VGG-16 - 2.0 English

Baseline Model

VGG is a network for large-scale image recognition. Refer to https://arxiv.org/abs/1409.1556 for the architecture of the VGG16.

Creating a Configuration File

Create a file named config.prototxt:

workspace: "examples/decent_p/"
gpu: "0,1,2,3"
test_iter: 100
acc_name: "top-1"
 
model: "examples/decent_p/vgg.prototxt"
weights: "examples/decent_p/vgg.caffemodel"
solver: "examples/decent_p/solver.prototxt"
 
rate: 0.1
 
pruner {
  method: REGULAR
}

Performing Model Analysis

$ ./vai_p_caffe ana –config config.prototxt

Pruning the Model

$ ./vai_p_caffe prune –config config.prototxt

Fine-tuning the Pruned Model

You can use the following solver settings to perform fine-tuning:

net: "vgg16/train_val.prototxt"
test_iter: 1250
test_interval: 1000
test_initialization: true
display: 100
average_loss: 100
base_lr: 0.004
lr_policy: "poly"
power: 1
gamma: 0.1
max_iter: 500000
momentum: 0.9
weight_decay: 0.0001
snapshot: 1000
snapshot_prefix: "vgg16/snapshot/res"
solver_mode: GPU
iter_size: 1

Use the following command to start fine-tuning:

$ ./vai_p_caffe finetune -config config.prototxt

Estimated time required: about 70 hours for 30 epochs using training set of ImageNet (ILSVRC2012) (1.2 million images, 4 x NVIDIA Tesla V100).

Getting Final Output

After a few pruning iterations, a pruned model with only 33% of required operations relative to the baseline is obtained.

To finalize the model, run the following:

$ ./vai_p_caffe transform –model baseline.prototxt –weights finetuned_model.caffemodel -output
final.caffemodel

Pruning Results

Dataset

ImageNet (ILSVRC2012)

Input Size

224 x 224

GPU Platform

4 x NVIDIA Tesla V100

FLOPs

30G

#Parameters

24M