{ "cells": [ { "cell_type": "markdown", "metadata": { "_cell_guid": "d02af827-edda-3054-375f-c3fbfbbe53f6" }, "source": [ "# Development notebook\n", "\n", "- data generator for Keras\n", "- random submission" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "ebc9ee4d-ccdb-f049-1ac8-8bc4488ced1a" }, "outputs": [], "source": [ "import os \n", "from glob import glob\n", "import numpy as np # linear algebra\n", "import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n", "import cv2\n", "\n", "import matplotlib.pyplot as plt\n", "\n", "from keras.backend import set_image_dim_ordering, image_dim_ordering\n", "\n", "set_image_dim_ordering('th')\n", "print(\"Image dim ordering : \", image_dim_ordering())" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "a1461181-304b-4fe8-112c-197dab930a6e" }, "outputs": [], "source": [ "TRAIN_DATA = \"../input/train\"\n", "TEST_DATA = \"../input/test\"\n", "ADDITIONAL_DATA = \"../input/additional\"\n", "\n", "\n", "type_1_files = glob(os.path.join(TRAIN_DATA, \"Type_1\", \"*.jpg\"))\n", "type_1_ids = np.array([s[len(os.path.join(TRAIN_DATA, \"Type_1\"))+1:-4] for s in type_1_files])\n", "type_2_files = glob(os.path.join(TRAIN_DATA, \"Type_2\", \"*.jpg\"))\n", "type_2_ids = np.array([s[len(os.path.join(TRAIN_DATA, \"Type_2\"))+1:-4] for s in type_2_files])\n", "type_3_files = glob(os.path.join(TRAIN_DATA, \"Type_3\", \"*.jpg\"))\n", "type_3_ids = np.array([s[len(os.path.join(TRAIN_DATA, \"Type_3\"))+1:-4] for s in type_3_files])\n", "\n", "additional_type_1_files = glob(os.path.join(ADDITIONAL_DATA, \"Type_1\", \"*.jpg\"))\n", "additional_type_1_ids = np.array([s[len(os.path.join(ADDITIONAL_DATA, \"Type_1\"))+1:-4] for s in additional_type_1_files])\n", "additional_type_2_files = glob(os.path.join(ADDITIONAL_DATA, \"Type_2\", \"*.jpg\"))\n", "additional_type_2_ids = np.array([s[len(os.path.join(ADDITIONAL_DATA, \"Type_2\"))+1:-4] for s in additional_type_2_files])\n", "additional_type_3_files = glob(os.path.join(ADDITIONAL_DATA, \"Type_3\", \"*.jpg\"))\n", "additional_type_3_ids = np.array([s[len(os.path.join(ADDITIONAL_DATA, \"Type_3\"))+1:-4] for s in additional_type_3_files])\n", "\n", "test_files = glob(os.path.join(TEST_DATA, \"*.jpg\"))\n", "test_ids = np.array([s[len(TEST_DATA)+1:-4] for s in test_files])" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "a0f5eddf-8cb8-6a0d-667a-efd618e030d3" }, "outputs": [], "source": [ "def get_filename(image_id, image_type):\n", " \"\"\"\n", " Method to get image file path from its id and type \n", " \"\"\"\n", " if image_type == \"Type_1\" or \\\n", " image_type == \"Type_2\" or \\\n", " image_type == \"Type_3\":\n", " data_path = os.path.join(TRAIN_DATA, image_type)\n", " elif image_type == \"Test\":\n", " data_path = TEST_DATA\n", " elif image_type == \"AType_1\" or \\\n", " image_type == \"AType_2\" or \\\n", " image_type == \"AType_3\":\n", " data_path = os.path.join(ADDITIONAL_DATA, image_type)\n", " else:\n", " raise Exception(\"Image type '%s' is not recognized\" % image_type)\n", "\n", " ext = 'jpg'\n", " return os.path.join(data_path, \"{}.{}\".format(image_id, ext))\n", "\n", "\n", "def get_image_data(image_id, image_type):\n", " \"\"\"\n", " Method to get image data as np.array specifying image id and type\n", " \"\"\"\n", " fname = get_filename(image_id, image_type)\n", " img = cv2.imread(fname)\n", " assert img is not None, \"Failed to read image : %s, %s\" % (image_id, image_type)\n", " img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n", " return img" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "6cf918d9-10a4-0076-d52c-55f78660b202" }, "outputs": [], "source": [ "type_to_index = {\n", " \"Type_1\": 0,\n", " \"Type_2\": 1,\n", " \"Type_3\": 2,\n", "}\n", "\n", "\n", "def data_iterator(image_id_type_list, batch_size, image_size, verbose=0, test_mode=False):\n", " \n", " while True:\n", " X = np.zeros((batch_size, 3) + image_size, dtype=np.float32)\n", " Y = np.zeros((batch_size, 3), dtype=np.uint8)\n", " image_ids = np.empty((batch_size,), dtype=np.object)\n", " counter = 0\n", " for i, (image_id, image_type) in enumerate(image_id_type_list):\n", " \n", " img = get_image_data(image_id, image_type)\n", " img = cv2.resize(img, dsize=image_size[::-1])\n", " img = img.transpose([2,0,1])\n", " img = img.astype(np.float32) / 255.0\n", " \n", " X[counter, :, :, :] = img \n", " if test_mode:\n", " image_ids[counter] = image_id\n", " else:\n", " Y[counter, type_to_index[image_type]] = 1 \n", " \n", " if verbose > 0:\n", " print(\"Image id/type:\", image_id, image_type)\n", " \n", " counter += 1\n", " if counter == batch_size:\n", " yield (X, Y) if not test_mode else (X, Y, image_ids)\n", " X = np.zeros((batch_size, 3) + image_size, dtype=np.float32)\n", " Y = np.zeros((batch_size, 3), dtype=np.uint8)\n", " image_ids = np.empty((batch_size,), dtype=np.object)\n", " counter = 0\n", " \n", " if counter > 0:\n", " X = X[:counter,:,:,:]\n", " Y = Y[:counter,:]\n", " image_ids = image_ids[:counter]\n", " yield (X, Y) if not test_mode else (X, Y, image_ids)\n", " \n", " if test_mode:\n", " break" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "3090c8b3-f1d9-30b4-85d2-a89ecb5a5c78" }, "outputs": [], "source": [ "val_split=0.3\n", "type_ids = [type_1_ids, type_2_ids, type_3_ids]\n", "image_types = [\"Type_1\", \"Type_2\", \"Type_3\"]\n", "train_ll = [int(len(ids) * (1.0 - val_split)) for ids in type_ids]\n", "val_ll = [int(len(ids) * (val_split)) for ids in type_ids]\n", "\n", "\n", "count = 0\n", "train_id_type_list = []\n", "train_ids = [ids[:l] for ids, l in zip(type_ids, train_ll)]\n", "max_size = max(train_ll)\n", "while count < max_size: \n", " for l, ids, image_type in zip(train_ll, train_ids, image_types): \n", " image_id = ids[count % l]\n", " train_id_type_list.append((image_id, image_type))\n", " count += 1\n", " \n", "\n", "count = 0\n", "val_id_type_list = []\n", "val_ids = [ids[tl:tl+vl] for ids, tl, vl in zip(type_ids, train_ll, val_ll)]\n", "max_size = max(val_ll)\n", "while count < max_size: \n", " for l, ids, image_type in zip(val_ll, val_ids, image_types): \n", " image_id = ids[count % l]\n", " val_id_type_list.append((image_id, image_type))\n", " count += 1\n", "\n", "assert len(set(train_id_type_list) & set(val_id_type_list)) == 0, \"WTF\" \n", "\n", " \n", "print(\"Train dataset contains : \")\n", "print(\"-\", train_ll, \" images of corresponding types\")\n", "print(\"Validation dataset contains : \")\n", "print(\"-\", val_ll, \" images of corresponding types\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "696d4b4f-17bc-ce9d-4f46-cbb81fd12e60" }, "outputs": [], "source": [ "image_size = (224, 224)\n", "batch_size = 15\n", "train_iter = data_iterator(train_id_type_list, batch_size=batch_size, image_size=image_size, verbose=1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "1202960d-1113-85a2-8211-142c8e020cc0" }, "outputs": [], "source": [ "for X, Y in train_iter:\n", " print(X.shape, X.dtype, Y.shape)\n", " n = 5\n", " for counter in range(batch_size):\n", " if counter % n == 0:\n", " plt.figure(figsize=(12, 4))\n", " plt.subplot(1, n, counter % n + 1)\n", " plt.imshow(X[counter, :, :, :].transpose([1, 2, 0]))\n", " plt.title(\"Type : {}\".format(Y[counter,:]))\n", " plt.axis('off')\n", " \n", " break" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "a24022e8-a665-e461-7fe3-e265ff75fdb7" }, "outputs": [], "source": [ "image_size = (224, 224)\n", "batch_size = 15\n", "val_iter = data_iterator(val_id_type_list, batch_size=batch_size, image_size=image_size, verbose=1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "8a643e65-083f-1f60-f6ed-38881193a236" }, "outputs": [], "source": [ "for X, Y in val_iter:\n", " print(X.shape, X.dtype, Y.shape)\n", " n = 5\n", " for counter in range(batch_size):\n", " if counter % n == 0:\n", " plt.figure(figsize=(12, 4))\n", " plt.subplot(1, n, counter % n + 1)\n", " plt.imshow(X[counter, :, :, :].transpose([1, 2, 0]))\n", " plt.title(\"Type : {}\".format(Y[counter,:]))\n", " plt.axis('off')\n", " \n", " break" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "969725f8-a841-152f-6a3c-a5e8a6ed9e02" }, "outputs": [], "source": "" }, { "cell_type": "markdown", "metadata": { "_cell_guid": "248665e2-fb26-4052-7c78-eec07199d44c" }, "source": [ "### Try ResNet from Keras applications" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "829aa560-ad9b-c46d-5c9f-16a399dea1c1" }, "outputs": [], "source": [ "from keras.applications.resnet50 import ResNet50\n", "from keras.layers import Dense, Flatten\n", "from keras.models import Model\n", "\n", "image_size = (224, 224)\n", "\n", "base_model = ResNet50(include_top=False, weights=None, input_tensor=None, input_shape=(3,) + image_size)\n", "x = Flatten()(base_model.output)\n", "output = Dense(3, activation='softmax')(x)\n", "model = Model(inputs=base_model.input, outputs=output)\n", "model.summary()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "d1a0a8e3-5884-dec2-5850-d58a36d92e13" }, "outputs": [], "source": [ "model.compile(loss='categorical_crossentropy', optimizer='adadelta', metrics=['accuracy',])" ] }, { "cell_type": "markdown", "metadata": { "_cell_guid": "91e407e4-79a5-8b65-76c3-e84625d228f9" }, "source": [ "### Random predictions" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "fb5f6ad0-4f20-d665-1962-ec9df3e72de0" }, "outputs": [], "source": [ "def logloss_mc(y_true, y_prob, epsilon=1e-15):\n", " \"\"\" Multiclass logloss\n", " This function is not officially provided by Kaggle, so there is no\n", " guarantee for its correctness.\n", " https://github.com/ottogroup/kaggle/blob/master/benchmark.py\n", " \"\"\"\n", " # normalize\n", " y_prob = y_prob / y_prob.sum(axis=1).reshape(-1, 1)\n", " y_prob = np.maximum(epsilon, y_prob)\n", " y_prob = np.minimum(1 - epsilon, y_prob)\n", " # get probabilities\n", " y = [y_prob[i, j] for (i, j) in enumerate(y_true)]\n", " ll = - np.mean(np.log(y))\n", " return ll" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "ad86c7b9-085a-938c-07c0-3480eff2650a" }, "outputs": [], "source": [ "batch_size = 16\n", "val_iter = data_iterator(val_id_type_list, batch_size=batch_size, image_size=image_size, test_mode=True)\n", "\n", "total_loss = 0.0\n", "total_counter = 0 \n", "for X, Y_true, _ in val_iter: \n", " s = Y_true.shape[0]\n", " total_counter += s\n", " #Y_pred = model.predict(X)\n", " Y_pred = 0.33333 * np.ones_like(Y_true)\n", " loss = logloss_mc(Y_true, Y_pred)\n", " print(\"--\", total_counter, \"batch loss : \", loss)\n", " total_loss += s * loss\n", "\n", "total_loss *= 1.0 / total_counter \n", "print(\"Total loss : \", total_loss)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "9af7139d-0f3b-7c91-e55d-5f8733c4e91d" }, "outputs": [], "source": [ "df = pd.DataFrame(columns=['image_name','Type_1','Type_2','Type_3'])\n", "def get_test_id_type_list():\n", " return [(image_id, 'Test') for image_id in test_ids]\n", "\n", "image_size = (224, 224)\n", "batch_size = 16\n", "test_id_type_list = get_test_id_type_list()\n", "test_iter = data_iterator(test_id_type_list, batch_size=batch_size, image_size=image_size, test_mode=True)\n", "\n", "\n", "df = pd.DataFrame(columns=['image_name','Type_1','Type_2','Type_3'])\n", "total_counter = 0\n", "for X, _, image_ids in test_iter: \n", " #Y_pred = model.predict(X) \n", " s = X.shape[0]\n", " total_counter += s\n", " Y_pred = 0.33333 * np.ones((s, 3))\n", " print(\"--\", total_counter, image_ids)\n", " for i in range(s):\n", " df.loc[total_counter + i, :] = (image_ids[i] + '.jpg', ) + tuple(Y_pred[i, :])" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "fe93644a-8a18-7c80-08c7-6e399fdd49be" }, "outputs": [], "source": [ "print(df.shape)\n", "df.head()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "eeefe47c-0567-7ee8-3995-751905183f74" }, "outputs": [], "source": [ "import datetime\n", "now = datetime.datetime.now()\n", "info = 'random_predictions'\n", "sub_file = 'submission_' + info + '_' + str(now.strftime(\"%Y-%m-%d-%H-%M\")) + '.csv'\n", "df.to_csv(sub_file, index=False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "a93b5287-f7f0-3aaa-6bbc-b65cb921299c" }, "outputs": [], "source": [ "!ls " ] }, { "cell_type": "markdown", "metadata": { "_cell_guid": "5b5a3c1c-7712-3a9a-e4a9-600dd88b872c" }, "source": [ "### Training phase\n", "\n", "Stops by Kaggle server due to 1200 timeout " ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "46037148-008f-e031-9732-1755e1a121c1" }, "outputs": [], "source": [ "batch_size=16\n", "samples_per_epoch = 512\n", "nb_val_samples = 64\n", "\n", "print(batch_size, samples_per_epoch, nb_val_samples)\n", "\n", "train_iter = data_iterator(train_id_type_list, batch_size=batch_size, image_size=image_size)\n", "val_iter = data_iterator(val_id_type_list, batch_size=batch_size, image_size=image_size)\n", "\n", "#history = model.fit_generator(\n", "# train_iter,\n", "# steps_per_epoch=samples_per_epoch, \n", "# epochs=1,\n", "# validation_data=val_iter,\n", "# validation_steps=nb_val_samples,\n", "# verbose=1\n", "#)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "6b4eb433-3a88-c08e-f717-eb1c3001b758" }, "outputs": [], "source": [ "type_3_ids" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "_cell_guid": "24ac2241-1168-4b34-81e1-3e978404778b" }, "outputs": [], "source": "" } ], "metadata": { "_change_revision": 0, "_is_fork": false, "kernelspec": { "display_name": "Python 3", "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": "3.6.0" } }, "nbformat": 4, "nbformat_minor": 0 }