| 60 | |
| 61 | |
| 62 | int main() try |
| 63 | { |
| 64 | // Let's instantiate our network in train mode. |
| 65 | model::train net; |
| 66 | |
| 67 | // We create a new scope so that resources from the loaded network are freed |
| 68 | // automatically when leaving the scope. |
| 69 | { |
| 70 | // Now, let's define the classic ResNet50 network and load the pretrained model on |
| 71 | // ImageNet. |
| 72 | resnet::train_50 resnet50; |
| 73 | std::vector<string> labels; |
| 74 | deserialize("resnet50_1000_imagenet_classifier.dnn") >> resnet50 >> labels; |
| 75 | |
| 76 | // For transfer learning, we are only interested in the ResNet50's backbone, which |
| 77 | // lays below the loss and the fc layers, so we can extract it as: |
| 78 | auto backbone = std::move(resnet50.subnet().subnet()); |
| 79 | |
| 80 | // We can now assign ResNet50's backbone to our network skipping the different |
| 81 | // layers, in our case, the loss layer and the fc layer: |
| 82 | net.subnet().subnet() = backbone; |
| 83 | |
| 84 | // An alternative way to use the pretrained network on a different |
| 85 | // network is to extract the relevant part of the network (we remove |
| 86 | // loss and fc layers), stack the new layers on top of it and assign |
| 87 | // the network. |
| 88 | using net_type = loss_metric<fc_no_bias<128, decltype(backbone)>>; |
| 89 | net_type net2; |
| 90 | net2.subnet().subnet() = backbone; |
| 91 | } |
| 92 | |
| 93 | // We can use the visit_layers function to modify the weight decay of the entire |
| 94 | // network: |
| 95 | visit_computational_layers(net, visitor_weight_decay_multiplier(0.001)); |
| 96 | |
| 97 | // We can also use predefined visitors to affect the learning rate of the whole |
| 98 | // network. |
| 99 | set_all_learning_rate_multipliers(net, 0.5); |
| 100 | |
| 101 | // Modifying the learning rates of a network is a common practice for fine tuning, for |
| 102 | // this reason it is already provided. However, it is implemented internally using a |
| 103 | // visitor that is very similar to the one defined in this example. |
| 104 | |
| 105 | // Usually, we want to freeze the network, except for the top layers: |
| 106 | visit_computational_layers(net.subnet().subnet(), visitor_weight_decay_multiplier(0)); |
| 107 | set_all_learning_rate_multipliers(net.subnet().subnet(), 0); |
| 108 | |
| 109 | // Alternatively, we can use the visit_layers_range to modify only a specific set of |
| 110 | // layers: |
| 111 | visit_computational_layers_range<0, 2>(net, visitor_weight_decay_multiplier(1)); |
| 112 | |
| 113 | // Sometimes we might want to set the learning rate differently throughout the network. |
| 114 | // Here we show how to use adjust the learning rate at the different ResNet50's |
| 115 | // convolutional blocks: |
| 116 | set_learning_rate_multipliers_range< 0, 2>(net, 1); |
| 117 | set_learning_rate_multipliers_range< 2, 38>(net, 0.1); |
| 118 | set_learning_rate_multipliers_range< 38, 107>(net, 0.01); |
| 119 | set_learning_rate_multipliers_range<107, 154>(net, 0.001); |
nothing calls this directly
no test coverage detected