Training Details

[SanthoshRajendiran]

Hello.. We looked into your output, for both 2class segmentation and style transfer. The output is promising. We tried replicating the same.. We could not get a good output even when training is done for 30k epochs, as it is done in the tensorflow deeplab site. Seems we are missing out some training details. It would be helpful if you could share the training details, such as number of epochs needed for decent enough output, for 2 class training. Thanks in advance.

[pschang-phy]

We use dataset from COCO (for those image with person). It has training data 64115 and validation data 2693.

You could download the dataset via script Segmentation/bin/cocoDownload.py

And we have initial training by using config as below

{
    "train_split" : "train",
    "model_variant" : "mobilenet_v2",
    "train_logdir" : "train_log",
    "dataset_dir" : "/home/jovyan/Dataset/COCO17_Person_tfrecord",
    "output_stride" : 16,
    "train_crop_size" : [513, 513],
    "train_batch_size" : 24,
    "depth_multiplier" : 0.5,
    "min_scale_factor" : 0.5,
    "max_scale_factor" : 2.0,
    "scale_factor_step_size" : 0.1,
    "base_learning_rate" : 0.001,
    "momentum" : 0.9,
    "fine_tune_batch_norm" : true,
    "training_number_of_steps" : 200000,

    "DatasetDescriptor" : {
        "name" : "coco17_person",
        "splits_to_sizes" : {
            "train" : 64115,
            "val" : 2693
        },

        "num_classes" : 2,
        "ignore_label" : 255
    }
}

If any question, please feel free let me know.