world-ecoregion/nn.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 matplotlib.pyplot as plt
import pandas as pd
import os.path

from utils import *
# from predict import predicted_map

RANDOM_SEED = 1

print(tf.__version__)

# tf.enable_eager_execution()

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

df = pd.read_pickle('data.p')

class MapHistory(keras.callbacks.Callback):
    def on_epoch_end(self, epoch, logs):
        print('EPOCH', epoch)
        predicted_map('maps/{}'.format(epoch))

class Model():
    def __init__(self, name, batch_size=16, shuffle_buffer_size=500, learning_rate=0.001, epochs=1):
        self.name = name
        self.path = "checkpoints/{}.hdf5".format(name)

        self.batch_size = batch_size
        self.shuffle_buffer_size = shuffle_buffer_size
        self.learning_rate = learning_rate 
        self.epochs = epochs

    def prepare_dataset(self, df, fn):
        self.dataset_fn = fn
        dataset_size, features, output_size, dataset = fn(df)
        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).batch(self.batch_size).repeat(),
                            self.dataset.skip(self.TRAIN_SIZE).batch(self.batch_size).repeat())

        # print(df.groupby(['biome_num']).agg({ 'biome_num': lambda x: x.count() / df.shape[0] }))

        print('dataset: size={}, train={}, test={}'.format(dataset_size, self.TRAIN_SIZE, self.TEST_SIZE))
        print('input_size={}'.format(features))

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

    def create_model(self, layers, out_activation=None):
        params = {
                'kernel_initializer': 'lecun_uniform',
                'bias_initializer': 'zeros',
                # 'kernel_regularizer': keras.regularizers.l2(l=0.01)
        }
        dropout = [keras.layers.Dropout(0.1, input_shape=[self.features])]
        # dropout = []
        self.model = keras.Sequential(dropout + [
            keras.layers.Dense(layers[0], activation=tf.nn.elu, **params)
        ] + [
            keras.layers.Dense(n, activation=tf.nn.elu, **params) for n in layers[1:]
        ] + [
            keras.layers.Dense(self.output_size, activation=out_activation, **params)
        ])

    def compile(self, loss='mse', metrics=['accuracy'], optimizer=tf.train.AdamOptimizer, load_weights=True):
        if load_weights:
            self.model.load_weights(self.path)

        optimizer = optimizer(self.learning_rate)

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

    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):
        self.model.summary()

        checkpoint = keras.callbacks.ModelCheckpoint(self.path, monitor='val_loss', verbose=1, mode='min', save_best_only=True)
        tensorboard = keras.callbacks.TensorBoard(log_dir='./logs', update_freq='epoch')
        # reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=0.0001)
        # map_callback = MapHistory()

        self.model.fit(
            self.training,
            batch_size=self.batch_size,
            epochs=self.epochs,
            steps_per_epoch=int(self.TRAIN_SIZE / self.batch_size),
            callbacks=[checkpoint, tensorboard],
            validation_data=self.test,
            validation_steps=int(self.TEST_SIZE / self.batch_size),
            verbose=1
        )

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

A = Model('a', epochs=2)
B = Model('b', learning_rate=0.0005, epochs=50)

# 24 so far
def compile_b():
    B.prepare_dataset(df, dataframe_to_dataset_biomes)
    B.create_model([12], tf.nn.softmax)
    B.compile(loss='sparse_categorical_crossentropy', load_weights=False)

def compile_a():
    A.prepare_dataset(df, dataframe_to_dataset_temp_precip)
    A.create_model([(4, tf.nn.elu)])
    # A.create_model([]) # linear model
    A.compile(metrics=['accuracy', 'mae'])

if __name__ == "__main__":
    compile_b()
    B.train()

    # for inp, out in B.test.take(1).make_one_shot_iterator():
        # print(inp, out)

    # print(np.unique(nums))
    # print(np.unique(predictions))
    # print('loss: {}, evaluation: {}'.format(*B.evaluate()))

    # compile_a()
    # A.train()