Datasets
Models
Downloads: 1
Diff of
/AttentonUnet.ipynb
[000000]
..
[00c919]
Switch to side-by-side view
--- a +++ b/AttentonUnet.ipynb @@ -0,0 +1,236 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 3, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Error loading .DS_Store or 0655[0]_47.png: cannot identify image file <_io.BytesIO object at 0x35adee660>. Skipping...\n", + "Epoch 1/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m384s\u001b[0m 2s/step - accuracy: 0.9061 - loss: 0.2485 - val_accuracy: 0.8808 - val_loss: 0.3486\n", + "Epoch 2/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m384s\u001b[0m 2s/step - accuracy: 0.9415 - loss: 0.1394 - val_accuracy: 0.8412 - val_loss: 0.4048\n", + "Epoch 3/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m378s\u001b[0m 2s/step - accuracy: 0.9457 - loss: 0.1280 - val_accuracy: 0.8718 - val_loss: 0.4388\n", + "Epoch 4/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m385s\u001b[0m 2s/step - accuracy: 0.9491 - loss: 0.1193 - val_accuracy: 0.8620 - val_loss: 0.4341\n", + "Epoch 5/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m378s\u001b[0m 2s/step - accuracy: 0.9492 - loss: 0.1185 - val_accuracy: 0.8636 - val_loss: 0.5675\n", + "Epoch 6/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m384s\u001b[0m 2s/step - accuracy: 0.9515 - loss: 0.1134 - val_accuracy: 0.8706 - val_loss: 0.5460\n", + "Epoch 7/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m384s\u001b[0m 2s/step - accuracy: 0.9568 - loss: 0.0998 - val_accuracy: 0.8562 - val_loss: 0.6479\n", + "Epoch 8/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m382s\u001b[0m 2s/step - accuracy: 0.9572 - loss: 0.0983 - val_accuracy: 0.8637 - val_loss: 1.0583\n", + "Epoch 9/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m391s\u001b[0m 2s/step - accuracy: 0.9601 - loss: 0.0928 - val_accuracy: 0.8689 - val_loss: 0.4872\n", + "Epoch 10/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m385s\u001b[0m 2s/step - accuracy: 0.9616 - loss: 0.0885 - val_accuracy: 0.8676 - val_loss: 0.6407\n", + "Epoch 11/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m373s\u001b[0m 2s/step - accuracy: 0.9648 - loss: 0.0807 - val_accuracy: 0.8683 - val_loss: 0.6889\n", + "Epoch 12/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m377s\u001b[0m 2s/step - accuracy: 0.9663 - loss: 0.0786 - val_accuracy: 0.8550 - val_loss: 0.7435\n", + "Epoch 13/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m374s\u001b[0m 2s/step - accuracy: 0.9703 - loss: 0.0703 - val_accuracy: 0.8677 - val_loss: 0.6834\n", + "Epoch 14/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m373s\u001b[0m 2s/step - accuracy: 0.9712 - loss: 0.0665 - val_accuracy: 0.8694 - val_loss: 0.5149\n", + "Epoch 15/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m379s\u001b[0m 2s/step - accuracy: 0.9716 - loss: 0.0672 - val_accuracy: 0.8633 - val_loss: 0.7259\n", + "Epoch 16/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m395s\u001b[0m 2s/step - accuracy: 0.9748 - loss: 0.0594 - val_accuracy: 0.8736 - val_loss: 0.6896\n", + "Epoch 17/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m380s\u001b[0m 2s/step - accuracy: 0.9767 - loss: 0.0545 - val_accuracy: 0.8695 - val_loss: 0.7535\n", + "Epoch 18/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m389s\u001b[0m 2s/step - accuracy: 0.9773 - loss: 0.0532 - val_accuracy: 0.8664 - val_loss: 0.8831\n", + "Epoch 19/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m376s\u001b[0m 2s/step - accuracy: 0.9781 - loss: 0.0512 - val_accuracy: 0.8720 - val_loss: 0.7170\n", + "Epoch 20/20\n", + "\u001b[1m193/193\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m384s\u001b[0m 2s/step - accuracy: 0.9790 - loss: 0.0487 - val_accuracy: 0.8707 - val_loss: 0.6628\n" + ] + } + ], + "source": [ + "import tensorflow as tf\n", + "from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, concatenate, Activation, BatchNormalization, Add, Multiply\n", + "from tensorflow.keras.models import Model\n", + "import os\n", + "import numpy as np\n", + "from tensorflow.keras.preprocessing.image import load_img, img_to_array\n", + "\n", + "def attention_block(x, g, inter_channel):\n", + " \"\"\"\n", + " Attention Block: Refines encoder features based on decoder signals.\n", + " x: Input tensor from the encoder (skip connection)\n", + " g: Gating signal from the decoder (upsampled tensor)\n", + " inter_channel: Number of intermediate channels (reduces computation)\n", + " \"\"\"\n", + " # 1x1 Convolution on input tensor\n", + " theta_x = Conv2D(inter_channel, kernel_size=(1, 1), strides=(1, 1), padding='same')(x)\n", + " # 1x1 Convolution on gating tensor\n", + " phi_g = Conv2D(inter_channel, kernel_size=(1, 1), strides=(1, 1), padding='same')(g)\n", + " \n", + " # Add the transformed inputs and apply ReLU\n", + " add_xg = Add()([theta_x, phi_g])\n", + " relu_xg = Activation('relu')(add_xg)\n", + " \n", + " # Another 1x1 Convolution to generate attention coefficients\n", + " psi = Conv2D(1, kernel_size=(1, 1), strides=(1, 1), padding='same')(relu_xg)\n", + " # Sigmoid activation to normalize attention weights\n", + " sigmoid_psi = Activation('sigmoid')(psi)\n", + " \n", + " # Multiply the input tensor with the attention weights\n", + " return Multiply()([x, sigmoid_psi])\n", + "\n", + "def conv_block(x, filters):\n", + " \"\"\"\n", + " Convolutional Block: Apply two 3x3 convolutions followed by BatchNorm and ReLU.\n", + " x: Input tensor\n", + " filters: Number of output filters for the convolutions\n", + " \"\"\"\n", + " x = Conv2D(filters, kernel_size=(3, 3), padding='same')(x)\n", + " x = BatchNormalization()(x)\n", + " x = Activation('relu')(x)\n", + " x = Conv2D(filters, kernel_size=(3, 3), padding='same')(x)\n", + " x = BatchNormalization()(x)\n", + " x = Activation('relu')(x)\n", + " return x\n", + "\n", + "def attention_unet(input_shape, num_classes):\n", + " \"\"\"\n", + " Attention U-Net model architecture.\n", + " input_shape: Shape of input images (H, W, C)\n", + " num_classes: Number of output segmentation classes\n", + " \"\"\"\n", + " # Input layer for the images\n", + " inputs = Input(input_shape)\n", + " \n", + " # Encoder (Downsampling path)\n", + " c1 = conv_block(inputs, 64) # First Conv Block\n", + " p1 = MaxPooling2D((2, 2))(c1) # Downsample by 2\n", + " \n", + " c2 = conv_block(p1, 128) # Second Conv Block\n", + " p2 = MaxPooling2D((2, 2))(c2) # Downsample by 2\n", + " \n", + " c3 = conv_block(p2, 256) # Third Conv Block\n", + " p3 = MaxPooling2D((2, 2))(c3) # Downsample by 2\n", + " \n", + " c4 = conv_block(p3, 512) # Fourth Conv Block\n", + " p4 = MaxPooling2D((2, 2))(c4) # Downsample by 2\n", + " \n", + " # Bottleneck (lowest level of the U-Net)\n", + " c5 = conv_block(p4, 1024)\n", + " \n", + " # Decoder (Upsampling path)\n", + " up6 = UpSampling2D((2, 2))(c5) # Upsample\n", + " att6 = attention_block(c4, up6, 512) # Attention Block\n", + " merge6 = concatenate([up6, att6], axis=-1) # Concatenate features\n", + " c6 = conv_block(merge6, 512) # Conv Block after concatenation\n", + " \n", + " up7 = UpSampling2D((2, 2))(c6)\n", + " att7 = attention_block(c3, up7, 256)\n", + " merge7 = concatenate([up7, att7], axis=-1)\n", + " c7 = conv_block(merge7, 256)\n", + " \n", + " up8 = UpSampling2D((2, 2))(c7)\n", + " att8 = attention_block(c2, up8, 128)\n", + " merge8 = concatenate([up8, att8], axis=-1)\n", + " c8 = conv_block(merge8, 128)\n", + " \n", + " up9 = UpSampling2D((2, 2))(c8)\n", + " att9 = attention_block(c1, up9, 64)\n", + " merge9 = concatenate([up9, att9], axis=-1)\n", + " c9 = conv_block(merge9, 64)\n", + " \n", + " # Output layer for segmentation\n", + " outputs = Conv2D(num_classes, (1, 1), activation='softmax' if num_classes > 1 else 'sigmoid')(c9)\n", + " \n", + " # Define the model\n", + " model = Model(inputs=inputs, outputs=outputs)\n", + " return model\n", + "\n", + "# Function to load and preprocess images and masks\n", + "def load_data(image_dir, mask_dir, image_size):\n", + " \"\"\"\n", + " Load and preprocess images and masks for training.\n", + " image_dir: Path to the directory containing input images\n", + " mask_dir: Path to the directory containing segmentation masks\n", + " image_size: Tuple specifying the size (height, width) to resize the images and masks\n", + " \"\"\"\n", + " images = []\n", + " masks = []\n", + " image_files = sorted(os.listdir(image_dir))\n", + " mask_files = sorted(os.listdir(mask_dir))\n", + " \n", + " for img_file, mask_file in zip(image_files, mask_files):\n", + " try:\n", + " # Load and preprocess images\n", + " img_path = os.path.join(image_dir, img_file)\n", + " mask_path = os.path.join(mask_dir, mask_file)\n", + " \n", + " img = load_img(img_path, target_size=image_size) # Resize image\n", + " mask = load_img(mask_path, target_size=image_size, color_mode='grayscale') # Resize mask\n", + " \n", + " # Convert to numpy arrays and normalize\n", + " img = img_to_array(img) / 255.0\n", + " mask = img_to_array(mask) / 255.0\n", + " mask = np.round(mask) # Ensure masks are binary\n", + " \n", + " images.append(img)\n", + " masks.append(mask)\n", + " except Exception as e:\n", + " print(f\"Error loading {img_file} or {mask_file}: {e}. Skipping...\")\n", + " \n", + " return np.array(images), np.array(masks)\n", + "\n", + "# Example usage\n", + "if __name__ == \"__main__\":\n", + " # Load data\n", + " image_dir = \"./images/\" # Replace with your image directory\n", + " mask_dir = \"./masks/\" # Replace with your mask directory\n", + " image_size = (128, 128) # Resize all images to 128x128\n", + " images, masks = load_data(image_dir, mask_dir, image_size)\n", + " \n", + " # Define the model\n", + " model = attention_unet(input_shape=(128, 128, 3), num_classes=1)\n", + " \n", + " # Compile the model\n", + " model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])\n", + " \n", + " # Train the model\n", + " model.fit(images, masks, batch_size=8, epochs=20, validation_split=0.1)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": "base", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.12.2" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +}