ResNet-50 lr_policy: "step" doesn't work, always divide by 10

Question

I'm using ResNet-50 of Caffe in order to classify images. I can finetune the it using ResNet-50_model.caffemodel, and the inital learning rate I set is 0.001, but the learning rate after each 500 iter is divide by 10 instead of reduce by 0.00005.

---

Here is my resnet_solver.prototxt

net: "data/resnet_data/resnet_50.prototxt"

test_iter: 320
test_interval: 100

base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 500
display: 100
max_iter: 2000
momentum: 0.9
weight_decay: 0.0005
snapshot: 500

snapshot_prefix: "data/resnet_data/model/resnet_grid"
solver_mode: GPU

And here is training screenshot: (iter 100 lr = 0.001, iter 500 lr = 0.0001, iter 1000 lr = 1e-05, and so on)

This is my first time to train ResNet. If you have any suggestion, please let me know. Many thanks!