./finetune.sh 0 with no output and finished within one second

[cicido]

What hardware and operating system/distribution are you running?

Operating system: CentOS7
CUDA version: 8.0
CUDNN version: .6.0
openCV version: 3.0
BLAS: open
Python version: 3.6.3
boost 1.65.1
boost-python: 1.65.1 with python3 support

after run ./combine.sh | xargs ./calflop.sh. I get the temp/cb_3c_3C4x_mem_bn_vgg.prototxt:
name: "VGG_ILSVRC_16_layers"
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
transform_param {
mirror: false
crop_size: 224
mean_value: 104.0
mean_value: 117.0
mean_value: 123.0
}
data_param {
source: "/home/duanxiping/channel-pruning/caffe/examples/imagenet/ilsvrc12_val_lmdb"
batch_size: 32
backend: LMDB
}
image_data_param {
source: "/data/mydata/sourcetrain.txt"
batch_size: 32
shuffle: true
new_dim: 256
bicubic: true
}
}
layer {
name: "conv1_1"
type: "Convolution"
bottom: "data"
top: "conv1_1"
convolution_param {
num_output: 64
pad: 1
kernel_size: 3
}
}
layer {
name: "relu1_1"
type: "ReLU"
bottom: "conv1_1"
top: "conv1_1"
}
layer {
name: "conv1_2_V"
type: "Convolution"
bottom: "conv1_1"
top: "conv1_2_V"
convolution_param {
num_output: 22
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv1_2_H"
type: "Convolution"
bottom: "conv1_2_V"
top: "conv1_2_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 22
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv1_2_P"
type: "Convolution"
bottom: "conv1_2_H"
top: "conv1_2_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 57
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu1_2"
type: "ReLU"
bottom: "conv1_2_P"
top: "conv1_2_P"
}
layer {
name: "pool1"
type: "Pooling"
bottom: "conv1_2_P"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv2_1_V"
type: "Convolution"
bottom: "pool1"
top: "conv2_1_V"
convolution_param {
num_output: 49
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv2_1_H"
type: "Convolution"
bottom: "conv2_1_V"
top: "conv2_1_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 49
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv2_1_P"
type: "Convolution"
bottom: "conv2_1_H"
top: "conv2_1_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 117
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu2_1"
type: "ReLU"
bottom: "conv2_1_P"
top: "conv2_1_P"
}
layer {
name: "conv2_2_V"
type: "Convolution"
bottom: "conv2_1_P"
top: "conv2_2_V"
convolution_param {
num_output: 62
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv2_2_H"
type: "Convolution"
bottom: "conv2_2_V"
top: "conv2_2_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 62
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv2_2_P"
type: "Convolution"
bottom: "conv2_2_H"
top: "conv2_2_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 120
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu2_2"
type: "ReLU"
bottom: "conv2_2_P"
top: "conv2_2_P"
}
layer {
name: "pool2"
type: "Pooling"
bottom: "conv2_2_P"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv3_1_V"
type: "Convolution"
bottom: "pool2"
top: "conv3_1_V"
convolution_param {
num_output: 110
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv3_1_H"
type: "Convolution"
bottom: "conv3_1_V"
top: "conv3_1_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 110
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv3_1_P"
type: "Convolution"
bottom: "conv3_1_H"
top: "conv3_1_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 228
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu3_1"
type: "ReLU"
bottom: "conv3_1_P"
top: "conv3_1_P"
}
layer {
name: "conv3_2_V"
type: "Convolution"
bottom: "conv3_1_P"
top: "conv3_2_V"
convolution_param {
num_output: 118
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv3_2_H"
type: "Convolution"
bottom: "conv3_2_V"
top: "conv3_2_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 118
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv3_2_P"
type: "Convolution"
bottom: "conv3_2_H"
top: "conv3_2_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 224
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu3_2"
type: "ReLU"
bottom: "conv3_2_P"
top: "conv3_2_P"
}
layer {
name: "conv3_3_V"
type: "Convolution"
bottom: "conv3_2_P"
top: "conv3_3_V"
convolution_param {
num_output: 141
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv3_3_H"
type: "Convolution"
bottom: "conv3_3_V"
top: "conv3_3_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 141
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv3_3_P"
type: "Convolution"
bottom: "conv3_3_H"
top: "conv3_3_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 256
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu3_3"
type: "ReLU"
bottom: "conv3_3_P"
top: "conv3_3_P"
}
layer {
name: "pool3"
type: "Pooling"
bottom: "conv3_3_P"
top: "pool3"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv4_1_V"
type: "Convolution"
bottom: "pool3"
top: "conv4_1_V"
convolution_param {
num_output: 233
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv4_1_H"
type: "Convolution"
bottom: "conv4_1_V"
top: "conv4_1_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 233
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv4_1_P"
type: "Convolution"
bottom: "conv4_1_H"
top: "conv4_1_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 468
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu4_1"
type: "ReLU"
bottom: "conv4_1_P"
top: "conv4_1_P"
}
layer {
name: "conv4_2_V"
type: "Convolution"
bottom: "conv4_1_P"
top: "conv4_2_V"
convolution_param {
num_output: 256
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv4_2_H"
type: "Convolution"
bottom: "conv4_2_V"
top: "conv4_2_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 256
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv4_2_P"
type: "Convolution"
bottom: "conv4_2_H"
top: "conv4_2_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 484
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu4_2"
type: "ReLU"
bottom: "conv4_2_P"
top: "conv4_2_P"
}
layer {
name: "conv4_3_V"
type: "Convolution"
bottom: "conv4_2_P"
top: "conv4_3_V"
convolution_param {
num_output: 302
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv4_3_H"
type: "Convolution"
bottom: "conv4_3_V"
top: "conv4_3_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 302
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "conv4_3_P"
type: "Convolution"
bottom: "conv4_3_H"
top: "conv4_3_P"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 1.0
decay_mult: 2.0
}
convolution_param {
num_output: 512
pad: 0
kernel_size: 1
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "relu4_3"
type: "ReLU"
bottom: "conv4_3_P"
top: "conv4_3_P"
}
layer {
name: "pool4"
type: "Pooling"
bottom: "conv4_3_P"
top: "pool4"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv5_1_V"
type: "Convolution"
bottom: "pool4"
top: "conv5_1_V"
convolution_param {
num_output: 398
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv5_1_H"
type: "Convolution"
bottom: "conv5_1_V"
top: "conv5_1_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 512
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "relu5_1"
type: "ReLU"
bottom: "conv5_1_H"
top: "conv5_1_H"
}
layer {
name: "conv5_2_V"
type: "Convolution"
bottom: "conv5_1_H"
top: "conv5_2_V"
convolution_param {
num_output: 390
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv5_2_H"
type: "Convolution"
bottom: "conv5_2_V"
top: "conv5_2_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 512
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "relu5_2"
type: "ReLU"
bottom: "conv5_2_H"
top: "conv5_2_H"
}
layer {
name: "conv5_3_V"
type: "Convolution"
bottom: "conv5_2_H"
top: "conv5_3_V"
convolution_param {
num_output: 379
pad: 1
pad: 0
kernel_size: 3
kernel_size: 1
stride: 1
}
}
layer {
name: "conv5_3_H"
type: "Convolution"
bottom: "conv5_3_V"
top: "conv5_3_H"
param {
lr_mult: 1.0
decay_mult: 1.0
}
convolution_param {
num_output: 512
bias_term: true
pad: 0
pad: 1
kernel_size: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
}
}
layer {
name: "relu5_3"
type: "ReLU"
bottom: "conv5_3_H"
top: "conv5_3_H"
}
layer {
name: "pool5"
type: "Pooling"
bottom: "conv5_3_H"
top: "pool5"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "fc6"
type: "InnerProduct"
bottom: "pool5"
top: "fc6"
inner_product_param {
num_output: 4096
}
}
layer {
name: "relu6"
type: "ReLU"
bottom: "fc6"
top: "fc6"
}
layer {
name: "drop6"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7"
type: "InnerProduct"
bottom: "fc6"
top: "fc7"
inner_product_param {
num_output: 4096
}
}
layer {
name: "relu7"
type: "ReLU"
bottom: "fc7"
top: "fc7"
}
layer {
name: "drop7"
type: "Dropout"
bottom: "fc7"
top: "fc7"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc8"
type: "InnerProduct"
bottom: "fc7"
top: "fc8"
inner_product_param {
num_output: 1000
}
}
layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "fc8"
bottom: "label"
top: "loss"
}
layer {
name: "accuracy/top1"
type: "Accuracy"
bottom: "fc8"
bottom: "label"
top: "accuracy@1"
accuracy_param {
top_k: 1
}
}
layer {
name: "accuracy/top5"
type: "Accuracy"
bottom: "fc8"
bottom: "label"
top: "accuracy@5"
accuracy_param {
top_k: 5
}
}

the ilsvrc12_val_lmdb file exists, and temp/model has no file. I don't know what' s wrong

[cicido]

I am not sure this problem may be with the bash_completion.

[ethanhe42]

use ./caffe/build/tools/caffe train -solver temp/solver.prototxt -weights temp/3c_vgg.caffemodel -gpu $1 instead of finetune.sh. Does it work?

[cicido]

it works. But I get a error as following:
F1206 09:59:24.308379 29489 net.cpp:757] Cannot copy param 0 weights from layer 'conv5_1_H'; shape mismatch. Source param shape is 398 398 1 3 (475212); target param shape is 512 398 1 3 (611328). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer.
*** Check failure stack trace: ***
@ 0x7f723767ae6d (unknown)
@ 0x7f723767cced (unknown)
@ 0x7f723767aa5c (unknown)
@ 0x7f723767d63e (unknown)
@ 0x7f723e7ce739 caffe::Net<>::CopyTrainedLayersFrom()
@ 0x7f723e7d50f2 caffe::Net<>::CopyTrainedLayersFromBinaryProto()
@ 0x7f723e7d5188 caffe::Net<>::CopyTrainedLayersFrom()
@ 0x40c405 CopyLayers()
@ 0x40ce6d train()
@ 0x409e0c main
@ 0x7f7225bb7b15 __libc_start_main
@ 0x40a7b1 (unknown)
Aborted (core dumped)

[cicido]

I note that code I have clone may be too old when I open a new issue able rankdic.
the code in my machine is :
for i in rankdic:
if 'conv5' in i:
break
rankdic[i] = int(rankdic[i] * 4. / speed_ratio)
I not the new code is :
the code in my machine is :
for i in rankdic:
if 'conv5' in i:
continue
rankdic[i] = int(rankdic[i] * 4. / speed_ratio)
since python dict has no order, this my produce unmatched shape error.
I will try later

[fgxfxpfzzfcc]

@cicido Have you solved this problem? I now encountered the same mistakes with you, if you solve,can you tell me how to solver this problem, thanks.

[cicido]

sorry for relay later. I have not solved this problem. but from the code:
if dcfgs.dic.vh and (conv in alldic or conv in pooldic) and (convnext in self.convs)

and the following code:
pooldic = ['conv1_2', 'conv2_2']#, 'conv3_3']

in PDB result , the alldic is:
(Pdb) p alldic
['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv3_2', 'conv4_2']

(Pdb) p list(zip(convs[1:], convs[2:]+['pool5']))
[('conv1_2', 'conv2_1'), ('conv2_1', 'conv2_2'), ('conv2_2', 'conv3_1'), ('conv3_1', 'conv3_2'), ('conv3_2', 'conv3_3'), ('conv3_3', 'conv4_1'), ('conv4_1', 'conv4_2'), ('conv4_2', 'conv4_3'), ('conv4_3', 'conv5_1'), ('conv5_1', 'conv5_2'), ('conv5_2', 'conv5_3'), ('conv5_3', 'pool5')]

so when in ('conv4_3', 'conv5_1') pair, there is no channeling pruning, only do VH_decompose andITQ_decompose.
I wander there may be something wrong with the monment of no chaneling pruning.

[cicido]

(64, 64) conv1_1
(64, [22,22,55]) conv1_2
(128,[49,49,117]) conv2_1
(128,[62,62,119]) conv2_2
(256,[110,110,230]) conv3_1
(256,[118,118,223]) conv3_2
(256,[141,141,256]) conv3_3
(512,[233,233,489]) conv4_1
(512,[256,256,470]) conv4_2
(512,[302,302,512]) conv4_3
(512,[398,398,512]) conv5_1
(512,[390,390,512]) conv5_2
(512,[379,379,512]) conv5_3

[cicido]

name: "conv1_1"
num_output: 64
name: "conv1_2_V"
num_output: 22
name: "conv1_2_H"
num_output: 22
name: "conv1_2_P"
num_output: 55
name: "conv2_1_V"
num_output: 49
name: "conv2_1_H"
num_output: 49
name: "conv2_1_P"
num_output: 117
name: "conv2_2_V"
num_output: 62
name: "conv2_2_H"
num_output: 62
name: "conv2_2_P"
num_output: 119
name: "conv3_1_V"
num_output: 110
name: "conv3_1_H"
num_output: 110
name: "conv3_1_P"
num_output: 230
name: "conv3_2_V"
num_output: 118
name: "conv3_2_H"
num_output: 118
name: "conv3_2_P"
num_output: 223
name: "conv3_3_V"
num_output: 141
name: "conv3_3_H"
num_output: 141
name: "conv3_3_P"
num_output: 256
name: "conv4_1_V"
num_output: 233
name: "conv4_1_H"
num_output: 233
name: "conv4_1_P"
num_output: 489
name: "conv4_2_V"
num_output: 256
name: "conv4_2_H"
num_output: 256
name: "conv4_2_P"
num_output: 470
name: "conv4_3_V"
num_output: 302
name: "conv4_3_H"
num_output: 302
name: "conv4_3_P"
num_output: 512
name: "conv5_1_V"
num_output: 398
name: "conv5_1_H"
num_output: 398
name: "conv5_1_P"
num_output: 512
name: "conv5_2_V"
num_output: 390
name: "conv5_2_H"
num_output: 390
name: "conv5_2_P"
num_output: 512
name: "conv5_3_V"
num_output: 379
name: "conv5_3_H"
num_output: 379
name: "conv5_3_P"
num_output: 512

[cicido]

should the parameter of layer conv*_V and conv*_H have the same num_output?

[zhaobaozi]

@cicido @yihui-he hi after run run ./combine.sh | xargs ./calflop.sh,I get the temp/cb_3c_3C4x_mem_bn_vgg.prototxt and cb_3c_vgg.caffemodel..so i run ./caffe/build/tools/caffe train -solver temp/solver.prototxt -weights temp/cb_3c_vgg.caffemodel -gpu $0 ..but i meet the same problem.. Cannot copy param 0 weights from layer 'conv5_1_H'; shape mismatch. Source param shape is 398 398 1 3 (475212); target param shape is 512 398 1 3 (611328). have you finish it??

[ethanhe42]

I downloaded my code and ran again. However, I failed to reproduce these problems.

[Ai-is-light]

@cicido @yihui-he @zhaobaozi would you mind sharing your ImageNet data,
like ,examples/imagenet/ilsvrc12_val_lmdb". I'm really instersted in this good work. but I have not the ImageNet dataset. Thank.

[jiaqun123]

@yihui-he
hi, I would like to know whether the zero weights after finetuning update or still maintain zero value. Hope to get your reply.

[JingliangGao]

@Ai-is-light I can give you

[JingliangGao]

I Input the command "./combine.sh | xargs ./calflop.sh", but meet this problem .

F0615 16:06:04.220142 25028 data_transformer.cpp:168] Check failed: height <= datum_height (224 vs. 107)

Please tell me how to solve the problem.

[JingliangGao]

@yihui-he

[JingliangGao]

@cicido @fgxfxpfzzfcc @zhaobaozi @Ai-is-light

[sunny5555]

I am running python3 train.py -action c3 -caffe [GPU1]，encountered this problem:*** Check failure stack trace: ***
How did you solve it?thank you very much.
@cicido