search

asanakoy / deeppose_tf

DeepPose implementation on TensorFlow. Original Paper http://arxiv.org/abs/1312.4659
Other
144 stars 61 forks source link

260k interations, mPCP on LSP is close to 0.42 with Imagenet initialization #7

Closed jjzhang10 closed 7 years ago

jjzhang10 commented 7 years ago

I train the LSP dataset with imagenet initialization. It has been 260k iterations, the train/pose_error is around[0.02 0.03]. The mPCP score (0.42) is still far from the result listed in the table. It does increase, but pretty slowly.

Does it mean it cannot converge for this time? I need restart the training? Thanks. wechat image_20170707094329

asanakoy commented 7 years ago

Looks like you initialized it randomly. On my experience Radom initialization achieves PCP score ~0.42

jjzhang10 commented 7 years ago

@asanakoy I forget to add the parameter "fix_conv_lr 10000". Training the conv and fc weights together at the very beginning may hurt the performance for the fine-tuning on imagenet. Now here is the result: qq 20170708182817

The train/pose_loss is already lower to 0.03 (you mentioned in the previous post) after 140k iterations, but the test/mPCP is around 0.47. Will wait and see whether it will achieve 0.55 By the way, will the program converge every time?

asanakoy commented 7 years ago

You need to train longer. Should converge