ML: Image segmentation with a U-Net-like architecture

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217

# https://keras.io/examples/vision/oxford_pets_image_segmentation/ import os input_dir="images/" target_dir="annotations/trimaps/" img_size=(160, 160) num_classes=3 batch_size=32 input_img_paths=sorted( [ os.path.join(input_dir, fname) for fname in os.listdir(input_dir) if fname.endswith(".jpg") ] ) target_img_paths=sorted( [ os.path.join(target_dir, fname) for fname in os.listdir(target_dir) if fname.endswith(".png") and not fname.startswith(".") ] ) print("Number of samples:", len(input_img_paths)) for input_path, target_path in zip(input_img_paths[:10], target_img_paths[:10]): print(input_path, "|", target_path) from IPython.display import Image, display from keras.utils import load_img from PIL import ImageOps import matplotlib.pyplot as plt # Display input image #10 display(Image(filename=input_img_paths[9])) # Display auto-contrast version of corresponding target (per-pixel categories) img=ImageOps.autocontrast(load_img(target_img_paths[9])) display(img) import keras import numpy as np from tensorflow import data as tf_data from tensorflow import image as tf_image from tensorflow import io as tf_io def get_dataset( batch_size, img_size, input_img_paths, target_img_paths, max_dataset_len=None, ): def load_img_masks(input_img_path, target_img_path): input_img=tf_io.read_file(input_img_path) # Reads the contents of file. input_img=tf_io.decode_png(input_img, channels=3) # Decode a PNG-encoded image to a uint8 or uint16 tensor. input_img=tf_image.resize(input_img, img_size) input_img=tf_image.convert_image_dtype(input_img, "float32") target_img=tf_io.read_file(target_img_path) target_img=tf_io.decode_png(target_img, channels=1) target_img=tf_image.resize(target_img, img_size, method="nearest") target_img=tf_image.convert_image_dtype(target_img, "uint8") # Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2: target_img=target_img-1 return input_img, target_img # For faster debugging, limit the size of data if max_dataset_len: # checks if max_dataset_len has a truthy value. If max_dataset_len is None, 0, or even an empty container like [] or '', the condition will be False input_img_paths=input_img_paths[: max_dataset_len] target_img_paths=target_img_paths[:max_dataset_len] dataset=tf_data.Dataset.from_tensor_slices((input_img_paths, target_img_paths)) # Represents a potentially large set of elements dataset=dataset.map(load_img_masks, num_parallel_calls=tf_data.AUTOTUNE) return dataset.batch(batch_size) from keras import layers def get_model(img_size, num_classes): inputs=keras.Input(shape=img_size+(3, )) # Entry block x=layers.Conv2D(32, 3, strides=2, padding="same")(inputs) # 2D convolution layer x=layers.BatchNormalization()(x) x=layers.Activation("relu")(x) previous_block_activation=x # Block 1, 2, 3 are identical apart from the feature depth for filters in [64, 128, 256]: x=layers.Activation("relu")(x) x=layers.SeparableConv2D(filters, 3, padding="same")(x) x=layers.BatchNormalization()(x) x=layers.Activation("relu")(x) x=layers.SeparableConv2D(filters, 3, padding="same")(x) x=layers.BatchNormalization()(x) x=layers.MaxPooling2D(3, strides=2, padding="same")(x) # Project residual residual=layers.Conv2D(filters, 1, strides=2, padding="same")( previous_block_activation ) x=layers.add([x, residual]) previous_block_activation=x for filters in [256, 128, 64,32]: x=layers.Activation("relu")(x) x=layers.Conv2DTranspose(filters, 3, padding="same")(x) x=layers.BatchNormalization()(x) x=layers.Activation("relu")(x) x=layers.Conv2DTranspose(filters, 3, padding="same")(x) x=layers.BatchNormalization()(x) x=layers.UpSampling2D(2)(x) # Project residual residual=layers.UpSampling2D(2)(previous_block_activation) residual=layers.Conv2D(filters, 1, padding="same")(residual) x=layers.add([x, residual]) previous_block_activation=x # Add a per-pixel classification layer outputs=layers.Conv2D(num_classes,3, activation="softmax", padding="same")(x) model=keras.Model(inputs, outputs) return model # Build model model = get_model(img_size, num_classes) # model=ComputeSumModel(get_model(img_size, num_classes)) model.summary() import random # Split our img paths into a training and a validation set val_samples=1000 random.Random(1337).shuffle(input_img_paths) random.Random(1337).shuffle(target_img_paths) train_input_img_paths=input_img_paths[:-val_samples] train_target_img_paths=target_img_paths[:-val_samples] val_input_img_paths=input_img_paths[-val_samples:] val_target_img_paths=target_img_paths[-val_samples:] # Instantiate dataset for each split # Limit input files in 'max_dataset_len' for faster epoch training time # Remove the 'max_dataset_len' arg when running with full dataset train_dataset=get_dataset( batch_size, img_size, train_input_img_paths, train_target_img_paths, max_dataset_len=1000, ) valid_dataset=get_dataset( batch_size, img_size, val_input_img_paths, val_target_img_paths ) # Configure the model for training. # We use the "sparse" version of categorical_crossentropy # because our target data is integers model.compile( optimizer=keras.optimizers.Adam(1e-4), loss="sparse_categorical_crossentropy" ) callbacks=[ keras.callbacks.ModelCheckpoint("oxford_segmentation.keras", save_best_only=True) ] # Train the model, doing validation at the end of each epoch epochs=1 model.fit( train_dataset, epochs=epochs, validation_data=valid_dataset, callbacks=callbacks, verbose=2, ) # Generate predictions for all images in the validation set val_dataset=get_dataset( batch_size, img_size, val_input_img_paths, val_target_img_paths ) val_preds=model.predict(val_dataset) def display_img(i): mask=np.argmax(val_preds[i], axis=-1) mask=np.expand_dims(mask, axis=-1) img=ImageOps.autocontrast(keras.utils.array_to_img(mask)) display(img) # Display results for validation image #10 i=10 # Display input image display(Image(filename=val_input_img_paths[i])) # Display ground-truth target mask img=ImageOps.autocontrast(load_img(val_target_img_paths[i])) display(img) # Display mask predicted by our model display_img(i)