// The contents of this file are in the public domain. See LICENSE_FOR_EXAMPLE_PROGRAMS.txt
/*
This is an example illustrating the use of the deep learning tools from the
dlib C++ Library. I'm assuming you have already read the dnn_introduction_ex.cpp
example. So in this example program I'm going to go over a number of more
advanced parts of the API, including:
- Using multiple GPUs
- Training on large datasets that don't fit in memory
- Defining large networks
- Accessing and configuring layers in a network
*/
#include <dlib/dnn.h>
#include <iostream>
#include <dlib/data_io.h>
using namespace std;
using namespace dlib;
// ----------------------------------------------------------------------------------------
// Let's start by showing how you can conveniently define large and complex
// networks. The most important tool for doing this are C++'s alias templates.
// These let us define new layer types that are combinations of a bunch of other
// layers. These will form the building blocks for more complex networks.
// So let's begin by defining the building block of a residual network (see
// Figure 2 in Deep Residual Learning for Image Recognition by He, Zhang, Ren,
// and Sun). We are going to decompose the residual block into a few alias
// statements. First, we define the core block.
// Here we have parameterized the "block" layer on a BN layer (nominally some
// kind of batch normalization), the number of filter outputs N, and the stride
// the block operates at.
template <
int N,
template <typename> class BN,
int stride,
typename SUBNET
>
using block = BN<con<N,3,3,1,1,relu<BN<con<N,3,3,stride,stride,SUBNET>>>>>;
// Next, we need to define the skip layer mechanism used in the residual network
// paper. They create their blocks by adding the input tensor to the output of
// each block. So we define an alias statement that takes a block and wraps it
// with this skip/add structure.
// Note the tag layer. This layer doesn't do any computation. It exists solely
// so other layers can refer to it. In this case, the add_prev1 layer looks for
// the tag1 layer and will take the tag1 output and add it to the input of the
// add_prev1 layer. This combination allows us to implement skip and residual
// style networks. We have also set the block stride to 1 in this statement.
// The significance of that is explained next.
template <
template <int,template<typename>class,int,typename> class block,
int N,
template<typename>class BN,
typename SUBNET
>
using residual = add_prev1<block<N,BN,1,tag1<SUBNET>>>;
// Some residual blocks do downsampling. They do this by using a stride of 2
// instead of 1. However, when downsampling we need to also take care to
// downsample the part of the network that adds the original input to the output
// or the sizes won't make sense (the network will still run, but the results
// aren't as good). So here we define a downsampling version of residual. In
// it, we make use of the skip1 layer. This layer simply outputs whatever is
// output by the tag1 layer. Therefore, the skip1 layer (there are also skip2,
// skip3, etc. in dlib) allows you to create branching network structures.
// residual_down creates a network structure like this:
/*
input from SUBNET
/ \
/ \
block downsample(using avg_pool)
\ /
\ /
add tensors (using add_prev2 which adds the output of tag2 with avg_pool's output)
|
output
*/
template <
template <int,template<typename>class,int,typename> class block,
int N,
template<typename>class BN,
typename SUBNET
>
using residual_down = add_prev2<avg_pool<2,2,2,2,skip1<tag2<block<N,BN,2,tag1<SUBNET>>>>>>;
// Now we can define 4 different residual blocks we will use in this example.
// The first two are non-downsampling residual blocks while the last two
// downsample. Also, res and res_down use batch normalization while ares and
// ares_down have had the batch normalization replaced with simple affine
// layers. We will use the affine version of the layers when testing our
// networks.
template <typename SUBNET> using res = relu<residual<block,8,bn_con,SUBNET>>;
template <typename SUBNET> using ares = relu<residual<block,8,affine,SUBNET>>;
template <typename SUBNET> using res_down = relu<residual_down<block,8,bn_con,SUBNET>>;
template <typename SUBNET> using ares_down = relu<residual_down<block,8,affine,SUBNET>>;
// Now that we have these convenient aliases, we can define a residual network
// without a lot of typing. Note the use of a repeat layer. This special layer
// type allows us to type repeat<9,res,SUBNET> instead of
// res<res<res<res<res<res<res<res<res<SUBNET>>>>>>>>>. It will also prevent
// the compiler from complaining about super deep template nesting when creating
// large networks.
const unsigned long number_of_classes = 10;
using net_type = loss_multiclass_log<fc<number_of_classes,
avg_pool_everything<
res<res<res<res_down<
repeat<9,res, // repeat this layer 9 times
res_down<
res<
input<matrix<unsigned char>>
>>>>>>>>>>;
// And finally, let's define a residual network building block that uses
// parametric ReLU units instead of regular ReLU.
template <typename SUBNET>
using pres = prelu<add_prev1<bn_con<con<8,3,3,1,1,prelu<bn_con<con<8,3,3,1,1,tag1<SUBNET>>>>>>>>;
// ----------------------------------------------------------------------------------------
int main(int argc, char** argv) try
{
if (argc != 2)
{
cout << "This example needs the MNIST dataset to run!" << endl;
cout << "You can get MNIST from http://yann.lecun.com/exdb/mnist/" << endl;
cout << "Download the 4 files that comprise the dataset, decompress them, and" << endl;
cout << "put them in a folder. Then give that folder as input to this program." << endl;
return 1;
}
std::vector<matrix<unsigned char>> training_images;
std::vector<unsigned long> training_labels;
std::vector<matrix<unsigned char>> testing_images;
std::vector<unsigned long> testing_labels;
load_mnist_dataset(argv[1], training_images, training_labels, testing_images, testing_labels);
// dlib uses cuDNN under the covers. One of the features of cuDNN is the
// option to use slower methods that use less RAM or faster methods that use
// a lot of RAM. If you find that you run out of RAM on your graphics card
// then you can call this function and we will request the slower but more
// RAM frugal cuDNN algorithms.
set_dnn_prefer_smallest_algorithms();
// Create a network as defined above. This network will produce 10 outputs
// because that's how we defined net_type. However, fc layers can have the
// number of outputs they produce changed at runtime.
net_type net;
// So if you wanted to use the same network but override the number of
// outputs at runtime you can do so like this:
net_type net2(num_fc_outputs(15));
// Now, let's imagine we wanted to replace some of the relu layers with
// prelu layers. We might do it like this:
using net_type2 = loss_multiclass_log<fc<number_of_classes,
avg_pool_everything<
pres<res<res<res_down< // 2 prelu layers here
tag4<repeat<9,pres, // 9 groups, each containing 2 prelu layers
res_down<
res<
input<matrix<unsigned char>>
>>>>>>>>>>>;
// prelu layers have a floating point parameter. If you want to set it to
// something other than its default value you can do so like this:
net_type2 pnet(prelu_(0.2),
prelu_(0.25),
repeat_group(prelu_(0.3),prelu_(0.4)) // Initialize all the prelu instances in the repeat
// layer. repeat_group() is needed to group the
// things that are part of repeat's block.
);
// As you can see, a network will greedily assign things given to its
// constructor to the layers inside itself. The assignment is done in the
// order the layers are defined, but it will skip layers where the
// assignment doesn't make sense.
// Now let's print the details of the pnet to the screen and inspect it.
cout << "The pnet has " << pnet.num_layers << " layers in it." << endl;
cout << pnet << endl;
// These print statements will output this (I've truncated it since it's
// long, but you get the idea):
/*
The pnet has 131 layers in it.
layer<0> loss_multiclass_log
layer<1> fc (num_outputs=10) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<2> avg_pool (nr=0, nc=0, stride_y=1, stride_x=1, padding_y=0, padding_x=0)
layer<3> prelu (initial_param_value=0.2)
layer<4> add_prev1
layer<5> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<6> con (num_filters=8, nr=3, nc=3, stride_y=1, stride_x=1, padding_y=1, padding_x=1) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<7> prelu (initial_param_value=0.25)
layer<8> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<9> con (num_filters=8, nr=3, nc=3, stride_y=1, stride_x=1, padding_y=1, padding_x=1) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<10> tag1
...
layer<34> relu
layer<35> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<36> con (num_filters=8, nr=3, nc=3, stride_y=2, stride_x=2, padding_y=0, padding_x=0) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<37> tag1
layer<38> tag4
layer<39> prelu (initial_param_value=0.3)
layer<40> add_prev1
layer<41> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
...
layer<118> relu
layer<119> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<120> con (num_filters=8, nr=3, nc=3, stride_y=2, stride_x=2, padding_y=0, padding_x=0) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<121> tag1
layer<122> relu
layer<123> add_prev1
layer<124> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<125> con (num_filters=8, nr=3, nc=3, stride_y=1, stride_x=1, padding_y=1, padding_x=1) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<126> relu
layer<127> bn_con eps=1e-05 learning_rate_mult=1 weight_decay_mult=0 bias_learning_rate_mult=1 bias_weight_decay_mult=1
layer<128> con (num_filters=8, nr=3, nc=3, stride_y=1, stride_x=1, padding_y=1, padding_x=1) learning_rate_mult=1 weight_decay_mult=1 bias_learning_rate_mult=1 bias_weight_decay_mult=0
layer<129> tag1
layer<130> input<matrix>
*/
// Now that we know the index numbers for each layer, we can access them
// individually using layer<index>(pnet). For example, to access the output
// tensor for the first prelu layer we can say:
layer<3>(pnet).get_output();
// Or to print the prelu parameter for layer 7 we can say:
cout << "prelu param: "<< layer<7>(pnet).layer_details().get_initial_param_value() << endl;
// We can also access layers by their type. This next statement finds the
// first tag1 layer in pnet, and is therefore equivalent to calling
// layer<10>(pnet):
layer<tag1>(pnet);
// The tag layers don't do anything at all and exist simply so you can tag
// parts of your network and access them by layer<tag>(). You can also
// index relative to a tag. So for example, to access the layer immediately
// after tag4 you can say:
layer<tag4,1>(pnet); // Equivalent to layer<38+1>(pnet).
// Or to access the layer 2 layers after tag4:
layer<tag4,2>(pnet);
// Tagging is a very useful tool for making complex network structures. For
// example, the add_prev1 layer is implemented internally by using a call to
// layer<tag1>().
// Ok, that's enough talk about defining and inspecting networks. Let's
// talk about training networks!
// The dnn_trainer will use SGD by default, but you can tell it to use
// different solvers like adam with a weight decay of 0.0005 and the given
// momentum parameters.
dnn_trainer<net_type,adam> trainer(net,adam(0.0005, 0.9, 0.999));
// Also, if you have multiple graphics cards you can tell the trainer to use
// them together to make the training faster. For example, replacing the
// above constructor call with this one would cause it to use GPU cards 0
// and 1.
//dnn_trainer<net_type,adam> trainer(net,adam(0.0005, 0.9, 0.999), {0,1});
trainer.be_verbose();
// While the trainer is running it keeps an eye on the training error. If
// it looks like the error hasn't decreased for the last 2000 iterations it
// will automatically reduce the learning rate by 0.1. You can change these
// default parameters to some other values by calling these functions. Or
// disable the automatic shrinking entirely by setting the shrink factor to 1.
trainer.set_iterations_without_progress_threshold(2000);
trainer.set_learning_rate_shrink_factor(0.1);
// The learning rate will start at 1e-3.
trainer.set_learning_rate(1e-3);
trainer.set_synchronization_file("mnist_resnet_sync", std::chrono::seconds(100));
// Now, what if your training dataset is so big it doesn't fit in RAM? You
// make mini-batches yourself, any way you like, and you send them to the
// trainer by repeatedly calling trainer.train_one_step().
//
// For example, the loop below stream MNIST data to out trainer.
std::vector<matrix<unsigned char>> mini_batch_samples;
std::vector<unsigned long> mini_batch_labels;
dlib::rand rnd(time(0));
// Loop until the trainer's automatic shrinking has shrunk the learning rate to 1e-6.
// Given our settings, this means it will stop training after it has shrunk the
// learning rate 3 times.
while(trainer.get_learning_rate() >= 1e-6)
{
mini_batch_samples.clear();
mini_batch_labels.clear();
// make a 128 image mini-batch
while(mini_batch_samples.size() < 128)
{
auto idx = rnd.get_random_32bit_number()%training_images.size();
mini_batch_samples.push_back(training_images[idx]);
mini_batch_labels.push_back(training_labels[idx]);
}
// Tell the trainer to update the network given this mini-batch
trainer.train_one_step(mini_batch_samples, mini_batch_labels);
// You can also feed validation data into the trainer by periodically
// calling trainer.test_one_step(samples,labels). Unlike train_one_step(),
// test_one_step() doesn't modify the network, it only computes the testing
// error which it records internally. This testing error will then be print
// in the verbose logging and will also determine when the trainer's
// automatic learning rate shrinking happens. Therefore, test_one_step()
// can be used to perform automatic early stopping based on held out data.
}
// When you call train_one_step(), the trainer will do its processing in a
// separate thread. This allows the main thread to work on loading data
// while the trainer is busy executing the mini-batches in parallel.
// However, this also means we need to wait for any mini-batches that are
// still executing to stop before we mess with the net object. Calling
// get_net() performs the necessary synchronization.
trainer.get_net();
net.clean();
serialize("mnist_res_network.dat") << net;
// Now we have a trained network. However, it has batch normalization
// layers in it. As is customary, we should replace these with simple
// affine layers before we use the network. This can be accomplished by
// making a network type which is identical to net_type but with the batch
// normalization layers replaced with affine. For example:
using test_net_type = loss_multiclass_log<fc<number_of_classes,
avg_pool_everything<
ares<ares<ares<ares_down<
repeat<9,ares,
ares_down<
ares<
input<matrix<unsigned char>>
>>>>>>>>>>;
// Then we can simply assign our trained net to our testing net.
test_net_type tnet = net;
// Or if you only had a file with your trained network you could deserialize
// it directly into your testing network.
deserialize("mnist_res_network.dat") >> tnet;
// And finally, we can run the testing network over our data.
std::vector<unsigned long> predicted_labels = tnet(training_images);
int num_right = 0;
int num_wrong = 0;
for (size_t i = 0; i < training_images.size(); ++i)
{
if (predicted_labels[i] == training_labels[i])
++num_right;
else
++num_wrong;
}
cout << "training num_right: " << num_right << endl;
cout << "training num_wrong: " << num_wrong << endl;
cout << "training accuracy: " << num_right/(double)(num_right+num_wrong) << endl;
predicted_labels = tnet(testing_images);
num_right = 0;
num_wrong = 0;
for (size_t i = 0; i < testing_images.size(); ++i)
{
if (predicted_labels[i] == testing_labels[i])
++num_right;
else
++num_wrong;
}
cout << "testing num_right: " << num_right << endl;
cout << "testing num_wrong: " << num_wrong << endl;
cout << "testing accuracy: " << num_right/(double)(num_right+num_wrong) << endl;
}
catch(std::exception& e)
{
cout << e.what() << endl;
}