How do you get 'num_examples_per_epoch'=5.32e4?

bilylee / SiamFC-TensorFlow

A TensorFlow implementation of the SiamFC tracker

MIT License

358 stars 112 forks source link

How do you get 'num_examples_per_epoch'=5.32e4? #29

Closed yiminglin-ai closed 6 years ago

yiminglin-ai commented 6 years ago

Hi Bily, can I ask how did you obtain the constant 'num_examples_per_epoch' in your config file? I looked into the training dataset by the following code snippet,

dataset = VID('data/train_imdb.pickle', 100)
print(dataset.epoch_size) # output is 8418

The epoch size is 8418 so the number of examples per epoch (num_examples_per_epoch) should be 8418 * 8 = 67,344?

Can you tell me if I mistook anything? Best regards, Yiming

bilylee commented 6 years ago

Hi,

It is directly taken from https://github.com/bertinetto/siamese-fc/blob/f586d4e32091184240ab7b496ea4df9cb81dd001/training/experiment.m#L24

By using 5.32e4, it means taking 5.32e4 / 8418 = 6 pairs per video for training. This is a hyper-parameter and doesn't necessarily have a meaningful value.