zi-angzhang / blah-blah Goto Github PK

Image Restoration includes image denoising, super resolution, inpainting, and so on.

Denoise networks looks similar to super resolution networks, Where is the node that I can bond them together?

My computer failed in booting and failed me.

What's momentum in optimizers? Where is the difference between optimizers?

Huawei launched its Ascend 910, which claims to outperform by twice V100, Here I investigate the GPU and TPU again, understood how memory is allocated by the memory. Wow, padding may lead to a waste of memory.

Chen Li's was beautiful!

I think CNN is nothing but pattern matching.

I brought a Mi 4 smart band, confused about its "card simulation" function. Till now, the mechanism of RTID card is understood while getting my card to simulate my pass card is waiting for my spare time.

Training with x2 batchSize lead to fastering training, that observation never goes beyond expection because one-shot loading is faster twice smaller loading. However, when training with batchSize = 64 takes 3 minutes to finish an epoch while batchSize = 8 acquires 16 minutes, I feel shocked and decide to investigate. Till Now, I found lmdb is slightly faster when loading and is safe with multiprocess.