search

hujie-frank / SENet

Squeeze-and-Excitation Networks
Apache License 2.0
3.38k stars 837 forks source link

top-1 accuracy is very low #45

Closed 408550969 closed 6 years ago

408550969 commented 6 years ago

Hi I test the SE-BN-Inception and use the pre-trained caffemodel, but the accuracy is 0.00018 on Imagenet2012. May I ask why this could happen, is it a problem with my dataset? Thanks.

hujie-frank commented 6 years ago

Please refer to #5.

408550969 commented 6 years ago

Thanks!

408550969 commented 6 years ago

Hi,I am training the SE-BN-Inception, but I find the acuuracy is only 68% now.(accurate learning rate is 0.0001) My learning rate is set wrong,I set it to 0.1 from the beginning. And I only use mirror true.Others are same. I am set the use_global_status to false in train.prototxt,and set it to true in test.prototxt. Do you know why the accuracy is so low?How to implement the augmentation list in the form?Thank you!

hujie-frank commented 6 years ago

@408550969 The data augmentation don't bring significant improvement. It must be something wrong in your experiments. Additionally, you can refer here to find the implementation of the augmentation in list.

408550969 commented 6 years ago

This is the top and bottom of the train.prototxt.Can you help me see what's wrong?Thanks! layer { name: "data" type: "Data" top: "data" top: "label" include { phase: TRAIN } transform_param { mirror: true crop_size: 224 mean_value: 104.0 mean_value: 117.0 mean_value: 123.0 } data_param { source: "/media/cll/Seagate/ilsvrc12_train_lmdb" batch_size: 32 backend: LMDB } }

layer { name: "classifier" type: "InnerProduct" bottom: "pool5/7x7_s1" top: "classifier" inner_product_param { num_output: 1000 } } layer { name: "loss" type: "SoftmaxWithLoss" bottom: "classifier" bottom: "label" top: "loss" }