load_tumor_image_data.py 文件源码

def __init__(self, hdf_filename, test_pct=0.25, neg_bias=0.5, batch_size=64,  normalize=False, malignancy_to_class=None, window_normalize=False):       
        neg_bias = 0.5 if neg_bias is None else neg_bias
        self._hdf_filename        = hdf_filename
        self._neg_bias            = neg_bias
        self._test_pct            = test_pct
        self._batch_size          = batch_size
        self._test_location       = 0
        self._train_location      = 0
        self._test_indices        = []
        self._train_indices       = []
        self._malignancy_to_class = malignancy_to_class
        self._normalize           = normalize
        self._Xmin                = None
        self._Xmax                = None
        self._window_normalize    = window_normalize
        if malignancy_to_class is not None and len(malignancy_to_class) != 6:
            raise Exception("malignancy_class mapping must contain exactly 6 values, one for each malignancy level 0 - 5")
        # Open the hdf file 
        self._hdf_file = h5py.File(self._hdf_filename, 'r')
        # Get info on classes and makeup of the dataset by examining the y values (classes):
        y          = self._hdf_file['nodule_classes'].value        
        if self._malignancy_to_class is not None:
            if malignancy_to_class is not None:
                mal = self._hdf_file['nodule_malignancy']          
                for i in range(len(y)):
                    y[i] = [malignancy_to_class[int(mal[i])]]
        if self._normalize:
            self._Xmin = self._hdf_file['nodule_pixel_min']
            self._Xmax = self._hdf_file['nodule_pixel_max']    
        n_examples = len(y)
        negatives  = [i for i in range(n_examples) if y[i] == [0]]
        positives  = [i for i in range(n_examples) if y[i][0] > 0]
        neg_count  = len(negatives)
        pos_count  = len(positives)
        n_examples = neg_count + pos_count
        neg_goal   = int(min(neg_count, round(neg_bias * n_examples)))
        pos_goal   = n_examples - neg_goal
        if pos_goal > pos_count:
            neg_goal = int(round(pos_count * (1-neg_bias+0.5)))
            pos_goal = pos_count
        # print("Before: neg count: {0}; goal: {1} - pos count: {2}; goal: {3}".format(neg_count, neg_goal, pos_count, pos_goal))
        # randomly choose neg_goal negatives and pos_goal positives:
        selected_indices = list(np.random.choice(negatives, size=(min(neg_count,neg_goal)), replace=False))
        selected_indices.extend(list(np.random.choice(positives, size=(min(pos_count, pos_goal)), replace=False)))
        # print("n examples: {0}".format(len(selected_indices)))
        n_examples = len(selected_indices)
        np.random.shuffle(selected_indices)
        test_examples = int(round(test_pct * n_examples))
        # print("test_examples: {0}".format(test_examples))
        self._test_indices  = selected_indices[0:test_examples]
        self._train_indices = selected_indices[test_examples:]