world-ecoregion/biomes/model.py

from __future__ import absolute_import, division, print_function

# TensorFlow and tf.keras
import tensorflow as tf
from tensorflow import keras

# Helper libraries
import numpy as np
import pandas as pd

from utils import *

RANDOM_SEED = 1

logger.debug('Tensorflow version: %s', tf.__version__)
logger.debug('Random Seed: %s', RANDOM_SEED)

tf.set_random_seed(RANDOM_SEED)
np.random.seed(RANDOM_SEED)

DEFAULT_BATCH_SIZE=256
DEFAULT_LAYERS = [512, 512]
DEFAULT_BUFFER_SIZE=500
DEFAULT_OUT_ACTIVATION = tf.nn.softmax
DEFAULT_LOSS = 'sparse_categorical_crossentropy'
DEFAULT_OPTIMIZER = tf.keras.optimizers.Adam(lr=0.001)
DEFAULT_METRICS = ['accuracy']

class Model():
    def __init__(self, name, epochs=1):
        self.name = name
        self.path = "checkpoints/{}.hdf5".format(name)

        self.epochs = epochs

    def prepare_dataset(self, df, fn, **kwargs):
        self.dataset_fn = fn

        self.set_dataset(*fn(df), **kwargs)

    def set_dataset(self, dataset_size, features, output_size, class_weight, dataset, shuffle_buffer_size=DEFAULT_BUFFER_SIZE, batch_size=DEFAULT_BATCH_SIZE):
        self.shuffle_buffer_size = shuffle_buffer_size

        self.class_weight = class_weight
        self.dataset = dataset.shuffle(self.shuffle_buffer_size)
        self.TRAIN_SIZE = int(dataset_size * 0.85)
        self.TEST_SIZE = dataset_size - self.TRAIN_SIZE
        (training, test) = (self.dataset.take(self.TRAIN_SIZE),
                            self.dataset.skip(self.TRAIN_SIZE))

        logger.debug('Model dataset info: size=%s, train=%s, test=%s', dataset_size, self.TRAIN_SIZE, self.TEST_SIZE)

        self.dataset_size = dataset_size
        self.features = features
        self.output_size = output_size
        self.training = training
        self.test = test

        logger.debug('Model input size: %s', self.features)
        logger.debug('Model output size: %s', self.output_size)

        self.batch_size = batch_size
        self.training_batched = self.training.batch(self.batch_size).repeat()
        self.test_batched = self.test.batch(self.batch_size).repeat()

    def create_model(self, layers=DEFAULT_LAYERS, out_activation=DEFAULT_OUT_ACTIVATION):
        params = {
                'kernel_initializer': 'lecun_uniform',
                'bias_initializer': 'zeros',
                # 'kernel_regularizer': keras.regularizers.l2(l=0.01)
                'input_shape': [self.features]
        }

        activation = tf.nn.elu

        logger.debug('Model layer parameters: %s', params)
        logger.debug('Model layer sizes: %s', layers)
        logger.debug('Model layer activation function: %s', activation)
        logger.debug('Model out activation function: %s', out_activation)


        self.model = keras.Sequential([
            keras.layers.Dense(n, activation=activation, **params) for n in layers
        ] + [
            keras.layers.Dense(self.output_size, activation=out_activation, **params)
        ])

    def compile(self, loss=DEFAULT_LOSS, metrics=DEFAULT_METRICS, optimizer=DEFAULT_OPTIMIZER):
        logger.debug('Model loss function: %s', loss)
        logger.debug('Model optimizer: %s', optimizer)
        logger.debug('Model metrics: %s', metrics)

        self.model.compile(loss=loss,
                    optimizer=optimizer,
                    metrics=metrics)

    def restore(self, path):
        logger.debug('Restoring model weights from path: %s', path)
        return self.model.load_weights(path)

    def save(self, path):
        logger.debug('Saving model weights to path: %s', path)
        self.model.save_weights(path + '/checkpoint.h5', save_format='h5')
        return path + '/checkpoint.h5'

    def evaluate(self):
        return self.model.evaluate(
            self.test,
            batch_size=self.batch_size,
            steps=int(self.dataset_size / self.batch_size),
            verbose=1
        )

    def evaluate_print(self):
        loss, accuracy = self.evaluate()
        print('Test evaluation: loss: {}, accuracy: {}'.format(loss, accuracy))

    def train(self, config):
        self.model.summary()

        # map_callback = MapHistory()

        extra_params = {}
        if self.class_weight:
            extra_params['class_weight'] = self.class_weight

        out = self.model.fit(
            self.training_batched,
            batch_size=self.batch_size,
            epochs=self.epochs,
            steps_per_epoch=int(self.TRAIN_SIZE / self.batch_size),
            validation_data=self.test_batched,
            validation_steps=int(self.TEST_SIZE / self.batch_size),
            verbose=1,
            **extra_params
        )

        return out

    def predict_class(self, a):
        return np.argmax(self.model.predict(a), axis=1)

    def predict(self, a):
        return self.model.predict(a)

    def prepare_for_use(self, df=None, batch_size=DEFAULT_BUFFER_SIZE, layers=DEFAULT_LAYERS, out_activation=DEFAULT_OUT_ACTIVATION, loss=DEFAULT_LOSS, optimizer=DEFAULT_OPTIMIZER, dataset_fn=dataframe_to_dataset_biomes, metrics=DEFAULT_METRICS):
        if df is None:
            df = pd.read_pickle('data.p')
        self.prepare_dataset(df, dataset_fn, batch_size=batch_size)
        self.create_model(layers=layers, out_activation=out_activation)
        self.compile(loss=loss, optimizer=optimizer, metrics=metrics)