zejiangh / filter-gap Goto Github PK

Hi! Thanks for the great work. I am currently trying to reproduce the ImageNet results but encounter some troubles. Is it possible to provide training(tensorflow for eg.) logs for the 77.4 results for debugging purpose? Thanks a lot!

Did you ever come across the situation in which the layer ratios for a certain layer reaches 0 (or 1 - not sure which way around it is), so that the mask sets all channels to 0, essentially rendering the layer useless? How do you mitigate this?

Note, I reimplemented the paper in tensorflow, so I have slightly different code.

你好，非常感谢你的开源代码！另外，请问如果我打算训练cifar10/100, 除了需要改数据集的路径，还需要改其他的训练参数吗？

因为受我的设备限制无法支持nvidia-dali，我在使用dataloader替换dali时遇到了困难，我无法得知您的代码中的cocoIteration返回的数据的shape，还请您告知

../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [0,0,0] Assertion idx_dim >= 0 && idx_dim < index_size && "index out of bounds" failed.
../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [1,0,0] Assertion idx_dim >= 0 && idx_dim < index_size && "index out of bounds" failed.
../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [2,0,0] Assertion idx_dim >= 0 && idx_dim < index_size && "index out of bounds" failed.
idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [122,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [123,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [124,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [125,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [126,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"failed. ../aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [0,0,0], thread: [127,0,0] Assertionidx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.
terminate called after throwing an instance of 'c10::CUDAError'
what(): CUDA error: device-side assert triggered
Exception raised from record at ../aten/src/ATen/cuda/CUDAEvent.h:119 (most recent call first):

Is there any suggestion?

作者你好，从论文介绍看，CHEX方法应该是结构化剪枝，但不同压缩率下提供的Checkpoints模型权重所占存储体积为什么都是一样的呢？例如：resnet50_1g/2g/3g都是195MB。

此外，readme中resnet50_2g的模型权重给成resnet50_3g了，麻烦更新下哈，谢谢！

prune_utils.py 中IS_update_channel_mask函数最后给当前模型BN层参数全部赋值为上一次剪枝的参数，然后prev_model为赋值后的当前模型深拷贝，这样每次更新的prev_model一直保存着最初的bn层权重

是否应该对MRU再生的通道对应的BN层参数赋值

是否在IS_update_channel_mask中更新bn层参数的代码逻辑有问题

请给出解释

十分感谢你的出色工作并开源！
我有一点困惑，在README中提到ssd detection会训练650epoch，这与一般的配置（120epoch）有些不同。

同时在补充材料中提到训练240k~129epoch（240k/(118287/64) ~129)。

哪种配置才是实际使用的呢？

Hi, I'm trying to reproduce the paper but am stuck at the SSD model's evaluation step.
I'm not claiming anything but I feel like the instruction for SSD in this repo is partially wrong:
First, an SSD300 is initialized as a full model:

Hi!

Thanks for your work. I am wondering how to prune maskrcnn as it uses frozen batchnorm in maskrcnn-benchmark. As the subnet exploration stage uses gamma as the criterion.