search

Bisonai / mobilenetv3-tensorflow

Unofficial implementation of MobileNetV3 architecture described in paper Searching for MobileNetV3.
Apache License 2.0
225 stars 81 forks source link

Training on simple classification tasks got nan result #1

Closed lucasjinreal closed 5 years ago

lucasjinreal commented 5 years ago 
INFO 05-13 18:44:56 classify_flowers_mbv3.py:94 - Epoch: 0, iter: 0, loss: 1.6094379425048828, train_acc: 0.0
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)
pred:  tf.Tensor([[[[nan nan nan nan nan]]]], shape=(1, 1, 1, 5), dtype=float32)

have you trained your implementation?

martinkersner commented 5 years ago

Hi @jinfagang!

We trained with our implementation. You can check our training script at https://github.com/Bisonai/mobilenetv3-tensorflow/blob/master/train.py. Currently, there are two examples for CIFAR10 and MNIST datasets.

Cheers, Martin