search

tensorflow / models

Models and examples built with TensorFlow
Other
76.95k stars 45.79k forks source link

Unclear about mobilenet v3 training config #9852

Open Kewenjing1020 opened 3 years ago

Kewenjing1020 commented 3 years ago

Hi, I tried to train mobilenet v3 following the configuration in page. However, some of the configurations are not defined in train_image_classifier.py:

init_stddev | 0.008 |  
dropout_keep_prob | 0.8 |  
bn_moving_average_decay | 0.997 |  
bn_epsilon | 0.001 |

Even though I tried to define these hyperparameters, I couldn't find where to define them. For example, I can't find where to define dropout_keep_prob, as the mobilenetv3 network doesn't include a dropout layer and the slim.conv2d layer has not a 'dropout' argument. Could you please further explain how to do these configurations? Or could you kindly provide the training scripts?

In addition, how do you train mobilenetv3 on multiple GPUs? I think train_image_classifier.py only supports single GPU training.

Kewenjing1020 commented 3 years ago

Hi, I found that 

init_stddev | 0.008 |  
dropout_keep_prob | 0.8 |  
bn_moving_average_decay | 0.997 |  
bn_epsilon | 0.001 |

are defined in nets/mobilenet/mobilenet.py.

Could you help me with multiple-GPU training? And, how many epochs have trained for each model(small/large/minimal)?