LAP/torchlap/experiments/entrypoint.py

import sys
from typing import Tuple
import argparse
import os
from abc import ABC, abstractmethod
from ..models.model import Model
from ..configs.base_config import BaseConfig, PhaseType


class BaseEntrypoint(ABC):
    """Base class for all entrypoints.
    """
    description = ''

    def __init__(self, phase_type: PhaseType) -> None:
        super().__init__()
        self.phase_type = phase_type

        self.conf, self.model = self._get_conf_model()
        self.parser = self._create_parser(sys.argv[0], sys.argv[1])
        self.conf.update(self.parser.parse_args(sys.argv[2:]))

    def _create_parser(self, command: str, entrypoint_name: str) -> argparse.ArgumentParser:
        parser = argparse.ArgumentParser(
            prog=f'{os.path.basename(command)} {entrypoint_name}',
            description=self.description or None
        )
        parser.add_argument('--device', default='cpu', type=str,
                            help='Analysis will run over device, use cpu, cuda:#, cuda (for all gpus)')

        parser.add_argument('--samples-dir', default=None, type=str,
                            help='The directory/dataGroup to be evaluated. In the case of dataGroup can be train/test/val. In the case of directory must contain 0 and 1 subdirectories. Use:FilterName to do over samples containing /FilterName/ in their path')

        parser.add_argument('--final-model-dir', default=None, type=str,
                            help='The directory to load the model from, when not given will be calculated!')

        parser.add_argument('--save-dir', default=None, type=str,
                            help='The directory to save the model from, when not given will be calculated!')

        parser.add_argument('--report-dir', default=None, type=str,
                            help='The dir to save reports per slice per sample in.')

        parser.add_argument('--epoch', default=None, type=str,
                            help='The epoch to load.')

        parser.add_argument('--try-name', default=None, type=str,
                            help='The run name specifying what run is doing')

        parser.add_argument('--try-num', default=None, type=int,
                            help='The try number to load')

        parser.add_argument('--data-separation', default=None, type=str,
                            help='The data_separation to be used.')

        parser.add_argument('--batch-size', default=None, type=int,
                            help='The batch size to be used.')

        parser.add_argument('--pretrained-model-file', default=None, type=str,
                            help='Address of .pt pretrained model')

        parser.add_argument('--max-epochs', default=None, type=int,
                            help='The maximum epochs of training!')

        parser.add_argument('--big-batch-size', default=None, type=int,
                            help='The big batch size (iteration per optimization)!')

        parser.add_argument('--iters-per-epoch', default=None, type=int,
                            help='The number of big batches per epoch!')

        parser.add_argument('--interpretation-method', default=None, type=str,
                            help='The method used for interpreting the results!')

        parser.add_argument('--cut-threshold', default=None, type=float,
                            help='The threshold for cutting interpretations!')

        parser.add_argument('--global-threshold', action='store_true',
                            help='Whether the given cut threshold must be applied global or to the relative values!')

        parser.add_argument('--dynamic-threshold', action='store_true',
                            help='Whether to use dynamic threshold in interpretation!')

        parser.add_argument('--class-label-for-interpretation', default=None, type=int,
                            help='The class label we want to explain why it has been chosen. None means the decision of the model!')

        parser.add_argument('--interpret-predictions-vs-gt', default='1', type=(lambda x: x == '1'),
                            help='If the class_label for_interpretation is None this will be considered. If 1, interpretations would be done for the predicted label, otherwise for the ground truth.')

        parser.add_argument('--mapped-labels-to-use', default=None, type=(lambda x: [int(y) for y in x.split(',')]),
                            help='The labels to do the analysis on them only (comma separated), default is all the labels.')

        parser.add_argument('--skip-overlay', action='store_true', default=None,
                            help='Passing this flag prevents the interpretation phase from storing overlay images.')

        parser.add_argument('--skip-raw', action='store_true', default=None,
                            help='Passing this flag prevents the interpretation phase from storing the raw interpretation values.')

        parser.add_argument('--overlay-only', action='store_true',
                            help='Passing this flag makes the interpretation phase to store just overlay images '
                            'and not the `.npy` files.')

        parser.add_argument('--save-by-file-name', action='store_true',
                            help='Saves sample-specific files by their file names only, not the whole path.')

        parser.add_argument('--n-interpretation-samples', default=None, type=int,
                            help='The max number of samples to be interpreted.')
        
        parser.add_argument('--interpretation-tag-to-evaluate', default=None, type=str,
                            help='The tag to be used as interpretation for evaluation phase, otherwise the first one will be used.')
        return parser

    @abstractmethod
    def _get_conf_model(self) -> Tuple[BaseConfig, Model]:
        pass