Query on segmentation modeling (deeplabv3)

1. In general, shouldn’t the input size and target size match? Why I need to downscale the target samples for training? For example, input size 128, target size 32.
2. I got the following warning message. Tested on colab.

WARNING:tensorflow:Gradients do not exist for variables [‘deep_lab_v3/feature_pyramid/output_P3/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P3/bias:0’, ‘deep_lab_v3/feature_pyramid/output_P4/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P4/bias:0’, ‘deep_lab_v3/feature_pyramid/output_P5/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P5/bias:0’] when minimizing the loss. If you’re using model.compile(), did you forget to provide a loss argument?

1. mixed precision makes loss value to nan (probably need to cast the output of segmentation head to tf.float32. You might want to introduce dtype param in here.)

Reproducible Code.

[qlzh727]

Thanks for the information. Note that for the accuracy, it might because the backbone we current use doesn’t have the pretrained weights, whereas the tutorial does use the keras application model with imagenet weights.

[qlzh727]

Thanks for the question.

1. In general, shouldn’t the input size and target size match? Why I need to downscale the target samples for training? For example, input size 128, target size 32.

Sure. Currently the model is just return the raw output from detection head. We could add extra layers to scale up the output to let it match the input shape. (for better comparison).

1. I got the following warning message. Tested on colab.

WARNING:tensorflow:Gradients do not exist for variables [‘deep_lab_v3/feature_pyramid/output_P3/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P3/bias:0’, ‘deep_lab_v3/feature_pyramid/output_P4/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P4/bias:0’, ‘deep_lab_v3/feature_pyramid/output_P5/kernel:0’, ‘deep_lab_v3/feature_pyramid/output_P5/bias:0’] when minimizing the loss. If you’re using model.compile(), did you forget to provide a loss argument?

Let me take a closer look. I guess the FPN weights is not tracked by model correctly. Thanks for the report.

1. mixed precision makes loss value to nan (probably need to cast the output of segmentation head to tf.float32. You might want to introduce dtype param in here.)

Thanks. Will do.

Reproducible Code.