kernels/a-keras-prototype-0-21174-on-pl.ipynb

{"cells": [{"execution_count": null, "outputs": [], "source": ["import numpy as np # linear algebra\n", "np.random.seed(666)\n", "import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n", "from sklearn.model_selection import train_test_split\n", "from subprocess import check_output\n", "print(check_output([\"ls\", \"../input\"]).decode(\"utf8\"))"], "cell_type": "code", "metadata": {"_cell_guid": "003a34b4-ea58-413c-8ee5-dda6c041ded7", "_uuid": "14112d83eff81bcfed3d8b0fdf8b4fc37af7d92b"}}, {"execution_count": null, "outputs": [], "source": ["#Load data\n", "train = pd.read_json(\"../input/train.json\")\n", "test = pd.read_json(\"../input/test.json\")\n", "train.inc_angle = train.inc_angle.replace('na', 0)\n", "train.inc_angle = train.inc_angle.astype(float).fillna(0.0)\n", "print(\"done!\")"], "cell_type": "code", "metadata": {"_cell_guid": "29a5f133-81c7-486a-82c7-5e643b648e22", "_uuid": "4b01d4e424cec7956507f5ac8179049bcd2f0235"}}, {"execution_count": null, "outputs": [], "source": ["# Train data\n", "x_band1 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in train[\"band_1\"]])\n", "x_band2 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in train[\"band_2\"]])\n", "X_train = np.concatenate([x_band1[:, :, :, np.newaxis]\n", "                          , x_band2[:, :, :, np.newaxis]\n", "                         , ((x_band1+x_band1)/2)[:, :, :, np.newaxis]], axis=-1)\n", "X_angle_train = np.array(train.inc_angle)\n", "y_train = np.array(train[\"is_iceberg\"])\n", "\n", "# Test data\n", "x_band1 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in test[\"band_1\"]])\n", "x_band2 = np.array([np.array(band).astype(np.float32).reshape(75, 75) for band in test[\"band_2\"]])\n", "X_test = np.concatenate([x_band1[:, :, :, np.newaxis]\n", "                          , x_band2[:, :, :, np.newaxis]\n", "                         , ((x_band1+x_band1)/2)[:, :, :, np.newaxis]], axis=-1)\n", "X_angle_test = np.array(test.inc_angle)\n", "\n", "\n", "X_train, X_valid, X_angle_train, X_angle_valid, y_train, y_valid = train_test_split(X_train\n", "                    , X_angle_train, y_train, random_state=123, train_size=0.75)"], "cell_type": "code", "metadata": {"_cell_guid": "2a419a87-0630-4794-8e5b-dff560d2d116", "_uuid": "b2eb784dbc8a40a145c5cdad2f6ca929e165ce95"}}, {"execution_count": null, "outputs": [], "source": ["X_train.shape"], "cell_type": "code", "metadata": {}}, {"execution_count": null, "outputs": [], "source": ["from matplotlib import pyplot\n", "from keras.preprocessing.image import ImageDataGenerator\n", "from keras.models import Sequential\n", "from keras.layers import Conv2D, MaxPooling2D, Dense, Dropout, Input, Flatten\n", "from keras.layers import GlobalMaxPooling2D\n", "from keras.layers.normalization import BatchNormalization\n", "from keras.layers.merge import Concatenate\n", "from keras.models import Model\n", "from keras.optimizers import Adam\n", "from keras.callbacks import ModelCheckpoint, Callback, EarlyStopping\n", "\n", "def get_callbacks(filepath, patience=2):\n", "    es = EarlyStopping('val_loss', patience=patience, mode=\"min\")\n", "    msave = ModelCheckpoint(filepath, save_best_only=True)\n", "    return [es, msave]\n", "    \n", "def get_model():\n", "    bn_model = 0\n", "    p_activation = \"elu\"\n", "    input_1 = Input(shape=(75, 75, 3), name=\"X_1\")\n", "    input_2 = Input(shape=[1], name=\"angle\")\n", "    \n", "    img_1 = Conv2D(16, kernel_size = (3,3), activation=p_activation) ((BatchNormalization(momentum=bn_model))(input_1))\n", "    img_1 = Conv2D(16, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = MaxPooling2D((2,2)) (img_1)\n", "    img_1 = Dropout(0.2)(img_1)\n", "    img_1 = Conv2D(32, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = Conv2D(32, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = MaxPooling2D((2,2)) (img_1)\n", "    img_1 = Dropout(0.2)(img_1)\n", "    img_1 = Conv2D(64, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = Conv2D(64, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = MaxPooling2D((2,2)) (img_1)\n", "    img_1 = Dropout(0.2)(img_1)\n", "    img_1 = Conv2D(128, kernel_size = (3,3), activation=p_activation) (img_1)\n", "    img_1 = MaxPooling2D((2,2)) (img_1)\n", "    img_1 = Dropout(0.2)(img_1)\n", "    img_1 = GlobalMaxPooling2D() (img_1)\n", "    \n", "    \n", "    img_2 = Conv2D(128, kernel_size = (3,3), activation=p_activation) ((BatchNormalization(momentum=bn_model))(input_1))\n", "    img_2 = MaxPooling2D((2,2)) (img_2)\n", "    img_2 = Dropout(0.2)(img_2)\n", "    img_2 = GlobalMaxPooling2D() (img_2)\n", "    \n", "    img_concat =  (Concatenate()([img_1, img_2, BatchNormalization(momentum=bn_model)(input_2)]))\n", "    \n", "    dense_ayer = Dropout(0.5) (BatchNormalization(momentum=bn_model) ( Dense(256, activation=p_activation)(img_concat) ))\n", "    dense_ayer = Dropout(0.5) (BatchNormalization(momentum=bn_model) ( Dense(64, activation=p_activation)(dense_ayer) ))\n", "    output = Dense(1, activation=\"sigmoid\")(dense_ayer)\n", "    \n", "    model = Model([input_1,input_2],  output)\n", "    optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)\n", "    model.compile(loss=\"binary_crossentropy\", optimizer=optimizer, metrics=[\"accuracy\"])\n", "    return model\n", "model = get_model()\n", "model.summary()"], "cell_type": "code", "metadata": {"_cell_guid": "48694991-a0af-4e4a-86c9-9e6a106ac21d", "collapsed": true, "_uuid": "d5787df36f2640ba2ae57dad49dc78b7c2a8f8b3"}}, {"execution_count": null, "outputs": [], "source": ["file_path = \".model_weights.hdf5\"\n", "callbacks = get_callbacks(filepath=file_path, patience=5)\n", "\n", "model = get_model()\n", "model.fit([X_train, X_angle_train], y_train, epochs=25\n", "          , validation_data=([X_valid, X_angle_valid], y_valid)\n", "         , batch_size=32\n", "         , callbacks=callbacks)"], "cell_type": "code", "metadata": {"_cell_guid": "ed3d20e2-9cd3-4aad-9664-c7f3c485cc1e", "collapsed": true, "_uuid": "7397f8e25535660a8149786218f0cc54070ddb7b"}}, {"execution_count": null, "outputs": [], "source": ["model.load_weights(filepath=file_path)\n", "\n", "print(\"Train evaluate:\")\n", "print(model.evaluate([X_train, X_angle_train], y_train, verbose=1, batch_size=200))\n", "print(\"####################\")\n", "print(\"watch list evaluate:\")\n", "print(model.evaluate([X_valid, X_angle_valid], y_valid, verbose=1, batch_size=200))"], "cell_type": "code", "metadata": {"_cell_guid": "aacdccc7-760d-4433-8a6f-554d357722fd", "collapsed": true, "_uuid": "f16cda9020fb9e07cb3325ba05061b786e603626"}}, {"execution_count": null, "outputs": [], "source": ["prediction = model.predict([X_test, X_angle_test], verbose=1, batch_size=200)"], "cell_type": "code", "metadata": {"_cell_guid": "78dbcf6c-a7ba-49c7-a459-20bed13a2219", "collapsed": true, "_uuid": "df1410c2fb4d4f124ba89602b3f669de7018728a"}}, {"execution_count": null, "outputs": [], "source": ["submission = pd.DataFrame({'id': test[\"id\"], 'is_iceberg': prediction.reshape((prediction.shape[0]))})\n", "submission.head(10)"], "cell_type": "code", "metadata": {"_cell_guid": "52766228-cf82-4cc8-9671-d8391d294bbb", "collapsed": true, "_uuid": "a656b27ff563fd4e69c43393280b9227e44aac74"}}, {"execution_count": null, "outputs": [], "source": ["submission.to_csv(\"./submission.csv\", index=False)"], "cell_type": "code", "metadata": {"_cell_guid": "da4a9987-2589-48b3-bfd1-5e2076f74f0c", "collapsed": true, "_uuid": "53646935d88ffa854e33aeb583c6d5c475e9380b"}}, {"metadata": {"_cell_guid": "2732f6a8-745c-49df-898b-8a55d22b0227", "collapsed": true, "_uuid": "321d861514ebae2d61817486dc3508dd350a047b"}, "source": ["Any comment will be welcome! ;)"], "cell_type": "markdown"}], "metadata": {"language_info": {"nbconvert_exporter": "python", "name": "python", "pygments_lexer": "ipython3", "file_extension": ".py", "mimetype": "text/x-python", "codemirror_mode": {"name": "ipython", "version": 3}, "version": "3.6.3"}, "kernelspec": {"name": "python3", "display_name": "Python 3", "language": "python"}}, "nbformat_minor": 1, "nbformat": 4}