Skip to content

Block train

block_train(config) ยค

Internal code block for AutoML training.

This function is called by the Command-Line-Interface (CLI) of AUCMEDI.

Parameters:

Name Type Description Default
config dict

Configuration dictionary containing all required parameters for performing an AutoML training.

required

The following attributes are stored in the config dictionary:

Attributes:

Name Type Description
path_imagedir str

Path to the directory containing the images.

path_modeldir str

Path to the output directory in which fitted models and metadata are stored.

path_gt str

Path to the index/class annotation file if required. (only for 'csv' interface).

analysis str

Analysis mode for the AutoML training. Options: ["minimal", "standard", "advanced"].

ohe bool

Boolean option whether annotation data is sparse categorical or one-hot encoded.

three_dim bool

Boolean, whether data is 2D or 3D.

shape_3D tuple of int

Desired input shape of 3D volume for architecture (will be cropped).

epochs int

Number of epochs. A single epoch is defined as one iteration through the complete data set.

batch_size int

Number of samples inside a single batch.

workers int

Number of workers/threads which preprocess batches during runtime.

metalearner str

Key for Metalearner or Aggregate function.

architecture str or list of str

Key (str) of a neural network model Architecture class instance.

Source code in aucmedi/automl/block_train.py
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
def block_train(config):
    """ Internal code block for AutoML training.

    This function is called by the Command-Line-Interface (CLI) of AUCMEDI.

    Args:
        config (dict):                      Configuration dictionary containing all required
                                            parameters for performing an AutoML training.

    The following attributes are stored in the `config` dictionary:

    Attributes:
        path_imagedir (str):                Path to the directory containing the images.
        path_modeldir (str):                Path to the output directory in which fitted models and metadata are stored.
        path_gt (str):                      Path to the index/class annotation file if required. (only for 'csv' interface).
        analysis (str):                     Analysis mode for the AutoML training. Options: `["minimal", "standard", "advanced"]`.
        ohe (bool):                         Boolean option whether annotation data is sparse categorical or one-hot encoded.
        three_dim (bool):                   Boolean, whether data is 2D or 3D.
        shape_3D (tuple of int):            Desired input shape of 3D volume for architecture (will be cropped).
        epochs (int):                       Number of epochs. A single epoch is defined as one iteration through
                                            the complete data set.
        batch_size (int):                   Number of samples inside a single batch.
        workers (int):                      Number of workers/threads which preprocess batches during runtime.
        metalearner (str):                  Key for Metalearner or Aggregate function.
        architecture (str or list of str):  Key (str) of a neural network model Architecture class instance.
    """
    # Obtain interface
    if config["path_gt"] is None : config["interface"] = "directory"
    else : config["interface"] = "csv"
    # Peak into the dataset via the input interface
    ds = input_interface(config["interface"],
                         config["path_imagedir"],
                         path_data=config["path_gt"],
                         training=True,
                         ohe=config["ohe"],
                         image_format=None)
    (index_list, class_ohe, class_n, class_names, image_format) = ds

    # Create output directory
    if not os.path.exists(config["path_modeldir"]):
        os.mkdir(config["path_modeldir"])

    # Identify task (multi-class vs multi-label)
    if np.sum(class_ohe) > class_ohe.shape[0] : config["multi_label"] = True
    else : config["multi_label"] = False

    # Sanity check on multi-label metalearner
    multilabel_metalearner_supported = ["mlp", "k_neighbors", "random_forest",
                                        "weighted_mean", "best_model",
                                        "decision_tree", "mean", "median"]
    if config["multi_label"] and config["analysis"] == "advanced" and \
       config["metalearner"] not in multilabel_metalearner_supported:
        raise ValueError("Non-compatible metalearner selected for multi-label"\
                         + " classification. Supported metalearner:",
                          multilabel_metalearner_supported)

    # Store meta information
    config["class_names"] = class_names
    path_meta = os.path.join(config["path_modeldir"], "meta.training.json")
    with open(path_meta, "w") as json_io:
        json.dump(config, json_io)

    # Define Callbacks
    callbacks = []
    if config["analysis"] == "standard":
        cb_loss = ModelCheckpoint(os.path.join(config["path_modeldir"],
                                               "model.best_loss.keras"),
                                  monitor="val_loss", verbose=1,
                                  save_best_only=True)
        callbacks.append(cb_loss)
    if config["analysis"] in ["minimal", "standard"]:
        cb_cl = CSVLogger(os.path.join(config["path_modeldir"],
                                       "logs.training.csv"),
                          separator=',', append=True)
        callbacks.append(cb_cl)
    if config["analysis"] != "minimal":
        cb_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=5,
                              verbose=1, mode='min', min_lr=1e-7)
        cb_es = EarlyStopping(monitor='val_loss', patience=12, verbose=1)
        callbacks.extend([cb_lr, cb_es])

    # Initialize loss function for multi-class
    if not config["multi_label"]:
        # Compute class weights
        class_weights, _ = compute_class_weights(ohe_array=class_ohe)
        # Initialize focal loss
        loss = categorical_focal_loss(class_weights)
    # Initialize loss function for multi-label
    else:
        # Compute class weights
        class_weights = compute_multilabel_weights(ohe_array=class_ohe)
        # Initialize focal loss
        loss = multilabel_focal_loss(class_weights)

    # Define neural network parameters
    nn_paras = {"n_labels": class_n,
                "channels": 3,
                "loss": loss,
                "metrics": [AUC(100)],
                "pretrained_weights": True,
    }
    # Select input shape for 3D
    if config["three_dim"] : nn_paras["input_shape"] = config["shape_3D"]
    # Select task type
    if config["multi_label"] : nn_paras["activation_output"] = "sigmoid"
    else : nn_paras["activation_output"] = "softmax"

    # Initialize Augmentation for 2D image data
    if not config["three_dim"]:
        data_aug = ImageAugmentation(flip=True, rotate=True, scale=False,
                                     brightness=True, contrast=True,
                                     saturation=False, hue=False, crop=False,
                                     grid_distortion=False, compression=False,
                                     gamma=True, gaussian_noise=False,
                                     gaussian_blur=False, downscaling=False,
                                     elastic_transform=True)
    # Initialize Augmentation for 3D volume data
    elif config["three_dim"]:
        data_aug = BatchgeneratorsAugmentation(image_shape=config["shape_3D"],
                        mirror=True, rotate=True, scale=True,
                        elastic_transform=True, gaussian_noise=False,
                        brightness=False, contrast=False, gamma=True)
    else : data_aug = None

    # Subfunctions
    sf_list = []
    if config["three_dim"]:
        sf_norm = Standardize(mode="grayscale")
        sf_pad = Padding(mode="constant", shape=config["shape_3D"])
        sf_crop = Crop(shape=config["shape_3D"], mode="random")
        sf_chromer = Chromer(target="rgb")
        sf_list.extend([sf_norm, sf_pad, sf_crop, sf_chromer])

    # Define parameters for DataGenerator
    paras_datagen = {
        "path_imagedir": config["path_imagedir"],
        "batch_size": config["batch_size"],
        "img_aug": data_aug,
        "subfunctions": sf_list,
        "prepare_images": False,
        "sample_weights": None,
        "seed": None,
        "image_format": image_format,
        "workers": config["workers"],
    }
    if not config["three_dim"] : paras_datagen["loader"] = image_loader
    else : paras_datagen["loader"] = sitk_loader

    # Gather training parameters
    paras_train = {
        "epochs": config["epochs"],
        "iterations": None,
        "callbacks": callbacks,
        "class_weights": None,
        "transfer_learning": True,
    }

    # Apply MIC pipelines
    if config["analysis"] == "minimal":
        # Setup neural network
        if not config["three_dim"] : arch_dim = "2D." + config["architecture"]
        else : arch_dim = "3D." + config["architecture"]
        model = NeuralNetwork(architecture=arch_dim, **nn_paras)

        # Build DataGenerator
        train_gen = DataGenerator(samples=index_list,
                                  labels=class_ohe,
                                  shuffle=True,
                                  resize=model.meta_input,
                                  standardize_mode=model.meta_standardize,
                                  **paras_datagen)

        # Start model training
        hist = model.train(training_generator=train_gen, **paras_train)
        # Store model
        path_model = os.path.join(config["path_modeldir"], "model.last.keras")
        model.dump(path_model)
    elif config["analysis"] == "standard":
        # Setup neural network
        if not config["three_dim"] : arch_dim = "2D." + config["architecture"]
        else : arch_dim = "3D." + config["architecture"]
        model = NeuralNetwork(architecture=arch_dim, **nn_paras)

        # Apply percentage split sampling
        ps_sampling = sampling_split(index_list, class_ohe,
                                     sampling=[0.85, 0.15],
                                     stratified=True, iterative=True,
                                     seed=0)

        # Build DataGenerator
        train_gen = DataGenerator(samples=ps_sampling[0][0],
                                  labels=ps_sampling[0][1],
                                  shuffle=True,
                                  resize=model.meta_input,
                                  standardize_mode=model.meta_standardize,
                                  **paras_datagen)
        val_gen = DataGenerator(samples=ps_sampling[1][0],
                                labels=ps_sampling[1][1],
                                shuffle=False,
                                resize=model.meta_input,
                                standardize_mode=model.meta_standardize,
                                **paras_datagen)

        # Start model training
        hist = model.train(training_generator=train_gen,
                           validation_generator=val_gen,
                           **paras_train)
        # Store model
        path_model = os.path.join(config["path_modeldir"], "model.last.keras")
        model.dump(path_model)
    else:
        # Sanity check of architecutre config
        if not isinstance(config["architecture"], list):
            raise ValueError("key 'architecture' in config has to be a list " \
                             + "if 'advanced' was selected as analysis.")
        # Build multi-model list
        model_list = []
        for arch in config["architecture"]:
            if not config["three_dim"] : arch_dim = "2D." + arch
            else : arch_dim = "3D." + arch
            model_part = NeuralNetwork(architecture=arch_dim, **nn_paras)
            model_list.append(model_part)
        el = Composite(model_list, metalearner=config["metalearner"],
                       k_fold=len(config["architecture"]))

        # Build DataGenerator
        train_gen = DataGenerator(samples=index_list,
                                  labels=class_ohe,
                                  shuffle=True,
                                  resize=None,
                                  standardize_mode=None,
                                  **paras_datagen)
        # Start model training
        hist = el.train(training_generator=train_gen, **paras_train)
        # Store model directory
        el.dump(config["path_modeldir"])

    # Plot fitting history
    evaluate_fitting(train_history=hist, out_path=config["path_modeldir"])