segmentation-final.py

from __future__ import division
import matplotlib.pyplot as plt
import os, glob
from sklearn.model_selection import train_test_split
import matplotlib.image as mpimg
from PIL import Image
import numpy as np
import json
from config import credentials

import tensorflow as tf
import tensorflow.contrib as tfcontrib
from tensorflow.python.keras import layers
from tensorflow.python.keras import losses
from tensorflow.python.keras import models
from tensorflow.python.keras import utils
from tensorflow.python.keras import backend as K

AWS_ACCESS_KEY_ID=credentials.key_id # Credentials only needed if connecting to a private endpoint
AWS_SECRET_ACCESS_KEY=credentials.secret_key
AWS_REGION='us-west-2'
S3_ENDPOINT='s3-us-west-2.amazonaws.com'       # Region for the S3 bucket, this is not always needed. Default is us-east-1.
S3_USE_HTTPS=1                          # Whether or not to use HTTPS. Disable with 0.
S3_VERIFY_SSL=1                         # If HTTPS is used, controls if SSL should be enabled. Disable with 0.

val_filename = "s3://street-tfrecord/val_ds.tfrecord"
train_filename = "s3://street-tfrecord/train_ds.tfrecord"

num_classes = 8
img_shape = (480, 480, 3)
batch_size = 1
epochs = 100

def _parse_image_function(example_proto):

    image_feature_description = {
        'image': tf.FixedLenFeature([], tf.string),
        'mask': tf.FixedLenFeature([], tf.string),
    }

    tfrecord_features =  tf.parse_single_example(example_proto, image_feature_description)

    img = tf.image.decode_jpeg(tfrecord_features['image'], channels=3)
    label_img = tf.image.decode_png(tfrecord_features['mask'], channels=3)

    label_img = label_img[:, :, 0]
    label_img = tf.expand_dims(label_img, axis=-1)
    label_img = tf.one_hot(label_img,8)
    label_img = tf.squeeze(label_img)
    label_img.set_shape((480,640,num_classes))

    img = tf.image.resize_images(img, [img_shape[0], img_shape[1]])
    label_img = tf.image.resize_images(label_img, [img_shape[0], img_shape[1]])

    img = tf.to_float(img) * 1/255.

    return img, label_img

def shift_img(output_img, label_img, width_shift_range, height_shift_range):
    if width_shift_range or height_shift_range:
        if width_shift_range:
            width_shift_range = tf.random_uniform([], -width_shift_range * img_shape[1],width_shift_range * img_shape[1])
        if height_shift_range:
            height_shift_range = tf.random_uniform([],-height_shift_range * img_shape[0],height_shift_range * img_shape[0])
      # Translate both
        output_img = tfcontrib.image.translate(output_img,
                                             [width_shift_range, height_shift_range])
        label_img = tfcontrib.image.translate(label_img,
                                             [width_shift_range, height_shift_range])
        return output_img, label_img

def flip_img(horizontal_flip, tr_img, label_img):
    if horizontal_flip:
        flip_prob = tf.random_uniform([], 0.0, 1.0)
        tr_img, label_img = tf.cond(tf.less(flip_prob, 0.5),
                                lambda: (tf.image.flip_left_right(tr_img), tf.image.flip_left_right(label_img)),
                                lambda: (tr_img, label_img))
        return tr_img, label_img


def train_preprocess(img,label_img):

    horizontal_flip = True
    width_shift_range = 0.1
    height_shift_range = 0.1
    hue_delta = 0.1

    img = tf.image.random_hue(img, hue_delta)
    img, label_img = flip_img(horizontal_flip, img, label_img)
    img, label_img = shift_img(img, label_img, width_shift_range, height_shift_range)

    return img, label_img

def get_baseline_dataset(tfrecord_filename, num_records, preprocess=True):

    dataset= tf.data.TFRecordDataset(tfrecord_filename)
    dataset = dataset.shuffle(num_records)
    dataset = dataset.repeat()
    dataset = dataset.map(_parse_image_function,num_parallel_calls=4)
    if preprocess:
        dataset = dataset.map(train_preprocess,num_parallel_calls=4)
    dataset = dataset.batch(batch_size)
    dataset = dataset.prefetch(1)

    return dataset

num_records_train = sum(1 for _ in tf.python_io.tf_record_iterator(train_filename))
num_records_val = sum(1 for _ in tf.python_io.tf_record_iterator(val_filename))

train_ds = get_baseline_dataset(train_filename,num_records_train, preprocess=True)
val_ds = get_baseline_dataset(val_filename,num_records_val, preprocess=False)

def conv_block(input_tensor, num_filters):
  encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(input_tensor)
  encoder = layers.BatchNormalization()(encoder)
  encoder = layers.Activation('relu')(encoder)
  encoder = layers.Conv2D(num_filters, (3, 3), padding='same')(encoder)
  encoder = layers.BatchNormalization()(encoder)
  encoder = layers.Activation('relu')(encoder)
  return encoder

def encoder_block(input_tensor, num_filters):
  encoder = conv_block(input_tensor, num_filters)
  encoder_pool = layers.MaxPooling2D((2, 2), strides=(2, 2))(encoder)

  return encoder_pool, encoder

def decoder_block(input_tensor, concat_tensor, num_filters):
  decoder = layers.Conv2DTranspose(num_filters, (2, 2), strides=(2, 2), padding='same')(input_tensor)
  decoder = layers.concatenate([concat_tensor, decoder], axis=-1)
  decoder = layers.BatchNormalization()(decoder)
  decoder = layers.Activation('relu')(decoder)
  decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)
  decoder = layers.BatchNormalization()(decoder)
  decoder = layers.Activation('relu')(decoder)
  decoder = layers.Conv2D(num_filters, (3, 3), padding='same')(decoder)
  decoder = layers.BatchNormalization()(decoder)
  decoder = layers.Activation('relu')(decoder)
  return decoder

inputs = layers.Input(shape=img_shape)
# 256

encoder0_pool, encoder0 = encoder_block(inputs, 32)
# 128

encoder1_pool, encoder1 = encoder_block(encoder0_pool, 64)
# 64

encoder2_pool, encoder2 = encoder_block(encoder1_pool, 128)
# 32

encoder3_pool, encoder3 = encoder_block(encoder2_pool, 256)
# 16

encoder4_pool, encoder4 = encoder_block(encoder3_pool, 512)
# 8

center = conv_block(encoder4_pool, 1024)
# center

decoder4 = decoder_block(center, encoder4, 512)
# 16

decoder3 = decoder_block(decoder4, encoder3, 256)
# 32

decoder2 = decoder_block(decoder3, encoder2, 128)
# 64

decoder1 = decoder_block(decoder2, encoder1, 64)
# 128

decoder0 = decoder_block(decoder1, encoder0, 32)
# 256

outputs = layers.Conv2D(num_classes, (1, 1), activation='softmax')(decoder0)

model = models.Model(inputs=[inputs], outputs=[outputs])

class MeanIoU(object):
    def __init__(self, num_classes):
        self.num_classes = num_classes

    def mean_iou(self, y_true, y_pred):
        return tf.py_func(self.np_mean_iou, [y_true, y_pred], tf.float32)

    def np_mean_iou(self, y_true, y_pred):
        target = np.argmax(y_true, axis=-1).ravel()
        predicted = np.argmax(y_pred, axis=-1).ravel()

        x = predicted + self.num_classes * target
        bincount_2d = np.bincount(x.astype(np.int32), minlength=self.num_classes**2)
        assert bincount_2d.size == self.num_classes**2
        conf = bincount_2d.reshape((self.num_classes, self.num_classes))

        true_positive = np.diag(conf)
        false_positive = np.sum(conf, 0) - true_positive
        false_negative = np.sum(conf, 1) - true_positive

        with np.errstate(divide='ignore', invalid='ignore'):
            iou = true_positive / (true_positive + false_positive + false_negative)
        iou[np.isnan(iou)] = 0

        return np.mean(iou).astype(np.float32)

miou_metric = MeanIoU(num_classes)

model.compile(optimizer='adam', loss=losses.categorical_crossentropy, metrics=[miou_metric.mean_iou])

save_model_path = 'weights.hdf5'
cp = tf.keras.callbacks.ModelCheckpoint(filepath=save_model_path, monitor='loss', save_best_only=True, verbose=1,period=10)

history = model.fit(train_ds,
                   steps_per_epoch=int(np.ceil(num_records_train / float(batch_size))),
                   epochs=epochs,
                   validation_data=val_ds,
                   validation_steps=int(np.ceil(num_records_val / float(batch_size))),
                   callbacks=[cp])

plt.style.use('ggplot')
f, (ax1, ax2) = plt.subplots(1, 2)
plt.tight_layout(pad=1.5,w_pad=0.2)

# Plot training & validation loss values
ax1.plot(history.history['loss'])
ax1.plot(history.history['val_loss'])
ax1.set_title('Model loss')
ax1.set_ylabel('Loss')
ax1.set_ylim(top=2)
ax1.set_xlabel('Epoch')
ax1.legend(['Train', 'Test'], loc='upper left')

# Plot training & validation mIOU
ax2.plot(history.history['mean_iou'])
ax2.plot(history.history['val_mean_iou'])
ax2.set_title('Model mIOU')
ax2.set_ylabel('mIOU')
ax2.set_xlabel('Epoch')
ax2.legend(['Train', 'Test'], loc='upper left')

f.set_size_inches(13, 5)
f.savefig('model_metrics.png', dpi=250)

with open('model_history.json', 'w') as f:
    json.dump(history.history, f)