train.py

import os
# os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
import wandb
from wandb.keras import WandbCallback
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow import keras
from tensorflow.keras.layers import Dense, Dropout, LSTM, LeakyReLU, Flatten, BatchNormalization, SimpleRNN, GRU, BatchNormalization, TimeDistributed, Lambda, ReLU, PReLU
import numpy as np
import random
import json
import matplotlib.pyplot as plt
import pickle
from sklearn.model_selection import train_test_split
from tensorflow.keras.models import load_model
import seaborn as sns
# from utils.auto_df_support import DataProcessor
from utils.BASEDIR import BASEDIR
from utils.tokenizer import split_list, tokenize
import ast
import matplotlib.gridspec as gridspec
# from keras.callbacks.tensorboard_v1 import TensorBoard

# config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 0.8
# set_session(tf.Session(config=config))
# sns.distplot(data_here)

# gpus = tf.config.experimental.list_physical_devices('GPU')
# if gpus:
#   tf.config.experimental.set_virtual_device_configuration(
#       gpus[0],
#       [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=2048)])
#   logical_gpus = tf.config.experimental.list_logical_devices('GPU')

if "CUDA_VISIBLE_DEVICES" not in os.environ:
  physical_devices = tf.config.list_physical_devices('GPU')
  tf.config.experimental.set_memory_growth(physical_devices[0], True)


# support = AutoDFSupport()
# support.init()
os.chdir(BASEDIR)

fig, ax = plt.subplots(5, 4)
ax = ax.flatten()


def show_pred_new(epoch=0, sample_idx=None, figure_idx=0):
  if sample_idx is None:
    sample_idx = random.randrange(len(auto.x_test))
  x = auto.x_test[sample_idx]
  y = auto.y_test[sample_idx]
  pred = auto.model.predict_on_batch(np.array([x]))[0]

  ax[figure_idx].cla()
  ax[figure_idx].bar(range(3), y, label='ground')
  ax[figure_idx].bar(range(3), pred, label='pred')
  # ax[figure_idx].legend()
  plt.show()
  plt.pause(0.01)
  # plt.savefig('models/model_imgs/{}'.format(epoch))


class ShowPredictions(tf.keras.callbacks.Callback):
  def __init__(self):
    super().__init__()
    self.every = 20
    self.sample_idxs = [random.randrange(len(auto.x_test)) for _ in range(len(ax))]

  def on_epoch_end(self, epoch, logs=None):
    if not (epoch + self.every) % self.every:
      for idx, sample_idx in enumerate(self.sample_idxs):
        show_pred_new(epoch, sample_idx, idx)


class AutoDynamicFollow:
  def __init__(self, cfg):
    self.test_size = 0.15
    self.config = cfg

  def start(self):
    self.load_data()
    self.train()

  def load_data(self):
    print("Loading data...", flush=True)
    self.x_train = np.load('model_data/x_train.npy', allow_pickle=True)
    self.y_train = np.load('model_data/y_train.npy', allow_pickle=True)
    # self.x_test = np.load('model_data/x_test.npy')
    # self.y_test = np.load('model_data/y_test.npy')
    with open("model_data/scales", "rb") as f:
      self.scales = json.load(f)

    samples = 'all'
    if samples != 'all':
      self.x_train = np.array(self.x_train[:samples])
      self.y_train = np.array(self.y_train[:samples])
    flatten = False  # already flattened in data processor
    if flatten:
      self.x_train = np.array([i.flatten() for i in self.x_train])

    self.x_train, self.x_test, self.y_train, self.y_test = train_test_split(self.x_train, self.y_train, test_size=self.test_size)
    print(self.x_train.shape)
    print(self.y_train.shape)

    print('{} traffic samples'.format(len([i for i in self.y_train if i[0] == 1])))
    print('{} relaxed samples'.format(len([i for i in self.y_train if i[1] == 1])))
    print('{} roadtrip samples'.format(len([i for i in self.y_train if i[2] == 1])))

  def train(self):
    # opt = keras.optimizers.Adadelta(lr=2) #lr=.000375)
    # opt = keras.optimizers.RMSprop(lr=0.001)#, decay=1e-5)
    # opt = keras.optimizers.Adagrad(lr=0.00025)
    # opt = keras.optimizers.SGD(lr=0.1, momentum=0.3)
    # opt = keras.optimizers.Adagrad()
    # opt = 'rmsprop'
    # opt = keras.optimizers.Adadelta()
    # opt = 'adam'
    opt = keras.optimizers.Adam(lr=self.config.learning_rate, amsgrad=True)

    a_function = self.config.activation

    self.model = Sequential()
    # model.add(Dropout(0.2))
    # model.add(GRU(128, return_sequences=True, input_shape=x_train.shape[1:]))
    # model.add(GRU(64, return_sequences=True))
    # model.add(GRU(64, return_sequences=True))
    # model.add(GRU(64, return_sequences=True))
    # model.add(GRU(y_train.shape[1], return_sequences=False))
    # model.add(Lambda(lambda x: x[:,0,:,:], output_shape=(1, 50, 1) + x_train.shape[2:]))

    # denses = [128, 64, 32]

    self.model.add(Dense(self.config.dense_one, activation=a_function, input_shape=self.x_train.shape[1:]))
    self.model.add(Dropout(config.dropout_one))

    self.model.add(Dense(self.config.dense_two, activation=a_function))
    self.model.add(Dropout(config.dropout_two))

    self.model.add(Dense(self.y_train.shape[1], activation='softmax'))

    self.model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['acc'])
    # name = ', '.join(['{}'.format(n) for n in denses])
    # print(name)
    w_and_b = WandbCallback()
    callbacks = [w_and_b, ShowPredictions()]
    self.model.fit(self.x_train, self.y_train,
                   shuffle=True,
                   batch_size=self.config.batch_size,
                   epochs=self.config.epochs,
                   validation_data=(self.x_test, self.y_test),
                   # sample_weight=np.full((len(y_train)), 100),
                   callbacks=callbacks)


hyperparameter_defaults = dict(
  dropout_one=0.175,
  dropout_two=0.26,

  dense_one=128,
  dense_two=96,

  activation='relu',
  batch_size=96,
  learning_rate=0.00044,
  epochs=140,
)

wandb.init(project="auto-df", config=hyperparameter_defaults)

config = wandb.config


auto = AutoDynamicFollow(config)
auto.start()

# preds = model.predict(x_test).reshape(1, -1)
# diffs = [abs(pred - ground) for pred, ground in zip(preds[0], y_test[0])]
#
# print("Test accuracy: {}%".format(round(np.interp(sum(diffs) / len(diffs), [0, 1], [1, 0]) * 100, 4)))
#
# for i in range(20):
#   c = random.randint(0, len(x_test))
#   print('Ground truth: {}'.format(support.unnorm(y_test[c][0], 'eps_torque')))
#   print('Prediction: {}'.format(support.unnorm(model.predict(np.array([x_test[c]]))[0][0], 'eps_torque')))
#   print()


def save_model(name='model'):
  auto.model.save('models/{}.h5'.format(name))