VGG net
How to set up the libtorch environment based on Visual Studio
https://blog.csdn.net/defi_wang/article/details/107450428
Code Introduce
https://blog.csdn.net/defi_wang/article/details/107889818 https://blog.csdn.net/defi_wang/article/details/108032208
Convert Image to Tensor
https://blog.csdn.net/defi_wang/article/details/107936757
How to run it?
VGGNet [options] command [arg...]
Commands
| command | description |
|---|---|
| state | Show the net state |
| train | train the network |
| verify | verify the pre-trained network with the test-set |
| classify | classify an input picture with the pre-trained network |
| help | show the help information |
options
| option | description |
|---|---|
| -v | Verbose mode to output more logs |
| -y | Proceed the operation without any prompt |
| -t nntype | Specify the neutral network mode, please see the "neutral network" table |
| -r image_set_root_path | specify the image set root path |
| -i imgset_type | specify the image-set type, at default, it is folder image set, please see the "image set type" table |
neutral network
| type | description | type | description |
|---|---|---|---|
| LENET | LeNet5 | RESNET18 | ResNet-18 |
| VGGA | VGG-A | RESNET34 | ResNet-34 |
| VGGA_LRN | VGG-A LRN | RESNET50 | ResNet-50 |
| VGGB | VGG-B | RESNET101 | ResNet-101 |
| VGGC | VGG-C | RESNET152 | ResNet-152 |
| VGGD | VGG-D | ||
| VGGE | VGG-E |
image set type
| type | description |
|---|---|
| folder | the train/test image set are organized with folder/files |
| MNIST | the MNIST hand-writing train and test set |
| CIFAR10 | The CIFAR10 image set |
| CIFAR100 | The CIFAR100 image set |
arguments for command
state
VGGNet state [--bn/-batchnorm] [-n numclass] [-s/--smallsize] [train_output]
If no arg is specified, it will print the VGG-D net at default.
examples:
VGGNet.exe state --bn --numclass 10 --smallsize
print the neutral network state with batchnorm layers, the output number of classes and use the 32x32 small input image instead the 224x224 image.
VGGNet.exe I:\catdog.pt
print the information of neutral network loading from I:\catdog.pt.
args
| name | shortname | arg | description |
|---|---|---|---|
| batchnorm bn |
n/a | n/a | enable batchnorm after CNN |
| numclass | n | num of classes | The specified final number of classes, the default value is 1000 |
| smallsize | s | n/a | Use 32x32 input instead of the original 224*224 |
train
NNUtil [-v ] [-y] [-r image_set_root_path] [-i imgset_type] train [train_output] [-b/--batchsize batchsize] [-e/--epochnum epochnum] [-l/--learningrate fixed_learningrate] [--bn/--batchnorm] [-n numclass] [-s/--smallsize] [--showloss once_num_batch] [--clean]
args
| name | shortname | arg | description |
|---|---|---|---|
| batchsize | b | batchsize | the batch size of sending to network |
| epochnum | e | epochnum | the number of train epochs |
| learningrate | l | learning rate | the fixed learning rate (*)if it is not specified, default learning rate is used, dynamic learning rate is used |
| lrdm | n/a | learning rate decay mode | learning rate manager mode, see learning rate decay mode table |
| lr_decay_steps | n/a | decay_steps | learning rate decay steps |
| lr_decay_rate | n/a | decay_rate | learning rate decay rate |
| lr_end | n/a | end_learning_rate | end at the specified learning rate |
| lr_cycle | n/a | the learning rate decline, and rise again, true or false | |
| lr_staircase | n/a | use the floor value for (global_step/decay_steps), true or false | |
| lr_alpha | n/a | the parameter used in cosine decay | |
| lr_beta | n/a | the parameter used in linear cosine decay | |
| lr_num_periods | n/a | fade cosine periods | |
| lr_power | n/a | the power used in polynomial decay | |
| lr_initial_variance | n/a | the noise initial variance | |
| lr_variance_decay | n/a | the decay noise variance | |
| batchnorm bn |
n/a | n/a | enable batchnorm after CNN |
| numclass | n | num of classes | The specified final number of classes, the default value is 1000 |
| smallsize | s | n/a | Use 32x32 input instead of the original 224*224 |
| showloss | n/a | once_num_batch | stat. the loss every num batch |
| clean | n/a | n/a | clean the previous pre-trained net state file |
| optimizer | o | optimizer | specify the optimizer, at default SGD will be used |
| weight_decay | w | weight decay for optimizer | the L2 Regularization |
| momentum | m | momentum | momentum |
| dampening | n/a | dampening | dampening |
| nesterov | n/a | nesterov | nesterov |
learning rate decay mode
| learning rate decay mode | name |
|---|---|
| exponent | exponent decay, lr = initial_learning_rate*decay_rate^(global_step/decay_steps) |
| natural_exp | natural exp decay |
| polynomial | polynomial decay |
| inversetime | inverse time decay |
| cosine | cosine decay |
| lcosine | linear cosine decay |
Train a network with the specified train image set, and the image set folder structure is like as
{image_set_root_path}
|-training_set
| |-tag1
| | |-- images......
| |-tag2
| | |-- images......
| ......
|_test_set
|-tag1
| |-- images......
Examples
NNUtil -r I:\CatDog train I:\catdog.pt --bn -b 64 -l 0.0001 --showloss 10
Train the image set lies at I:\CatDog, and save the output to I:\catdog.pt, the batchnorm layers will be introduced, and the batch size is 64, the learning rate use the fixed 0.0001, and show the loss rate every 10 batches.
NNUtil -t RESNET18 -r I:\CIFAR\cifar-10-batches-bin -d CIFAR10 train I:\cifar_resnet18.pt -b 64 -l 0.0001 --showloss 10 --optim adam
Train the CIFAR10 image set, and save the train result to I:\cifar_resnet18.pt, batch size is 64, learning rate is 0.0001, and show the loss every 10 batches, and use adam optimizer.
verify
NNUtil [-v] [-y] [-r image_set_root_pah] [-i imgset_type] verify image_set_root_path pretrain_network* Verify the test-set and show the pass-rate and other information
NNUtil -r I:\CatDog verify I:\catdog.pt
classify
NNUtil classify pretrain_network image_file_path* With the specified pre-trained network, classify a image.
NNUtil classify I:\catdog.pt PIC_001.png