Notebook

ground-segmentation¶

Learning a deep neural network for ground segmentation.

Preprocessing¶

Get dataset¶

In [1]:

from utils.get_dataset import get_cmu_corridor_dataset

get_cmu_corridor_dataset(dataset_path='./dataset')

Compressing dataset: 100%|██████████| 2906/2906 [00:01<00:00, 1495.70it/s]

Load dataset¶

In [1]:

from utils.load_dataset import load_cmu_corridor_dataset

X_train, X_test, y_train, y_test = load_cmu_corridor_dataset(dataset_path='./dataset/cmu_corridor_dataset',
                                                             train_test_split=True,
                                                             image_size=(240, 320),
                                                             gray=True,
                                                             verbose=True)
print('X_train: %s, y_train: %s' % (X_train.shape, y_train.shape))
print('X_test: %s, y_test: %s' % (X_test.shape, y_test.shape))

Loading images: 100%|██████████| 967/967 [00:01<00:00, 551.44it/s]
Loading labels: 100%|██████████| 967/967 [00:00<00:00, 1785.36it/s]

X_train: (725, 240, 320, 1), y_train: (725, 240, 320, 1)
X_test: (242, 240, 320, 1), y_test: (242, 240, 320, 1)

Insight dataset¶

In [27]:

from matplotlib import pyplot as plt
plt.rcParams['figure.figsize'] = [5, 8]

def plot_image(ax, img, title):
    ax.axis('off')
    ax.set_title(title)
    if img.shape[-1] == 1:
        img = img[..., 0]
        ax.imshow(img, cmap='gray')
    else:
        ax.imshow(img)
    
def plot_label(ax, label, title):
    ax.axis('off')
    ax.set_title(title)
    label = label[..., 0]
    ax.imshow(label, cmap='gray')
    
NUM_PLOTTING = 4

fig = plt.figure("Insight dataset")

images, labels = X_train, y_train

for index in range(NUM_PLOTTING):
    image, label = images[index], labels[index]
    image_title = '' if index != 0 else 'Raw image'
    label_title = '' if index != 0 else 'Ground segmentation'

    ax = fig.add_subplot(NUM_PLOTTING, 2, index*2 + 1)
    plot_image(ax, image, title=image_title)
    ax = fig.add_subplot(NUM_PLOTTING, 2, index*2 + 2)
    plot_label(ax, label, title=label_title)

plt.show()

Define Model¶

unet¶

In [2]:

img_height, img_width, img_channel = X_train[0].shape

In [3]:

from tensorflow.keras import backend as K
import tensorflow as tf
from tensorflow.keras import metrics
import numpy as np

# Compatible with tensorflow backend
def focal_loss(gamma=2., alpha=.75):
	def focal_loss_fixed(y_true, y_pred):
		pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
		pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
		return -K.mean(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1)) - K.mean((1 - alpha) * K.pow(pt_0, gamma) * K.log(1. - pt_0))
	return focal_loss_fixed

def mean_IoU(target, pred):
    m = metrics.MeanIoU(num_classes=2)
    return m(target, pred > 0.5)

In [4]:

from tensorflow.keras import Input
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import metrics
from utils.unet import get_unet

input_img = Input((img_height, img_width, img_channel), name='img')
model = get_unet(input_img, n_filters=16, dropout=0.05, batchnorm=True)

model.compile(optimizer=Adam(), loss="binary_crossentropy",
              metrics=["accuracy", mean_IoU])
# model.compile(optimizer=Adam(), loss=focal_loss(gamma=2, alpha=.75),
#               metrics=["accuracy", mean_IoU])

In [5]:

model.summary()

Model: "unet"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
img (InputLayer)                [(None, 240, 320, 1) 0                                            
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 240, 320, 16) 160         img[0][0]                        
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 240, 320, 16) 64          conv2d_1[0][0]                   
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 240, 320, 16) 0           batch_normalization_1[0][0]      
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D)    (None, 120, 160, 16) 0           activation_1[0][0]               
__________________________________________________________________________________________________
dropout (Dropout)               (None, 120, 160, 16) 0           max_pooling2d[0][0]              
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 120, 160, 32) 4640        dropout[0][0]                    
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 120, 160, 32) 128         conv2d_3[0][0]                   
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 120, 160, 32) 0           batch_normalization_3[0][0]      
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D)  (None, 60, 80, 32)   0           activation_3[0][0]               
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 60, 80, 32)   0           max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 60, 80, 64)   18496       dropout_1[0][0]                  
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 60, 80, 64)   256         conv2d_5[0][0]                   
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 60, 80, 64)   0           batch_normalization_5[0][0]      
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D)  (None, 30, 40, 64)   0           activation_5[0][0]               
__________________________________________________________________________________________________
dropout_2 (Dropout)             (None, 30, 40, 64)   0           max_pooling2d_2[0][0]            
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 30, 40, 128)  73856       dropout_2[0][0]                  
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 30, 40, 128)  512         conv2d_7[0][0]                   
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 30, 40, 128)  0           batch_normalization_7[0][0]      
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D)  (None, 15, 20, 128)  0           activation_7[0][0]               
__________________________________________________________________________________________________
dropout_3 (Dropout)             (None, 15, 20, 128)  0           max_pooling2d_3[0][0]            
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 15, 20, 256)  295168      dropout_3[0][0]                  
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 15, 20, 256)  1024        conv2d_9[0][0]                   
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 15, 20, 256)  0           batch_normalization_9[0][0]      
__________________________________________________________________________________________________
conv2d_transpose (Conv2DTranspo (None, 30, 40, 128)  295040      activation_9[0][0]               
__________________________________________________________________________________________________
concatenate (Concatenate)       (None, 30, 40, 256)  0           conv2d_transpose[0][0]           
                                                                 activation_7[0][0]               
__________________________________________________________________________________________________
dropout_4 (Dropout)             (None, 30, 40, 256)  0           concatenate[0][0]                
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 30, 40, 128)  295040      dropout_4[0][0]                  
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 30, 40, 128)  512         conv2d_11[0][0]                  
__________________________________________________________________________________________________
activation_11 (Activation)      (None, 30, 40, 128)  0           batch_normalization_11[0][0]     
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 60, 80, 64)   73792       activation_11[0][0]              
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 60, 80, 128)  0           conv2d_transpose_1[0][0]         
                                                                 activation_5[0][0]               
__________________________________________________________________________________________________
dropout_5 (Dropout)             (None, 60, 80, 128)  0           concatenate_1[0][0]              
__________________________________________________________________________________________________
conv2d_13 (Conv2D)              (None, 60, 80, 64)   73792       dropout_5[0][0]                  
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 60, 80, 64)   256         conv2d_13[0][0]                  
__________________________________________________________________________________________________
activation_13 (Activation)      (None, 60, 80, 64)   0           batch_normalization_13[0][0]     
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 120, 160, 32) 18464       activation_13[0][0]              
__________________________________________________________________________________________________
concatenate_2 (Concatenate)     (None, 120, 160, 64) 0           conv2d_transpose_2[0][0]         
                                                                 activation_3[0][0]               
__________________________________________________________________________________________________
dropout_6 (Dropout)             (None, 120, 160, 64) 0           concatenate_2[0][0]              
__________________________________________________________________________________________________
conv2d_15 (Conv2D)              (None, 120, 160, 32) 18464       dropout_6[0][0]                  
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 120, 160, 32) 128         conv2d_15[0][0]                  
__________________________________________________________________________________________________
activation_15 (Activation)      (None, 120, 160, 32) 0           batch_normalization_15[0][0]     
__________________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTrans (None, 240, 320, 16) 4624        activation_15[0][0]              
__________________________________________________________________________________________________
concatenate_3 (Concatenate)     (None, 240, 320, 32) 0           conv2d_transpose_3[0][0]         
                                                                 activation_1[0][0]               
__________________________________________________________________________________________________
dropout_7 (Dropout)             (None, 240, 320, 32) 0           concatenate_3[0][0]              
__________________________________________________________________________________________________
conv2d_17 (Conv2D)              (None, 240, 320, 16) 4624        dropout_7[0][0]                  
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 240, 320, 16) 64          conv2d_17[0][0]                  
__________________________________________________________________________________________________
activation_17 (Activation)      (None, 240, 320, 16) 0           batch_normalization_17[0][0]     
__________________________________________________________________________________________________
conv2d_18 (Conv2D)              (None, 240, 320, 1)  17          activation_17[0][0]              
==================================================================================================
Total params: 1,179,121
Trainable params: 1,177,649
Non-trainable params: 1,472
__________________________________________________________________________________________________

Training¶

In [8]:

from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, TensorBoard

callbacks = [
    EarlyStopping(patience=10, verbose=1),
    ReduceLROnPlateau(factor=0.1, patience=5, min_lr=0.00001, verbose=1),
    ModelCheckpoint('models/model-gs-unet.h5', verbose=1, save_best_only=True, save_weights_only=False),
    TensorBoard(log_dir='.logs')
]

In [9]:

results = model.fit(X_train, y_train, batch_size=16, epochs=50, callbacks=callbacks,\
                    validation_data=(X_test, y_test))

WARNING: Logging before flag parsing goes to stderr.
W0718 12:10:35.686528 140018677303104 deprecation.py:323] From /home/kerry/anaconda3/lib/python3.7/site-packages/tensorflow/python/ops/math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

Train on 725 samples, validate on 242 samples
Epoch 1/50
720/725 [============================>.] - ETA: 0s - loss: 0.5008 - accuracy: 0.7831 - mean_IoU: 0.4471
Epoch 00001: val_loss improved from inf to 0.67455, saving model to models/model-gs-unet.h5
725/725 [==============================] - 14s 20ms/sample - loss: 0.5000 - accuracy: 0.7835 - mean_IoU: 0.4498 - val_loss: 0.6746 - val_accuracy: 0.8255 - val_mean_IoU: 0.5803
Epoch 2/50
720/725 [============================>.] - ETA: 0s - loss: 0.3085 - accuracy: 0.9031 - mean_IoU: 0.6228
Epoch 00002: val_loss improved from 0.67455 to 0.43366, saving model to models/model-gs-unet.h5
725/725 [==============================] - 11s 15ms/sample - loss: 0.3086 - accuracy: 0.9030 - mean_IoU: 0.6234 - val_loss: 0.4337 - val_accuracy: 0.8157 - val_mean_IoU: 0.6415
Epoch 3/50
720/725 [============================>.] - ETA: 0s - loss: 0.2576 - accuracy: 0.9147 - mean_IoU: 0.6545
Epoch 00003: val_loss did not improve from 0.43366
725/725 [==============================] - 10s 14ms/sample - loss: 0.2580 - accuracy: 0.9144 - mean_IoU: 0.6548 - val_loss: 0.8242 - val_accuracy: 0.6535 - val_mean_IoU: 0.6575
Epoch 4/50
720/725 [============================>.] - ETA: 0s - loss: 0.2305 - accuracy: 0.9193 - mean_IoU: 0.6606
Epoch 00004: val_loss did not improve from 0.43366
725/725 [==============================] - 10s 14ms/sample - loss: 0.2304 - accuracy: 0.9193 - mean_IoU: 0.6608 - val_loss: 0.7008 - val_accuracy: 0.6848 - val_mean_IoU: 0.6643
Epoch 5/50
720/725 [============================>.] - ETA: 0s - loss: 0.1978 - accuracy: 0.9308 - mean_IoU: 0.6702
Epoch 00005: val_loss did not improve from 0.43366
725/725 [==============================] - 10s 14ms/sample - loss: 0.1978 - accuracy: 0.9308 - mean_IoU: 0.6704 - val_loss: 0.5901 - val_accuracy: 0.7582 - val_mean_IoU: 0.6759
Epoch 6/50
720/725 [============================>.] - ETA: 0s - loss: 0.1891 - accuracy: 0.9314 - mean_IoU: 0.6809
Epoch 00006: val_loss did not improve from 0.43366
725/725 [==============================] - 10s 14ms/sample - loss: 0.1888 - accuracy: 0.9316 - mean_IoU: 0.6811 - val_loss: 0.5757 - val_accuracy: 0.7628 - val_mean_IoU: 0.6856
Epoch 7/50
720/725 [============================>.] - ETA: 0s - loss: 0.1821 - accuracy: 0.9313 - mean_IoU: 0.6902
Epoch 00007: val_loss improved from 0.43366 to 0.39979, saving model to models/model-gs-unet.h5
725/725 [==============================] - 11s 15ms/sample - loss: 0.1817 - accuracy: 0.9314 - mean_IoU: 0.6904 - val_loss: 0.3998 - val_accuracy: 0.8377 - val_mean_IoU: 0.6960
Epoch 8/50
720/725 [============================>.] - ETA: 0s - loss: 0.1663 - accuracy: 0.9382 - mean_IoU: 0.7012
Epoch 00008: val_loss improved from 0.39979 to 0.39095, saving model to models/model-gs-unet.h5
725/725 [==============================] - 11s 14ms/sample - loss: 0.1667 - accuracy: 0.9380 - mean_IoU: 0.7014 - val_loss: 0.3909 - val_accuracy: 0.8437 - val_mean_IoU: 0.7066
Epoch 9/50
720/725 [============================>.] - ETA: 0s - loss: 0.1549 - accuracy: 0.9423 - mean_IoU: 0.7112
Epoch 00009: val_loss improved from 0.39095 to 0.30219, saving model to models/model-gs-unet.h5
725/725 [==============================] - 11s 15ms/sample - loss: 0.1550 - accuracy: 0.9422 - mean_IoU: 0.7113 - val_loss: 0.3022 - val_accuracy: 0.8775 - val_mean_IoU: 0.7166
Epoch 10/50
720/725 [============================>.] - ETA: 0s - loss: 0.1466 - accuracy: 0.9447 - mean_IoU: 0.7214
Epoch 00010: val_loss did not improve from 0.30219
725/725 [==============================] - 10s 14ms/sample - loss: 0.1466 - accuracy: 0.9448 - mean_IoU: 0.7215 - val_loss: 0.3188 - val_accuracy: 0.8721 - val_mean_IoU: 0.7255
Epoch 11/50
720/725 [============================>.] - ETA: 0s - loss: 0.1385 - accuracy: 0.9481 - mean_IoU: 0.7295
Epoch 00011: val_loss did not improve from 0.30219
725/725 [==============================] - 10s 14ms/sample - loss: 0.1381 - accuracy: 0.9483 - mean_IoU: 0.7296 - val_loss: 0.5125 - val_accuracy: 0.8174 - val_mean_IoU: 0.7325
Epoch 12/50
720/725 [============================>.] - ETA: 0s - loss: 0.1347 - accuracy: 0.9486 - mean_IoU: 0.7352
Epoch 00012: val_loss improved from 0.30219 to 0.26944, saving model to models/model-gs-unet.h5
725/725 [==============================] - 11s 15ms/sample - loss: 0.1342 - accuracy: 0.9488 - mean_IoU: 0.7353 - val_loss: 0.2694 - val_accuracy: 0.9013 - val_mean_IoU: 0.7391
Epoch 13/50
720/725 [============================>.] - ETA: 0s - loss: 0.1290 - accuracy: 0.9506 - mean_IoU: 0.7426
Epoch 00013: val_loss improved from 0.26944 to 0.19169, saving model to models/model-gs-unet.h5
725/725 [==============================] - 10s 14ms/sample - loss: 0.1289 - accuracy: 0.9506 - mean_IoU: 0.7427 - val_loss: 0.1917 - val_accuracy: 0.9251 - val_mean_IoU: 0.7465
Epoch 14/50
720/725 [============================>.] - ETA: 0s - loss: 0.1337 - accuracy: 0.9478 - mean_IoU: 0.7498
Epoch 00014: val_loss improved from 0.19169 to 0.18700, saving model to models/model-gs-unet.h5
725/725 [==============================] - 10s 14ms/sample - loss: 0.1344 - accuracy: 0.9476 - mean_IoU: 0.7499 - val_loss: 0.1870 - val_accuracy: 0.9274 - val_mean_IoU: 0.7528
Epoch 15/50
720/725 [============================>.] - ETA: 0s - loss: 0.1306 - accuracy: 0.9499 - mean_IoU: 0.7558
Epoch 00015: val_loss did not improve from 0.18700
725/725 [==============================] - 10s 14ms/sample - loss: 0.1303 - accuracy: 0.9501 - mean_IoU: 0.7559 - val_loss: 0.2278 - val_accuracy: 0.9132 - val_mean_IoU: 0.7585
Epoch 16/50
720/725 [============================>.] - ETA: 0s - loss: 0.1163 - accuracy: 0.9555 - mean_IoU: 0.7614
Epoch 00016: val_loss did not improve from 0.18700
725/725 [==============================] - 10s 14ms/sample - loss: 0.1166 - accuracy: 0.9554 - mean_IoU: 0.7614 - val_loss: 0.1888 - val_accuracy: 0.9268 - val_mean_IoU: 0.7641
Epoch 17/50
720/725 [============================>.] - ETA: 0s - loss: 0.1129 - accuracy: 0.9562 - mean_IoU: 0.7670
Epoch 00017: val_loss did not improve from 0.18700
725/725 [==============================] - 10s 14ms/sample - loss: 0.1129 - accuracy: 0.9563 - mean_IoU: 0.7670 - val_loss: 0.2084 - val_accuracy: 0.9233 - val_mean_IoU: 0.7695
Epoch 18/50
720/725 [============================>.] - ETA: 0s - loss: 0.1143 - accuracy: 0.9558 - mean_IoU: 0.7717
Epoch 00018: val_loss did not improve from 0.18700
725/725 [==============================] - 10s 14ms/sample - loss: 0.1150 - accuracy: 0.9556 - mean_IoU: 0.7718 - val_loss: 0.3585 - val_accuracy: 0.8675 - val_mean_IoU: 0.7733
Epoch 19/50
720/725 [============================>.] - ETA: 0s - loss: 0.1091 - accuracy: 0.9579 - mean_IoU: 0.7749
Epoch 00019: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.

Epoch 00019: val_loss did not improve from 0.18700
725/725 [==============================] - 10s 14ms/sample - loss: 0.1094 - accuracy: 0.9577 - mean_IoU: 0.7750 - val_loss: 0.2544 - val_accuracy: 0.9046 - val_mean_IoU: 0.7769
Epoch 20/50
720/725 [============================>.] - ETA: 0s - loss: 0.0964 - accuracy: 0.9637 - mean_IoU: 0.7789
Epoch 00020: val_loss improved from 0.18700 to 0.16215, saving model to models/model-gs-unet.h5
725/725 [==============================] - 10s 14ms/sample - loss: 0.0963 - accuracy: 0.9638 - mean_IoU: 0.7789 - val_loss: 0.1622 - val_accuracy: 0.9348 - val_mean_IoU: 0.7813
Epoch 21/50
720/725 [============================>.] - ETA: 0s - loss: 0.0885 - accuracy: 0.9666 - mean_IoU: 0.7837
Epoch 00021: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0886 - accuracy: 0.9666 - mean_IoU: 0.7837 - val_loss: 0.1775 - val_accuracy: 0.9279 - val_mean_IoU: 0.7858
Epoch 22/50
720/725 [============================>.] - ETA: 0s - loss: 0.0844 - accuracy: 0.9687 - mean_IoU: 0.7880
Epoch 00022: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0845 - accuracy: 0.9687 - mean_IoU: 0.7880 - val_loss: 0.1651 - val_accuracy: 0.9323 - val_mean_IoU: 0.7901
Epoch 23/50
720/725 [============================>.] - ETA: 0s - loss: 0.0838 - accuracy: 0.9690 - mean_IoU: 0.7921
Epoch 00023: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0842 - accuracy: 0.9689 - mean_IoU: 0.7921 - val_loss: 0.1622 - val_accuracy: 0.9328 - val_mean_IoU: 0.7940
Epoch 24/50
720/725 [============================>.] - ETA: 0s - loss: 0.0845 - accuracy: 0.9686 - mean_IoU: 0.7959
Epoch 00024: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0843 - accuracy: 0.9687 - mean_IoU: 0.7959 - val_loss: 0.1669 - val_accuracy: 0.9315 - val_mean_IoU: 0.7978
Epoch 25/50
720/725 [============================>.] - ETA: 0s - loss: 0.0829 - accuracy: 0.9689 - mean_IoU: 0.7994
Epoch 00025: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.

Epoch 00025: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0829 - accuracy: 0.9690 - mean_IoU: 0.7995 - val_loss: 0.1628 - val_accuracy: 0.9334 - val_mean_IoU: 0.8011
Epoch 26/50
720/725 [============================>.] - ETA: 0s - loss: 0.0791 - accuracy: 0.9708 - mean_IoU: 0.8028
Epoch 00026: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0791 - accuracy: 0.9708 - mean_IoU: 0.8028 - val_loss: 0.1805 - val_accuracy: 0.9260 - val_mean_IoU: 0.8043
Epoch 27/50
720/725 [============================>.] - ETA: 0s - loss: 0.0808 - accuracy: 0.9703 - mean_IoU: 0.8059
Epoch 00027: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0807 - accuracy: 0.9704 - mean_IoU: 0.8059 - val_loss: 0.1920 - val_accuracy: 0.9214 - val_mean_IoU: 0.8073
Epoch 28/50
720/725 [============================>.] - ETA: 0s - loss: 0.0802 - accuracy: 0.9704 - mean_IoU: 0.8086
Epoch 00028: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0802 - accuracy: 0.9704 - mean_IoU: 0.8086 - val_loss: 0.2008 - val_accuracy: 0.9183 - val_mean_IoU: 0.8098
Epoch 29/50
720/725 [============================>.] - ETA: 0s - loss: 0.0786 - accuracy: 0.9712 - mean_IoU: 0.8111
Epoch 00029: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0785 - accuracy: 0.9713 - mean_IoU: 0.8112 - val_loss: 0.2049 - val_accuracy: 0.9170 - val_mean_IoU: 0.8124
Epoch 30/50
720/725 [============================>.] - ETA: 0s - loss: 0.0788 - accuracy: 0.9711 - mean_IoU: 0.8135
Epoch 00030: ReduceLROnPlateau reducing learning rate to 1e-05.

Epoch 00030: val_loss did not improve from 0.16215
725/725 [==============================] - 10s 14ms/sample - loss: 0.0787 - accuracy: 0.9711 - mean_IoU: 0.8135 - val_loss: 0.2074 - val_accuracy: 0.9163 - val_mean_IoU: 0.8147
Epoch 00030: early stopping

In [19]:

plt.figure(figsize=(8, 8))
plt.title("Learning curve")
plt.plot(results.history["loss"], label="loss")
plt.plot(results.history["val_loss"], label="val_loss")
plt.plot( np.argmin(results.history["val_loss"]), np.min(results.history["val_loss"]), marker="x", color="r", label="best model")
plt.xlabel("Epochs")
plt.ylabel("log_loss")
plt.legend();

Inference¶

In [6]:

from tensorflow.keras.optimizers import Adam
from tensorflow.keras import metrics

from tensorflow.keras.models import load_model

model = load_model('models/model-gs-unet.h5', compile=False)
model.compile(optimizer=Adam(), loss="binary_crossentropy", metrics=["accuracy", mean_IoU])

In [7]:

# Evaluate on train set (this must be equals to the best log_loss)
model.evaluate(X_train, y_train, verbose=1)

725/725 [==============================] - 4s 5ms/sample - loss: 0.1232 - accuracy: 0.9499 - mean_IoU: 0.8752

Out[7]:

[0.12322666003786284, 0.9499445, 0.875184]

In [8]:

# Evaluate on test set (this must be equals to the best log_loss)
model.evaluate(X_test, y_test, verbose=1)

242/242 [==============================] - 1s 5ms/sample - loss: 0.1622 - accuracy: 0.9348 - mean_IoU: 0.8639

Out[8]:

[0.16215478450306192, 0.9348493, 0.863872]

In [9]:

# Predict on train and test
preds_train = model.predict(X_train, verbose=1)
preds_test = model.predict(X_test, verbose=1)

725/725 [==============================] - 2s 2ms/sample
242/242 [==============================] - 1s 2ms/sample

Predictions on training set¶

In [10]:

from utils.plot import plot_samples

In [11]:

plot_samples(X_train, y_train, preds_train, num_samples=4, prob=0.5, seed=1)

Predictions on test set¶

In [12]:

plot_samples(X_test, y_test, preds_test, num_samples=4, prob=0.5, seed=1)

Predictions on selected images¶

In [15]:

import glob
from path import Path
from tqdm import tqdm_notebook
import tensorflow as tf
from tensorflow.keras.preprocessing.image import array_to_img, img_to_array, load_img

DEMO_PATH = Path('./demo')

image_files = glob.glob(DEMO_PATH / '*.*')
X_misc = np.zeros((len(image_files), img_height, img_width, img_channel), dtype=np.float32)

for i, image_file in enumerate(tqdm_notebook(image_files, total=len(image_files), desc="Loading images")):
    img = load_img(image_file, grayscale=False if img_channel == 3 else True)
    x_img = img_to_array(img)
    x_img = tf.image.resize(x_img, (img_height, img_width))
    X_misc[i] = x_img / 255.0

HBox(children=(IntProgress(value=0, description='Loading images', max=9, style=ProgressStyle(description_width…

In [16]:

preds_misc = model.predict(X_misc, verbose=1)

9/9 [==============================] - 0s 39ms/sample

In [17]:

indexs = list(range(len(X_misc)))

plot_samples(X_misc, None, preds_misc, indexs=indexs[:len(indexs)//2], prob=0.5)
plot_samples(X_misc, None, preds_misc, indexs=indexs[len(indexs)//2:], prob=0.5)

In [18]:

from tensorflow.keras.preprocessing.image import save_img
from tensorflow.image import resize
import os

OUTPUT_PATH = DEMO_PATH / 'predict'

if not os.path.exists(OUTPUT_PATH): os.mkdir(OUTPUT_PATH)

binary_preds = (preds_misc > 0.5).astype(np.uint8)
for index, (image, label) in enumerate((tqdm_notebook(zip(X_misc, binary_preds), total=len(X_misc), desc="Saving images and labels"))):
    image_fname = OUTPUT_PATH / 'rgb_%04d.png' % index
    label_fname = OUTPUT_PATH / 'label_%04d.png' % index
    
    image = resize(image, (480, 640))
    save_img(image_fname, image, scale=True)
    
    label = resize(label, (480, 640))
    save_img(label_fname, label, scale=True)

HBox(children=(IntProgress(value=0, description='Saving images and labels', max=9, style=ProgressStyle(descrip…

In [ ]: