search

keras-team / keras-applications

Reference implementations of popular deep learning models.
Other
2k stars 910 forks source link

Self-train These Models on ImageNet #103

Closed fzyzcjy closed 5 years ago

fzyzcjy commented 5 years ago

Hi! Thank you very much for these wonderful libraries! I want to have a mobilenet with alpha=0.125 (or possibly smaller) to make it even more light-weight. Since there is no pretrained models (only with alpha=0.25 and so on), I want to first pretrain it on mobilenet, and then use it for fine-tuning. Therefore, I would appreciate it if I could be given the code you use to pretrain it on mobilenet. Thank you very much!

taehoonlee commented 5 years ago

@fzyzcjy,

The keras-applications is designed for inference only, so don't provide training details such as data augmentation (e.g., rotating, shifting), optimization hyperparameters (e.g., lr, decay), and a release number of ImageNet used for training.

For such information, you can check the original repositories shown in the table in README.

fzyzcjy commented 5 years ago

Thank you very much!

fzyzcjy commented 5 years ago

@taehoonlee Sorry for bothering again. So I just ask the training details on corresponding TensorFlow GitHub. However, could you please give the conversion from TF to Keras? Thank you very much for your reply!

taehoonlee commented 5 years ago

@fzyzcjy, The conversions have been done by contributors. They must have the conversion script.

fzyzcjy commented 5 years ago

@taehoonlee Thanks very much! I wonder how can I contact them?

taehoonlee commented 5 years ago

@fzyzcjy, You can find individual authors on the keras commit logs.

fzyzcjy commented 5 years ago

@taehoonlee Thanks! (sorry i forgot to reply 2 months ago when seeing your message)