Support for Image Input?

[cwozny]

I'm trying to adapt the example in the readme markdown to accept a grayscale input image with a resolution of 512x448 in the range 0-1 and an output of 8 values between 0-1. My dimensions seem to be incompatible somewhere along the line, but I can't figure out exactly where.

const int WIDTH = 512;
const int HEIGHT = 448;
const int DEPTH = 100;
const int KEYPAD = 8;

Eigen::MatrixXd screen(WIDTH*HEIGHT,DEPTH);
Eigen::MatrixXd keypad(KEYPAD,DEPTH);
Eigen::MatrixXd pred;

// Construct a network object
MiniDNN::Network net;

// Create three layers
// Layer 1 -- convolutional, input size 512x448x1, 1 output channels, filter size 5x5
MiniDNN::Layer* layer1 = new MiniDNN::Convolutional<MiniDNN::ReLU>(WIDTH, HEIGHT, 1, 1, 5, 5);
// Layer 2 -- max pooling, input size 16x16x1, pooling window size 3x3
MiniDNN::Layer* layer2 = new MiniDNN::MaxPooling<MiniDNN::ReLU>(16, 16, 1, 3, 3);
// Layer 3 -- fully connected, input size 5x5x1, output size 8
MiniDNN::Layer* layer3 = new MiniDNN::FullyConnected<MiniDNN::Identity>(5 * 5 * 1, KEYPAD);

// Add layers to the network object
net.add_layer(layer1);
net.add_layer(layer2);
net.add_layer(layer3);

// Set output layer
net.set_output(new MiniDNN::RegressionMSE());

// Create optimizer object
MiniDNN::RMSProp opt;
opt.m_lrate = 0.001;

// (Optional) set callback function object
MiniDNN::VerboseCallback callback;
net.set_callback(callback);

net.init(0, 0.01, 123);

// Populate observations and responses here...
...
...

net.fit(opt, screen, keypad, 100, 10, 123);

The above code when executed yields to the following error:

terminate called after throwing an instance of 'std::invalid_argument'
what(): Unit sizes do not match
Aborted (core dumped)

Any insight on what I am doing incorrectly @giovastabile or @yixuan?

[TrevorBlythe]

Is layer ones output size the same as layer twos input size???