.7/site-packages/keras/engine/base_layer.py", line 474, in __call__
    output_shape = self.compute_output_shape(input_shape)
  File "/test/venv/lib/python3.7/site-packages/keras/layers/core.py", line 398, in compute_output_shape
    input_shape[1:], self.target_shape)
  File "/test/venv/lib/python3.7/site-packages/keras/layers/core.py", line 386, in _fix_unknown_dimension
    raise ValueError(msg)
ValueError: total size of new array must be unchanged

My guess is i would need to change how the Reshape layer is added, what could I change to make it work?

Hello, i was wondering why the input shape to loss function is (13, 13, 5, 25), and you opt to leave out batch dimension, or am i looking at it wrong? I see ProcessGroundTruth returns y_true of shape (13, 13, 5, 25) which is input to loss function. In general, y_true and y_pred in loss function are going to have batch_size as their first dimension, or in this case number_of_samples/steps_per_epoch.

Hi,
I am trying to train the TinyYOLOv2 model with VOC07 dataset. I've defined the function to fetch bounding boxes and labels for all the training images. But after fitting the model the output returns empty arrays when I make tiny.m.forward(image)[0] for a test image.

The training set is (318, 416, 416, 3) and the y_true is (318, 13, 13, 5, 25).

Am I doing something wrong? During fitting the loss information per epoch was always NaN.

I have seen other loss functions implemented, yours is more similar to that of Yolo V1, my question is:

1. Why not use categorical crosss entrophy in class predictions ?
2. In the IoU filter, maked in the computation of noobjs_loss you put that
   higthst_iou = tf.math.reduce_max(iou, axis=-1)
mask = (tf.cast(higthst_iou < 0.6, dtype=tf.float32)[..., None] * (1 - true_object)) # noobj mask
   but, if we have to penalize higest IoU, why put < 0.6 and not > 0.6

thanks a lot! :D

I started the training however after certain epochs i am getting loss NaN . just before getting nan the loss was converging .