configs.md

Configuration files

In order to train and evaluate a model implemented in this project, you must create a configuration file that define these processes. The configs should follow the form general structure

model { ... }

train_config { ... }

train_input_reader { ... }

eval_config { ... }

eval_input_reader { ... }

The easiest way to get starting with your own configs is to modify the sample config located at configs/sample_icnet_resnet_v1.config.

Fine-tuning

Fine-tune ICNet from ResNet-50 ImageNet weights

If you wish to train ICNet from the original ResNet50 weights as done by the author of ICNet, you must first download the associated TF Slim ResNet50 model checkpoints and add point to them in your config file.

First download the TF Slim ResNet-50 weights with

wget http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
tar -zxf resnet_v1_50_2016_08_28.tar.gz -C /tmp

Fine-tuning options are specified in training configs. You must modify two fields within your train_config. We first specify the fine_tune_checkpoint_type as classification (instead of segmentation) and enter the the location of the checkpoint in the fine_tune_checkpoint field. Your config should look like

train_config: {
    ...

    fine_tune_checkpoint_type: "classification"
    fine_tune_checkpoint: "/tmp/resnet_v1_50.ckpt" # resnet50 weights

    ...
}

Fine-tune ICNet from Cityscapes weights

If you wish to use the supplied Cityscapes checkpoints and fine-tune on your own data, then you must specify the fine_tune_checkpoint_type field as segmentation. Your config should be in the form

train_config: {
    ...

    fine_tune_checkpoint_type: "segmentation"
    fine_tune_checkpoint: "/tmp/train/model.ckpt-YYYY" # cityscapes weights

    ...
}