Data Modeling [04]: Keras and Tensorflow
Published:
Workflow of Keras (Tensorflow). Tutorial from Keras on the Kaggle Cats vs Dogs binary classification dataset.
Datasets Loading
Download and unzip dataset.
!curl -O https://download.microsoft.com/download/3/E/1/3E1C3F21-ECDB-4869-8368-6DEBA77B919F/kagglecatsanddogs_3367a.zip
!unzip -q kagglecatsanddogs_3367a.zip
!ls
kagglecatsanddogs_3367a.zip
kearas.py
MSR-LA - 3467.docx
PetImages
readme[1].txt
!ls PetImages
Cat
Dog
We now have a PetImages folder containing two subfolders Cat and Dog. Next we delete the corrupted images:
import os
num_skipped = 0
for folder_name in ("Cat", "Dog"):
folder_path = os.path.join("PetImages", folder_name)
for fname in os.listdir(folder_path):
fpath = os.path.join(folder_path, fname)
try:
fobj = open(fpath, "rb") # read mode 'r' with binary I/O 'b'.
is_jfif = tf.compat.as_bytes("JFIF") in fobj.peek(10)
# JPG files contain the string "JFIF" at the beginning of the file, encoded as bytes
# fobj.peek(10) returns the first 10 baytes of the file
finally:
fobj.close()
if not is_jfif:
num_skipped += 1
# Delete corrupted image
os.remove(fpath)
print("Deleted %d images" % num_skipped)
Deleted 1578 images
Training & Testing Dataset Generation
image_size = (180, 180)
batch_size = 32
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"PetImages",
validation_split=0.2,
subset="training",
seed=1337,
image_size=image_size,
batch_size=batch_size,
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"PetImages",
validation_split=0.2,
subset="validation",
seed=1337,
image_size=image_size,
batch_size=batch_size,
)
Found 23422 files belonging to 2 classes.
Using 18738 files for training.
Found 23422 files belonging to 2 classes.
Using 4684 files for validation.
train_ds
Out[11]: <BatchDataset element_spec=(TensorSpec(shape=(None, 180, 180, 3), dtype=tf.float32, name=None), TensorSpec(shape=(None,), dtype=tf.int32, name=None))>
The type of the data in the dataset is tensor tuple. The first element is a 3D tensor storing the information of the color images. The second element stores the labels with ‘1’ being ‘dog’ and ‘0’ for ‘cat’.
type(next(iter(train_ds)))
Out[16]: tuple
next(iter(train_ds))
Out[15]:
(<tf.Tensor: shape=(32, 180, 180, 3), dtype=float32, numpy=
array([[[[160.11111 , 155.11111 , 126.111115 ],
[159.33333 , 154.33333 , 125.333336 ],
[158.74692 , 153.74692 , 124.74692 ],
...,
[158.43056 , 150.43056 , 127.43056 ],
[155.76387 , 148.4305 , 125.09718 ],
[156.43056 , 150.43056 , 126.43056 ]],
[[161. , 156. , 127. ],
[160.09721 , 155.09721 , 126.09722 ],
[159. , 154. , 125. ],
...,
[160.29167 , 152.29167 , 129.29167 ],
[157.19446 , 149.86108 , 126.52776 ],
[157. , 151. , 127. ]],
[[162. , 157. , 128. ],
[161. , 156. , 127. ],
[160. , 155. , 126. ],
...,
[161.47067 , 153.47067 , 130.47067 ],
[158.56483 , 151.23146 , 127.89815 ],
[158. , 152. , 128. ]],
...,
[[189.73611 , 180.73611 , 149.43057 ],
[187.05092 , 178.53702 , 153.49077 ],
[189.58334 , 182.29167 , 165.83492 ],
...,
[241.69446 , 238. , 239. ],
[241.31023 , 237.61577 , 238.61577 ],
[240.54169 , 236.84723 , 237.84723 ]],
[[187.4722 , 176.34717 , 149.4722 ],
[190.45836 , 181.08336 , 161.8195 ],
[205.1992 , 196.88437 , 186.8196 ],
...,
[241.29166 , 237.29166 , 236.87497 ],
[240.95834 , 236.95834 , 236.54166 ],
[240.29166 , 236.29166 , 235.87497 ]],
[[188.0278 , 176.0278 , 153.16672 ],
[197.42134 , 187.18523 , 172.70378 ],
[224.15906 , 215.79486 , 210.6313 ],
...,
[238.86108 , 234.86108 , 233.86108 ],
[238.71758 , 234.71758 , 233.71758 ],
[238.43054 , 234.43054 , 233.43054 ]]],
[[[ 12.382964 , 13.017286 , 16.294075 ],
[ 31.420378 , 32.509266 , 34.14667 ],
[ 31.04445 , 32.04445 , 30.772844 ],
...,
[ 30.949389 , 31.99692 , 30.582722 ],
[ 30.335556 , 32.30223 , 30.213333 ],
[ 12.977433 , 14.572387 , 12.343099 ]],
[[ 82.092964 , 82.66371 , 84.87778 ],
[251.55112 , 253.16333 , 251.84222 ],
[252.10556 , 253.69629 , 249.37407 ],
...,
[253.50742 , 254.75371 , 250.14075 ],
[247.13661 , 249.10329 , 245.42772 ],
[ 81.839966 , 82.839966 , 81.51776 ]],
[[ 82.108025 , 83.108025 , 82.82716 ],
[254. , 255. , 251.84444 ],
[143.85802 , 146.70679 , 139.26234 ],
...,
[129.9105 , 132.13272 , 124.688286 ],
[250.61664 , 251.99257 , 247.05182 ],
[ 81.839966 , 83.25417 , 81.123955 ]],
...,
[[ 82.44257 , 82.44257 , 82.44257 ],
[253.39626 , 253.39626 , 251.46292 ],
[110.7682 , 110.7682 , 108.7682 ],
...,
[ 98.0678 , 99.305466 , 97.203575 ],
[252.87965 , 254.87965 , 251.87965 ],
[ 81.18265 , 83.18265 , 80.44624 ]],
[[ 81.81224 , 81.81224 , 81.81224 ],
[252.32114 , 252.32114 , 250.3878 ],
[254.57224 , 254.57224 , 251.79814 ],
...,
[253.0815 , 254.55186 , 251.63336 ],
[253.5967 , 254.96335 , 252.6389 ],
[ 81.36742 , 82.73407 , 81.45335 ]],
[[ 14.206032 , 14.206032 , 14.206032 ],
[ 32.62375 , 32.62375 , 30.690414 ],
[ 32.869766 , 33.406204 , 31.793259 ],
...,
[ 31.529648 , 33.1136 , 30.871002 ],
[ 34.434906 , 36.283447 , 33.472305 ],
[ 11.786362 , 13.381325 , 12.4207 ]]],
[[[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
...,
[ 94.194496 , 100.194496 , 98.194496 ],
[ 88.87489 , 95.87483 , 94.208145 ],
[ 88.23617 , 97.23617 , 96.23617 ]],
[[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
...,
[102.43056 , 108.43056 , 106.43056 ],
[ 90.8332 , 97.83314 , 96.16645 ],
[ 91.65283 , 100.65283 , 99.65283 ]],
[[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
[ 23. , 15. , 12. ],
...,
[ 85.13889 , 91.13889 , 89.13889 ],
[ 91.20825 , 98.20819 , 96.541504 ],
[ 92.18058 , 101.18058 , 100.18058 ]],
...,
[[ 22.88889 , 23.88889 , 17.88889 ],
[ 23.666666 , 24.666666 , 18.666666 ],
[ 24.444445 , 25.444445 , 19.444445 ],
...,
[ 21. , 13. , 10. ],
[ 20.194443 , 13.194383 , 10.527757 ],
[ 20. , 15. , 10.416687 ]],
[[ 22.88889 , 23.88889 , 17.88889 ],
[ 23.666666 , 24.666666 , 18.666666 ],
[ 24.444445 , 25.444445 , 19.444445 ],
...,
[ 21.375 , 13.375 , 10.375 ],
[ 20.916695 , 13.916634 , 10.583321 ],
[ 20. , 15. , 9. ]],
[[ 22.88889 , 23.88889 , 17.88889 ],
[ 23.666666 , 24.666666 , 18.666666 ],
[ 24.444445 , 25.444445 , 19.444445 ],
...,
[ 22. , 14. , 11. ],
[ 21.638926 , 14.6388645, 10.638926 ],
[ 20. , 15.458313 , 8.083374 ]]],
...,
[[[179.45139 , 163.59723 , 147.13426 ],
[168.74652 , 158.46529 , 143.15973 ],
[153.13889 , 146.2604 , 131.50462 ],
...,
[169.28354 , 85.74996 , 39.51964 ],
[179.93735 , 99.152596 , 50.919983 ],
[205.23584 , 124.38401 , 64.75918 ]],
[[174.07985 , 160.04514 , 146.65625 ],
[163.1875 , 156.08333 , 139.40625 ],
[149.75 , 144. , 123.208336 ],
...,
[207.09029 , 129.24649 , 81.9722 ],
[206.65623 , 133.60416 , 87.14579 ],
[205.53476 , 134.25346 , 78.75018 ]],
[[163.60301 , 154.29745 , 143.83333 ],
[151.3646 , 147.00694 , 133.2639 ],
[148.22223 , 141.32986 , 125.3588 ],
...,
[194.9248 , 120.857666 , 72.093765 ],
[192.89922 , 120.54506 , 75.61791 ],
[200.35065 , 127.52542 , 84.99767 ]],
...,
[[252.2442 , 215.329 , 170.30817 ],
[254.1771 , 221.55212 , 184.96884 ],
[253.70262 , 170.11118 , 130.9874 ],
...,
[254.32887 , 220.88106 , 183.56517 ],
[242.63867 , 182.59323 , 139.37448 ],
[238.10547 , 185.81163 , 140.59872 ]],
[[251.78473 , 164.71875 , 122.5 ],
[251.10417 , 165.64583 , 122.010414 ],
[252.9618 , 173.84027 , 126.802086 ],
...,
[241.7326 , 171.10411 , 122.78468 ],
[247.79161 , 182.0416 , 133.06238 ],
[248.2813 , 185.01738 , 140.53123 ]],
[[253.21645 , 181.70467 , 127.34817 ],
[252.8993 , 179.89238 , 128.44455 ],
[253.0382 , 177.73972 , 128.71077 ],
...,
[243.54869 , 178.37743 , 122.54984 ],
[234.27773 , 164.17004 , 115.9755 ],
[240.52554 , 172.87396 , 129.46648 ]]],
[[[ 29.949074 , 79.94907 , 138.94907 ],
[ 30.847221 , 80.84722 , 139.84723 ],
[ 31.541666 , 81.541664 , 140.54167 ],
...,
[ 27. , 78. , 141. ],
[ 27.208353 , 76.541725 , 139.20842 ],
[ 28. , 74. , 136. ]],
[[ 30.625 , 80.625 , 139.625 ],
[ 31.875 , 81.875 , 140.875 ],
[ 32.625 , 82.625 , 141.625 ],
...,
[ 27.791672 , 78.79167 , 141.79167 ],
[ 27.875008 , 77.20838 , 139.87506 ],
[ 29.555573 , 75.55557 , 137.55557 ]],
[[ 32.708332 , 82.708336 , 141.70833 ],
[ 33.708332 , 83.708336 , 142.70833 ],
[ 34.708332 , 84.708336 , 143.70833 ],
...,
[ 28.837967 , 79.83797 , 142.83797 ],
[ 28.76386 , 78.56947 , 141.23616 ],
[ 28.708332 , 76.708336 , 138.70833 ]],
...,
[[142.96751 , 153.96751 , 156.55083 ],
[151.70827 , 162.70827 , 167.09714 ],
[155.49063 , 166.19897 , 173.07394 ],
...,
[126.06967 , 137.75485 , 154.06046 ],
[121.819405 , 134.15273 , 149.76382 ],
[111.319305 , 124.319305 , 140.3193 ]],
[[144.95833 , 155.95833 , 160.70833 ],
[153.83333 , 164.58333 , 170.91667 ],
[165.25 , 175.25 , 184.25 ],
...,
[139.7777 , 151.98602 , 163.91661 ],
[126.62484 , 139.08315 , 152.87479 ],
[117.49997 , 130.49997 , 146.49997 ]],
[[135.6944 , 146.1759 , 153.78233 ],
[145.43053 , 155.43053 , 164.88884 ],
[164.25917 , 174.25917 , 183.71748 ],
...,
[133.21252 , 146.21252 , 154.75421 ],
[128.98642 , 142.31973 , 152.65298 ],
[111.6572 , 124.6572 , 140.65721 ]]],
[[[191.825 , 171.825 , 110.825 ],
[185.17375 , 164.32376 , 106.72375 ],
[148.75 , 127. , 75.25 ],
...,
[ 28.29375 , 28.29375 , 16.79375 ],
[ 26.025005 , 26.025005 , 16.025005 ],
[ 25.175 , 25.175 , 15.175 ]],
[[191. , 171. , 110. ],
[185.05 , 164.2 , 106.6 ],
[148.75 , 127. , 75.25 ],
...,
[ 28.36875 , 28.63125 , 17. ],
[ 26.325006 , 27.296259 , 17.0075 ],
[ 25.475 , 26. , 17.05 ]],
[[191. , 171. , 110. ],
[185.79375 , 165.6875 , 106.6 ],
[149.625 , 128.09375 , 75.90625 ],
...,
[ 28.03125 , 28.3125 , 18.53125 ],
[ 24.843761 , 25.950012 , 17.63126 ],
[ 23.25 , 24.25 , 16.25 ]],
...,
[[123. , 101.875 , 56.375 ],
[130.54375 , 109.41875 , 63.91875 ],
[128.875 , 107.9375 , 62.34375 ],
...,
[203.125 , 183.4375 , 152.625 ],
[199.48752 , 179.98752 , 147.07501 ],
[198. , 178.5 , 145.375 ]],
[[130.95001 , 108.95001 , 59.475006 ],
[133.81879 , 111.81879 , 62.34378 ],
[143.96257 , 123.46257 , 73.237564 ],
...,
[192.81876 , 167.25 , 136.2375 ],
[189.36124 , 165.07872 , 132.37497 ],
[188.14996 , 164.67496 , 130.67496 ]],
[[133.65 , 111.649994 , 61.649994 ],
[142.95747 , 120.95747 , 70.95747 ],
[159.69371 , 139.19371 , 88.44371 ],
...,
[197.09373 , 170.34373 , 139.59373 ],
[189.53003 , 163.83002 , 130.43254 ],
[184.175 , 160.175 , 124.52501 ]]]], dtype=float32)>,
<tf.Tensor: shape=(32,), dtype=int32, numpy=
array([0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 1])>)
Visualization
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 10))
for images, labels in train_ds.take(1): # take the first batch only
for i in range(9):
ax = plt.subplot(3, 3, i + 1)
plt.imshow(images[i].numpy().astype("uint8"))
label = 'dog' if labels[i] else 'cat'
plt.title(label)
plt.axis("off")
Dataset Preprocessing
Firstly, it’s a good practice to artificially introduce sample diversity by applying random yet realistic transformations to the training images, such as random horizontal flipping or small random rotations.
data_augmentation = keras.Sequential(
[
layers.RandomFlip("horizontal"),
layers.RandomRotation(0.1),
]
)
Secondly, we should standardize the pixel values to be in the ‘[0, 1]’ by using a Rescaling layer at the start of the model.
For GPU
inputs = keras.Input(shape=input_shape)
X = data_augmentation(inputs)
X = layers.Rescaling(1./255)(X)
- With this option, data augmentation will happen on device, synchronously with the rest of the model execution.
- Input samples will only be augmented during
fit(), not when callingevaluate()orpredict().
For CPU
augmented_train_ds = train_ds.map(
lambda x, y: (data_augmentation(x, training=True), y))
- With this option, data augmentation will happen on CPU asynchronously, and will be buffered before going into the model.
Model Building
Prefetching configuration:
train_ds = train_ds.prefetch(buffer_size=32)
val_ds = val_ds.prefetch(buffer_size=32)
Xception neural network:
def make_model(input_shape, num_classes):
inputs = keras.Input(shape=input_shape)
# Image augmentation block
x = data_augmentation(inputs) # flip and rotation
# Normalization
x = layers.Rescaling(1.0 / 255)(x)
# Entry block - 2 Conv layers
x = layers.Conv2D(32, 3, strides=2, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
previous_block_activation = x # Set aside residual
# Each Block: 2 Conv + 1 Pool + 1 Res
for size in [128, 256, 512, 728]:
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
# Separable convolutions consist of first performing a depthwise spatial convolution
# followed by a pointwise convolution which mixes the resulting output channels.
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.MaxPooling2D(3, strides=2, padding="same")(x)
# Project residual
residual = layers.Conv2D(size, 1, strides=2, padding="same")(
previous_block_activation
)
x = layers.add([x, residual]) # Add back residual
previous_block_activation = x # Set aside next residual
x = layers.SeparableConv2D(1024, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.GlobalAveragePooling2D()(x)
if num_classes == 2:
activation = "sigmoid"
units = 1
else:
activation = "softmax"
units = num_classes
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(units, activation=activation)(x)
return keras.Model(inputs, outputs)
model = make_model(input_shape=image_size + (3,), num_classes=2)
dot_img_file = 'model_xception.png'
tf.keras.utils.plot_model(model, to_file=dot_img_file, show_shapes=True)
Model summary:
model.summary()
Model: "model_4"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_5 (InputLayer) [(None, 180, 180, 3)] 0 []
sequential_4 (Sequential) (None, 180, 180, 3) 0 ['input_5[0][0]']
rescaling_4 (Rescaling) (None, 180, 180, 3) 0 ['sequential_4[0][0]']
conv2d_24 (Conv2D) (None, 90, 90, 32) 896 ['rescaling_4[0][0]']
batch_normalization_44 (BatchNormalization) (None, 90, 90, 32) 128 ['conv2d_24[0][0]']
activation_44 (Activation) (None, 90, 90, 32) 0 ['batch_normalization_44[0][0]']
conv2d_25 (Conv2D) (None, 90, 90, 64) 18496 ['activation_44[0][0]']
batch_normalization_45 (BatchNormalization) (None, 90, 90, 64) 256 ['conv2d_25[0][0]']
activation_45 (Activation) (None, 90, 90, 64) 0 ['batch_normalization_45[0][0]']
activation_46 (Activation) (None, 90, 90, 64) 0 ['activation_45[0][0]']
separable_conv2d_36 (SeparableConv2D) (None, 90, 90, 128) 8896 ['activation_46[0][0]']
batch_normalization_46 (BatchNormalization) (None, 90, 90, 128) 512 ['separable_conv2d_36[0][0]']
activation_47 (Activation) (None, 90, 90, 128) 0 ['batch_normalization_46[0][0]']
separable_conv2d_37 (SeparableConv2D) (None, 90, 90, 128) 17664 ['activation_47[0][0]']
batch_normalization_47 (BatchNormalization) (None, 90, 90, 128) 512 ['separable_conv2d_37[0][0]']
max_pooling2d_16 (MaxPooling2D) (None, 45, 45, 128) 0 ['batch_normalization_47[0][0]']
conv2d_26 (Conv2D) (None, 45, 45, 128) 8320 ['activation_45[0][0]']
add_16 (Add) (None, 45, 45, 128) 0 ['max_pooling2d_16[0][0]', 'conv2d_26[0][0]']
activation_48 (Activation) (None, 45, 45, 128) 0 ['add_16[0][0]']
separable_conv2d_38 (SeparableConv2D) (None, 45, 45, 256) 34176 ['activation_48[0][0]']
batch_normalization_48 (BatchNormalization) (None, 45, 45, 256) 1024 ['separable_conv2d_38[0][0]']
activation_49 (Activation) (None, 45, 45, 256) 0 ['batch_normalization_48[0][0]']
separable_conv2d_39 (SeparableConv2D) (None, 45, 45, 256) 68096 ['activation_49[0][0]']
batch_normalization_49 (BatchNormalization) (None, 45, 45, 256) 1024 ['separable_conv2d_39[0][0]']
max_pooling2d_17 (MaxPooling2D) (None, 23, 23, 256) 0 ['batch_normalization_49[0][0]']
conv2d_27 (Conv2D) (None, 23, 23, 256) 33024 ['add_16[0][0]']
add_17 (Add) (None, 23, 23, 256) 0 ['max_pooling2d_17[0][0]', 'conv2d_27[0][0]']
activation_50 (Activation) (None, 23, 23, 256) 0 ['add_17[0][0]']
separable_conv2d_40 (SeparableConv2D) (None, 23, 23, 512) 133888 ['activation_50[0][0]']
batch_normalization_50 (BatchNormalization) (None, 23, 23, 512) 2048 ['separable_conv2d_40[0][0]']
activation_51 (Activation) (None, 23, 23, 512) 0 ['batch_normalization_50[0][0]']
separable_conv2d_41 (SeparableConv2D) (None, 23, 23, 512) 267264 ['activation_51[0][0]']
batch_normalization_51 (BatchNormalization) (None, 23, 23, 512) 2048 ['separable_conv2d_41[0][0]']
max_pooling2d_18 (MaxPooling2D) (None, 12, 12, 512) 0 ['batch_normalization_51[0][0]']
conv2d_28 (Conv2D) (None, 12, 12, 512) 131584 ['add_17[0][0]']
add_18 (Add) (None, 12, 12, 512) 0 ['max_pooling2d_18[0][0]', 'conv2d_28[0][0]']
activation_52 (Activation) (None, 12, 12, 512) 0 ['add_18[0][0]']
separable_conv2d_42 (SeparableConv2D) (None, 12, 12, 728) 378072 ['activation_52[0][0]']
batch_normalization_52 (BatchNormalization) (None, 12, 12, 728) 2912 ['separable_conv2d_42[0][0]']
activation_53 (Activation) (None, 12, 12, 728) 0 ['batch_normalization_52[0][0]']
separable_conv2d_43 (SeparableConv2D) (None, 12, 12, 728) 537264 ['activation_53[0][0]']
batch_normalization_53 (BatchNormalization) (None, 12, 12, 728) 2912 ['separable_conv2d_43[0][0]']
max_pooling2d_19 (MaxPooling2D) (None, 6, 6, 728) 0 ['batch_normalization_53[0][0]']
conv2d_29 (Conv2D) (None, 6, 6, 728) 373464 ['add_18[0][0]']
add_19 (Add) (None, 6, 6, 728) 0 ['max_pooling2d_19[0][0]', 'conv2d_29[0][0]']
separable_conv2d_44 (SeparableConv2D) (None, 6, 6, 1024) 753048 ['add_19[0][0]']
batch_normalization_54 (BatchNormalization) (None, 6, 6, 1024) 4096 ['separable_conv2d_44[0][0]']
activation_54 (Activation) (None, 6, 6, 1024) 0 ['batch_normalization_54[0][0]']
global_average_pooling2d_4 (GlobalAveragePooling2D) (None, 1024) 0 ['activation_54[0][0]']
dropout_4 (Dropout) (None, 1024) 0 ['global_average_pooling2d_4[0][0]']
dense_4 (Dense) (None, 1) 1025 ['dropout_4[0][0]']
==================================================================================================
Total params: 2,782,649
Trainable params: 2,773,913
Non-trainable params: 8,736
__________________________________________________________________________________________________
Training
epochs = 50
callbacks = [
keras.callbacks.ModelCheckpoint("save_at_{epoch}.h5"),
]
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="binary_crossentropy",
metrics=["accuracy"],
)
history = model.fit(
train_ds, epochs=epochs, callbacks=callbacks, validation_data=val_ds,
)
It took about 7645.4 seconds to finish 50 epochs, and we obtain an accuracy of 96.3% on the validation dataset.
Epoch 50/50
loss: 0.0495 - accuracy: 0.9814 - val_loss: 0.1077 - val_accuracy: 0.9633
Total time cost: 7645.395040399999 s
Plot the results:
figure = plt.figure()
ax = figure.add_subplot(1,1,1)
ax.plot(history.history['loss'],linewidth=2,label='train_loss')
ax.plot(history.history['accuracy'],'--',linewidth=2,label='train accuracy')
ax.plot(history.history['val_loss'],'-.',linewidth=2,label='validation loss')
ax.plot(history.history['val_accuracy'],':',linewidth=2,label='validation accuracy')
ax.set_xlabel('epochs')
ax.set_ylabel('loss/accuracy')
ax.legend()
Testing
img = keras.preprocessing.image.load_img(
"PetImages/Cat/6779.jpg", target_size=image_size
)
plt.imshow(img)
img_array = keras.preprocessing.image.img_to_array(img)
img_array = tf.expand_dims(img_array, 0) # Create batch axis
predictions = model.predict(img_array)
score = predictions[0]
print(
"This image is %.2f percent cat and %.2f percent dog."
% (100 * (1 - score), 100 * score)
)
Result:
This image is 96.00 percent cat and 4.00 percent dog.
Complete Code
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import os
# =============================================================================
# Load Dataset
# =============================================================================
# =============================================================================
# !curl -O https://download.microsoft.com/download/3/E/1/3E1C3F21-ECDB-4869-8368-6DEBA77B919F/kagglecatsanddogs_3367a.zip
#
# !unzip -q kagglecatsanddogs_3367a.zip
# !ls
# !ls PetImages
# =============================================================================
num_skipped = 0
for folder_name in ("Cat", "Dog"):
folder_path = os.path.join("PetImages", folder_name)
for fname in os.listdir(folder_path):
fpath = os.path.join(folder_path, fname)
try:
fobj = open(fpath, "rb")
is_jfif = tf.compat.as_bytes("JFIF") in fobj.peek(10)
finally:
fobj.close()
if not is_jfif:
num_skipped += 1
# Delete corrupted image
os.remove(fpath)
print("Deleted %d images" % num_skipped)
# =============================================================================
# Training and Validation Dataset
# =============================================================================
image_size = (180, 180)
batch_size = 32
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"PetImages",
validation_split=0.2,
subset="training",
seed=1337,
image_size=image_size,
batch_size=batch_size,
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"PetImages",
validation_split=0.2,
subset="validation",
seed=1337,
image_size=image_size,
batch_size=batch_size,
)
# =============================================================================
data_augmentation = keras.Sequential(
[
layers.RandomFlip("horizontal"),
layers.RandomRotation(0.1),
]
)
# =============================================================================
# Model
# =============================================================================
def make_model(input_shape, num_classes):
inputs = keras.Input(shape=input_shape)
# Image augmentation block
x = data_augmentation(inputs) # flip and rotation
# Entry block
x = layers.Rescaling(1.0 / 255)(x)
x = layers.Conv2D(32, 3, strides=2, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
previous_block_activation = x # Set aside residual
for size in [128, 256, 512, 728]:
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.MaxPooling2D(3, strides=2, padding="same")(x)
# Project residual
residual = layers.Conv2D(size, 1, strides=2, padding="same")(
previous_block_activation
)
x = layers.add([x, residual]) # Add back residual
previous_block_activation = x # Set aside next residual
x = layers.SeparableConv2D(1024, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.GlobalAveragePooling2D()(x)
if num_classes == 2:
activation = "sigmoid"
units = 1
else:
activation = "softmax"
units = num_classes
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(units, activation=activation)(x)
return keras.Model(inputs, outputs)
model = make_model(input_shape=image_size + (3,), num_classes=2)
# =============================================================================
# Train
# =============================================================================
epochs = 50
callbacks = [
keras.callbacks.ModelCheckpoint("save_at_{epoch}.h5"),
]
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="binary_crossentropy",
metrics=["accuracy"],
)
history = model.fit(
train_ds, epochs=epochs, callbacks=callbacks, validation_data=val_ds,
)
# =============================================================================
# Test
# =============================================================================
img = keras.preprocessing.image.load_img(
"PetImages/Cat/6779.jpg", target_size=image_size
)
plt.imshow(img)
img_array = keras.preprocessing.image.img_to_array(img)
img_array = tf.expand_dims(img_array, 0) # Create batch axis
predictions = model.predict(img_array)
score = predictions[0]
print(
"This image is %.2f percent cat and %.2f percent dog."
% (100 * (1 - score), 100 * score)
)
Cuda Configuaration
- OS: Windows 10
- GPU: GTX 1660 Ti, 6.0 G
- Cuda: cuda_11.3.0_465.89
- CuDNN: Download cuDNN v8.2.1, for CUDA 11.x
- Tensorflow: 2.8
Using cuda, it took about 156 s to run a single epoch:
98/586 [====>.........................] - ETA: 2:01 - loss: 0.7277 - accuracy: 0.5842
142/586 [======>.......................] - ETA: 1:50 - loss: 0.7136 - accuracy: 0.5900
293/586 [==============>...............] - ETA: 1:13 - loss: 0.6683 - accuracy: 0.6272
325/586 [===============>..............] - ETA: 1:05 - loss: 0.6631 - accuracy: 0.6319
464/586 [======================>.......] - ETA: 30s - loss: 0.6369 - accuracy: 0.6544
544/586 [==========================>...] - ETA: 10s - loss: 0.6276 - accuracy: 0.6612
586/586 [==============================] - ETA: 0s - loss: 0.6202 - accuracy: 0.6674
586/586 [==============================] - 156s 264ms/step - loss: 0.6202 - accuracy: 0.6674 - val_loss: 0.7241 - val_accuracy: 0.6008
Comments