在线时间:8:00-16:00
迪恩网络APP
随时随地掌握行业动态
扫描二维码
关注迪恩网络微信公众号
开源软件名称(OpenSource Name):smichalowski/google_inception_v3_for_caffe开源软件地址(OpenSource Url):https://github.com/smichalowski/google_inception_v3_for_caffe开源编程语言(OpenSource Language):开源软件介绍(OpenSource Introduction):Google Inception V3 for Cafferevision 2IntroductionThis model is a replication of the model described in the Rethinking the Inception Architecture for Computer Vision If you wish to train this model on ILSVRC2012 dataset remember to prepare LMDB with 300px images instead of 256px. Hardware and TrainingOriginal implementation from paper uses 32 batch size for 100 epochs using RMSProp with learning rate of 0.045. You need some Titan X or K40 with more than 10GB of RAM. Provided train_val.prototxt uses batch_size = 22 and fits into Titan X. Please use NVIDIA/caffe for training. I was UNABLE to achieve good results using regular Caffe (probably because of different BatchNorm implementation). Training on TINY SETI have trained this model on ImageNet subset of 18 categories for 11 epochs using NVIDIA/caffe branch 0.15.5:
If you want to try it yourself you can find it here. Remember that this link provides model trained using ONLY 18 categories. DIGITSIf you want to train this network using NVIDIA/DIGITS compatible train_val.prototxt is provided in digits folder for your pleasure. Please be advised that currently DIGITS web interface doesnt allow you to set following parameters for solver that allowed me to achieve such good results on tiny set:
You can force DIGITS to use these parameters hardcoding these values into train_caffe.py Just paste this code:
Also if you will use DIGITS to create "New Image Classification Dataset" be sure to set Image Encoding to None. Training on full ImageNetRight now Im training it on full ImageNet set using provided solver.txt. I will publish it when it`s done. LicenseThis model is released for unrestricted use. |
2023-10-27
2022-08-15
2022-08-17
2022-09-23
2022-08-13
请发表评论