W tensorflow/core/grappler/utils/graph_view.cc:830] No registered 'swish_f32' OpKernel for GPU devices compatible with node {{node model_1/swish_38/swish_f32}}

[bitwangdan]

This error occurred when I was training on the GPU，Has anyone else experienced this problem?

[KaqiKaqi]

same warning.

[bitwangdan]

@KaqiKaqi have you solved it?

[mingxingtan]

Hmm, I am surprised that swish doesn't work on GPU. If so, could you replace:

automl/efficientdet/utils.py

Line 30 in e90a471

relu_fn = tf.nn.swish

with "relu_fn = x * tf.sigmoid(x)"? (they are equivalent but tf.nn.swish is supposed to be more memory efficient)

[mingxingtan]

As shown in this tutorial, it seems to run well with GPU. closing this issue for now.

[Cocoalate]

HI, i'm facing the same problem and replace tf.nn.swish with tf.sigmoid works for me. I suppose it's happenning because I'm using tf.compat.v1 to train main.py under tf2.0 environment