Ejemplo n.º 1
0
def parse_command_line():
    parser = argparse.ArgumentParser(description="Trains Caffe model against prepared data")
    parser.add_argument("--log_path", help="The path to where to place log files",
        type=str, default="logs")
    parser.add_argument("--log_num", help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""", type=int,
        default=1)
    parser.add_argument("--note", help="Adds extra note into training logs.", type=str,
        default=None)
    parser.add_argument("--solver", help="The path to our Caffe solver prototxt file",
        type=str, default="src/caffe_model/bvlc_alexnet/solver.prototxt")
    parser.add_argument("--input_weight_file", help="""A pre-trained Caffe model that we will use
        to start training with in order to fine-tune from""", type=str,
        default="src/caffe_model/bvlc_alexnet/bvlc_alexnet.caffemodel")
    parser.add_argument("--output_weight_file", help="""Where to place the final, trained Caffe
        model""", type=str, default="logs/latest_bvlc_alexnet_finetuned.caffemodel")

    args = vars(parser.parse_args())

    caffe_home = utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix, output_log_file) = utils.get_log_path_details(log_path, log_num)

    solver = os.path.abspath(args["solver"])
    input_weight_file = os.path.abspath(args["input_weight_file"])
    output_weight_file = os.path.abspath(args["output_weight_file"])
    train(caffe_home, log_path, output_log_file, solver, input_weight_file, output_weight_file,
        args["note"])
Ejemplo n.º 2
0
def parse_command_line():
    parser = argparse.ArgumentParser(description="""Tests a trained Caffe model to see how well
        it does, generating quality graphs and statistics""")
    parser.add_argument("--log_path", help="The path to where to place log files and graphs",
        type=str, default="logs")
    parser.add_argument("--log_num", help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""", type=int,
        default=1)
    parser.add_argument("--input_weight_file", help="""The trained and fine-tuned Caffe model that
        we will be testing; defaults to the last trained model from train.py""", type=str,
        default="logs/latest_bvlc_alexnet_finetuned.caffemodel")
    parser.add_argument("--note", help="Adds extra note onto generated quality graphs.", type=str,
        default="")
    parser.add_argument("--solver", help="The path to our Caffe solver prototxt file",
        type=str, default="src/caffe_model/bvlc_alexnet/solver.prototxt")
    parser.add_argument("--deploy", help="""Path to our Caffe deploy/inference time prototxt file""",
        type=str, default="src/caffe_model/bvlc_alexnet/deploy.prototxt")
    parser.add_argument("--threshold", help="""The percentage threshold over which we assume
        something is a cloud. Note that this value is from 0.0 to 100.0""", type=float, default=0.1)
    parser.add_argument("--validation_leveldb", help="""Path to where the validation leveldb file is""",
        type=str, default="data/leveldb/validation_leveldb")
    parser.add_argument("--width", help="Width of image during training", type=int, default=256)
    parser.add_argument("--height", help="Height of image during training", type=int, default=256)
    parser.add_argument("--inference_width", help="Width of image during training", type=int,
        default=227)
    parser.add_argument("--inference_height", help="Height of image during training", type=int,
        default=227)
    parser.add_argument("--training_mean_pickle", help="Path to pickled mean values", type=str,
        default="data/imagenet/imagenet_mean.npy")

    args = vars(parser.parse_args())

    print "Testing trained model..."

    caffe_home = utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix, output_log_file) = utils.get_log_path_details(log_path, log_num)
    output_graph_path = output_log_prefix

    (training_details, validation_details) = utils.parse_logs(log_path, output_log_file)

    plot_results(training_details, validation_details, args["note"], output_graph_path, args["solver"])
    validation_leveldb = os.path.abspath(args["validation_leveldb"])
    deploy = os.path.abspath(args["deploy"])
    input_weight_file = os.path.abspath(args["input_weight_file"])
    training_mean_pickle = os.path.abspath(args["training_mean_pickle"])
    predict.test_validation(args["threshold"], output_log_prefix, validation_leveldb,
        deploy, args["width"], args["height"], args["inference_width"],
        args["inference_height"], input_weight_file, training_mean_pickle)
Ejemplo n.º 3
0
def parse_command_line():
    parser = argparse.ArgumentParser(description="""Prepares data for training via Caffe""")
    parser.add_argument("--input_metadata", help="Path to where our bounding box metadata is",
        type=str, default="data/planetlab/metadata/annotated.json")
    parser.add_argument("--input_images", help="Path to where our unbounded images are",
        type=str, default="data/planetlab/metadata")
    parser.add_argument("--output_images", help="Path to place our cropped, bounded images",
        type=str, default="data/planetlab/images/bounded")
    parser.add_argument("--output_leveldb", help="Path to place our prepared leveldb directories",
        type=str, default="data/leveldb")
    parser.add_argument("--width", help="Width of image at training time (it will be scaled to this)",
        type=int, default=256)
    parser.add_argument("--height", help="Height of image at training time (it will be scaled to this)",
        type=int, default=256)
    parser.add_argument("--log_path", help="The path to where to place log files",
        type=str, default="logs")
    parser.add_argument("--log_num", help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""", type=int,
        default=1)
    parser.add_argument("--do_augmentation", help="Whether to do data augmentation",
        dest="do_augmentation", action="store_true")

    parser.set_defaults(do_augmentation=False)
    args = vars(parser.parse_args())

    utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix, output_log_file) = utils.get_log_path_details(log_path, log_num)

    input_metadata = os.path.abspath(args["input_metadata"])
    input_images = os.path.abspath(args["input_images"])
    output_images = os.path.abspath(args["output_images"])
    output_leveldb = os.path.abspath(args["output_leveldb"])
    prepare_data(input_metadata, input_images, output_images, output_leveldb, args["width"],
        args["height"], args["do_augmentation"], output_log_prefix)
Ejemplo n.º 4
0
def parse_command_line():
    parser = argparse.ArgumentParser(
        description="Trains Caffe model against prepared data")
    parser.add_argument("--log_path",
                        help="The path to where to place log files",
                        type=str,
                        default="logs")
    parser.add_argument(
        "--log_num",
        help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""",
        type=int,
        default=1)
    parser.add_argument("--note",
                        help="Adds extra note into training logs.",
                        type=str,
                        default=None)
    parser.add_argument("--solver",
                        help="The path to our Caffe solver prototxt file",
                        type=str,
                        default="src/caffe_model/bvlc_alexnet/solver.prototxt")
    parser.add_argument(
        "--input_weight_file",
        help="""A pre-trained Caffe model that we will use
        to start training with in order to fine-tune from""",
        type=str,
        default="src/caffe_model/bvlc_alexnet/bvlc_alexnet.caffemodel")
    parser.add_argument(
        "--output_weight_file",
        help="""Where to place the final, trained Caffe
        model""",
        type=str,
        default="logs/latest_bvlc_alexnet_finetuned.caffemodel")

    args = vars(parser.parse_args())

    caffe_home = utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix,
     output_log_file) = utils.get_log_path_details(log_path, log_num)

    solver = os.path.abspath(args["solver"])
    input_weight_file = os.path.abspath(args["input_weight_file"])
    output_weight_file = os.path.abspath(args["output_weight_file"])
    train(caffe_home, log_path, output_log_file, solver, input_weight_file,
          output_weight_file, args["note"])
Ejemplo n.º 5
0
def parse_command_line():
    parser = argparse.ArgumentParser(
        description="""Prepares data for training via Caffe""")
    parser.add_argument("--input_metadata",
                        help="Path to where our bounding box metadata is",
                        type=str,
                        default="data/planetlab/metadata/annotated.json")
    parser.add_argument("--input_images",
                        help="Path to where our unbounded images are",
                        type=str,
                        default="data/planetlab/metadata")
    parser.add_argument("--output_images",
                        help="Path to place our cropped, bounded images",
                        type=str,
                        default="data/planetlab/images/bounded")
    parser.add_argument("--output_leveldb",
                        help="Path to place our prepared leveldb directories",
                        type=str,
                        default="data/leveldb")
    parser.add_argument(
        "--width",
        help="Width of image at training time (it will be scaled to this)",
        type=int,
        default=256)
    parser.add_argument(
        "--height",
        help="Height of image at training time (it will be scaled to this)",
        type=int,
        default=256)
    parser.add_argument("--log_path",
                        help="The path to where to place log files",
                        type=str,
                        default="logs")
    parser.add_argument(
        "--log_num",
        help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""",
        type=int,
        default=1)
    parser.add_argument("--do_augmentation",
                        help="Whether to do data augmentation",
                        dest="do_augmentation",
                        action="store_true")

    parser.set_defaults(do_augmentation=False)
    args = vars(parser.parse_args())

    utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix,
     output_log_file) = utils.get_log_path_details(log_path, log_num)

    input_metadata = os.path.abspath(args["input_metadata"])
    input_images = os.path.abspath(args["input_images"])
    output_images = os.path.abspath(args["output_images"])
    output_leveldb = os.path.abspath(args["output_leveldb"])
    prepare_data(input_metadata, input_images, output_images, output_leveldb,
                 args["width"], args["height"], args["do_augmentation"],
                 output_log_prefix)
Ejemplo n.º 6
0
def parse_command_line():
    parser = argparse.ArgumentParser(
        description="""Tests a trained Caffe model to see how well
        it does, generating quality graphs and statistics""")
    parser.add_argument("--log_path",
                        help="The path to where to place log files and graphs",
                        type=str,
                        default="logs")
    parser.add_argument(
        "--log_num",
        help="""Number that will be appended to log files; this will
        be automatically padded and added with zeros, such as output00001.log""",
        type=int,
        default=1)
    parser.add_argument(
        "--input_weight_file",
        help="""The trained and fine-tuned Caffe model that
        we will be testing; defaults to the last trained model from train.py""",
        type=str,
        default="logs/latest_bvlc_alexnet_finetuned.caffemodel")
    parser.add_argument("--note",
                        help="Adds extra note onto generated quality graphs.",
                        type=str,
                        default="")
    parser.add_argument("--solver",
                        help="The path to our Caffe solver prototxt file",
                        type=str,
                        default="src/caffe_model/bvlc_alexnet/solver.prototxt")
    parser.add_argument(
        "--deploy",
        help="""Path to our Caffe deploy/inference time prototxt file""",
        type=str,
        default="src/caffe_model/bvlc_alexnet/deploy.prototxt")
    parser.add_argument("--threshold",
                        help="""The percentage threshold over which we assume
        something is a cloud. Note that this value is from 0.0 to 100.0""",
                        type=float,
                        default=0.1)
    parser.add_argument(
        "--validation_leveldb",
        help="""Path to where the validation leveldb file is""",
        type=str,
        default="data/leveldb/validation_leveldb")
    parser.add_argument("--width",
                        help="Width of image during training",
                        type=int,
                        default=256)
    parser.add_argument("--height",
                        help="Height of image during training",
                        type=int,
                        default=256)
    parser.add_argument("--inference_width",
                        help="Width of image during training",
                        type=int,
                        default=227)
    parser.add_argument("--inference_height",
                        help="Height of image during training",
                        type=int,
                        default=227)
    parser.add_argument("--training_mean_pickle",
                        help="Path to pickled mean values",
                        type=str,
                        default="data/imagenet/imagenet_mean.npy")

    args = vars(parser.parse_args())

    print "Testing trained model..."

    caffe_home = utils.assert_caffe_setup()

    # Ensure the random number generator always starts from the same place for consistent tests.
    random.seed(0)

    log_path = os.path.abspath(args["log_path"])
    log_num = args["log_num"]
    (output_ending, output_log_prefix,
     output_log_file) = utils.get_log_path_details(log_path, log_num)
    output_graph_path = output_log_prefix

    (training_details,
     validation_details) = utils.parse_logs(log_path, output_log_file)

    plot_results(training_details, validation_details, args["note"],
                 output_graph_path, args["solver"])
    validation_leveldb = os.path.abspath(args["validation_leveldb"])
    deploy = os.path.abspath(args["deploy"])
    input_weight_file = os.path.abspath(args["input_weight_file"])
    training_mean_pickle = os.path.abspath(args["training_mean_pickle"])
    predict.test_validation(args["threshold"], output_log_prefix,
                            validation_leveldb, deploy, args["width"],
                            args["height"], args["inference_width"],
                            args["inference_height"], input_weight_file,
                            training_mean_pickle)