{"cells":[{"metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","collapsed":true,"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","trusted":false},"cell_type":"markdown","source":"# Keras MobileNet Benchmark\n\nIn a previous benchmark we used a simple three layer ConvNet. This time we use a deeper MobileNet architecture on greyscale strokes. \n\nThis kernel has two main components:\n\n* MobileNet\n* Fast and memory efficient Image Generator with temporal colored strokes\n\nI did some paramer search but it should not be hard to improve the current score."},{"metadata":{"_uuid":"f7f2a9516140a84124bf7bbf538ee4c30860b778"},"cell_type":"markdown","source":"## Setup\nImport the necessary libraries and a few helper functions."},{"metadata":{"trusted":true,"_uuid":"ce6d2aa7de1fa341144def7d3a5b1ffdea26bc91","_kg_hide-input":true},"cell_type":"code","source":"%matplotlib inline\nfrom IPython.core.interactiveshell import InteractiveShell\nInteractiveShell.ast_node_interactivity = \"all\"\nimport os\nimport ast\nimport datetime as dt\nimport matplotlib.pyplot as plt\nplt.rcParams['figure.figsize'] = [16, 10]\nplt.rcParams['font.size'] = 14\nimport seaborn as sns\nimport cv2\nimport pandas as pd\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D\nfrom tensorflow.keras.layers import Dense, Dropout, Flatten, Activation\nfrom tensorflow.keras.metrics import categorical_accuracy, top_k_categorical_accuracy, categorical_crossentropy\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau, ModelCheckpoint\nfrom tensorflow.keras.optimizers import Adam\nfrom tensorflow.keras.applications import MobileNet\nfrom tensorflow.keras.applications.mobilenet import preprocess_input\nstart = dt.datetime.now()","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true,"_uuid":"978b1e827e598c53df3ef09838a6d85591d83052"},"cell_type":"code","source":"DP_DIR = '../input/shuffle-animal-csvs/'\nINPUT_DIR = '../input/quickdraw-doodle-recognition/'\n\nBASE_SIZE = 256\nNCSVS = 100\nnp.random.seed(seed=1987)\ntf.set_random_seed(seed=1987)\n\ndef f2cat(filename: str) -> str:\n    return filename.split('.')[0]\n\ndef list_all_categories():\n    files = os.listdir(os.path.join(INPUT_DIR, 'train_simplified'))\n    return sorted([f2cat(f) for f in files], key=str.lower)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"656c4f3454f799bfeb2973e6cf85c0504cc5c802"},"cell_type":"code","source":"animals = ['ant', 'bat', 'bear', 'bee', 'bird', 'butterfly', 'camel', 'cat', 'cow',\n           'crab', 'crocodile', 'dog', 'dolphin', 'dragon', 'duck', 'elephant', 'fish',\n           'flamingo', 'frog', 'giraffe', 'hedgehog', 'horse', 'kangaroo', 'lion',\n           'lobster', 'monkey', 'mosquito', 'mouse', 'octopus', 'owl', 'panda',\n           'parrot', 'penguin', 'pig', 'rabbit', 'raccoon', 'rhinoceros', 'scorpion',\n           'sea turtle', 'shark', 'sheep', 'snail', 'snake', 'spider', 'squirrel',\n           'swan', 'teddy-bear', 'tiger', 'whale', 'zebra']\nNCATS = len(animals)\nprint('We have {} animals'.format(NCATS))","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true,"_uuid":"b2fcd1a08ae1ae0619be38a113a244eb6515b63b"},"cell_type":"code","source":"def apk(actual, predicted, k=3):\n    \"\"\"\n    Source: https://github.com/benhamner/Metrics/blob/master/Python/ml_metrics/average_precision.py\n    \"\"\"\n    if len(predicted) > k:\n        predicted = predicted[:k]\n    score = 0.0\n    num_hits = 0.0\n    for i, p in enumerate(predicted):\n        if p in actual and p not in predicted[:i]:\n            num_hits += 1.0\n            score += num_hits / (i + 1.0)\n    if not actual:\n        return 0.0\n    return score / min(len(actual), k)\n\ndef mapk(actual, predicted, k=3):\n    \"\"\"\n    Source: https://github.com/benhamner/Metrics/blob/master/Python/ml_metrics/average_precision.py\n    \"\"\"\n    return np.mean([apk(a, p, k) for a, p in zip(actual, predicted)])\n\ndef preds2catids(predictions):\n    return pd.DataFrame(np.argsort(-predictions, axis=1)[:, :3], columns=['a', 'b', 'c'])\n\ndef top_3_accuracy(y_true, y_pred):\n    return top_k_categorical_accuracy(y_true, y_pred, k=3)","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"264156422a95e4b350886d558d516ae8bd2e25c0"},"cell_type":"markdown","source":"## MobileNet\n\nMobileNets are based on a streamlined architecture that uses depthwise separable convolutions to build light weight deep neural networks.\n\n[MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/pdf/1704.04861.pdf)"},{"metadata":{"trusted":true,"_uuid":"54e5f0c637195b6624e2f3e6db5e7f8990e14eb7"},"cell_type":"code","source":"size = 64\nbatchsize = 512","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"scrolled":false,"_uuid":"0860ec35bee03f0c5cd21202dc7471c2d201cf5f","_kg_hide-output":true},"cell_type":"code","source":"model = MobileNet(input_shape=(size, size, 1), alpha=1., weights=None, classes=NCATS)\nmodel.compile(optimizer=Adam(lr=0.002), loss='categorical_crossentropy',\n              metrics=[categorical_crossentropy, categorical_accuracy, top_3_accuracy])\nprint(model.summary())","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"ab1834ea2757a53d602a3508efffcc34bc190dc7"},"cell_type":"markdown","source":"## Training with Image Generator"},{"metadata":{"trusted":true,"_uuid":"f6455bf9555b8381b6a4292098a64a0eb7ff54dc"},"cell_type":"code","source":"def draw_cv2(raw_strokes, size=256, lw=6, time_color=True):\n    img = np.zeros((BASE_SIZE, BASE_SIZE), np.uint8)\n    for t, stroke in enumerate(raw_strokes):\n        for i in range(len(stroke[0]) - 1):\n            color = 255 - min(t, 10) * 13 if time_color else 255\n            _ = cv2.line(img, (stroke[0][i], stroke[1][i]),\n                         (stroke[0][i + 1], stroke[1][i + 1]), color, lw)\n    if size != BASE_SIZE:\n        return cv2.resize(img, (size, size))\n    else:\n        return img\n\ndef image_generator_xd(size, batchsize, ks, lw=6, time_color=True):\n    while True:\n        for k in np.random.permutation(ks):\n            filename = os.path.join(DP_DIR, 'train_k{}.csv.gz'.format(k))\n            for df in pd.read_csv(filename, chunksize=batchsize):\n                df['drawing'] = df['drawing'].apply(ast.literal_eval)\n                x = np.zeros((len(df), size, size, 1))\n                for i, raw_strokes in enumerate(df.drawing.values):\n                    x[i, :, :, 0] = draw_cv2(raw_strokes, size=size, lw=lw,\n                                             time_color=time_color)\n                x = preprocess_input(x).astype(np.float32)\n                y = keras.utils.to_categorical(df.y, num_classes=NCATS)\n                yield x, y\n\ndef df_to_image_array_xd(df, size, lw=6, time_color=True):\n    df['drawing'] = df['drawing'].apply(ast.literal_eval)\n    x = np.zeros((len(df), size, size, 1))\n    for i, raw_strokes in enumerate(df.drawing.values):\n        x[i, :, :, 0] = draw_cv2(raw_strokes, size=size, lw=lw, time_color=time_color)\n    x = preprocess_input(x).astype(np.float32)\n    return x","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"98ff512e1a1b5e86e86d9eef4127525bedf3b9e1"},"cell_type":"code","source":"valid_df = pd.read_csv(os.path.join(DP_DIR, 'train_k{}.csv.gz'.format(NCSVS - 1)))\nx_valid = df_to_image_array_xd(valid_df, size)\ny_valid = keras.utils.to_categorical(valid_df.y, num_classes=NCATS)\nprint(x_valid.shape, y_valid.shape)\nprint('Validation array memory {:.2f} GB'.format(x_valid.nbytes / 1024.**3 ))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"d80ad7f4d378ea7f30479221d604eeeed559cae4"},"cell_type":"code","source":"train_datagen = image_generator_xd(size=size, batchsize=batchsize, ks=range(NCSVS - 1))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"9ce5fb89fbb77777316d6fca7689b6636c0e6021"},"cell_type":"code","source":"x, y = next(train_datagen)\nn = 8\nfig, axs = plt.subplots(nrows=n, ncols=n, sharex=True, sharey=True, figsize=(12, 12))\nfor i in range(n**2):\n    ax = axs[i // n, i % n]\n    (-x[i]+1)/2\n    ax.imshow((-x[i, :, :, 0] + 1)/2, cmap=plt.cm.gray)\n    ax.axis('off')\nplt.tight_layout()\nfig.savefig('gs.png', dpi=300)\nplt.show();","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"f9be146924d48eacf919db87f862e4b673fd591a"},"cell_type":"code","source":"%%timeit\nx, y = next(train_datagen)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"da72d70fc1781e80427d45a80c07b3571dda0b36"},"cell_type":"code","source":"callbacks = [\n    EarlyStopping(monitor='val_top_3_accuracy', patience=7, min_delta=0.001, mode='max'),\n    ReduceLROnPlateau(monitor='val_top_3_accuracy', factor=0.5, patience=5, min_delta=0.005,\n                      mode='max', cooldown=3),\n    ModelCheckpoint('gs_animal_mobile.h5', monitor='val_top_3_accuracy', mode='max',\n                    save_best_only=True, save_weights_only=True),\n]\nhists = []\nhist = model.fit_generator(\n    train_datagen, steps_per_epoch=500, epochs=50, verbose=1,\n    validation_data=(x_valid, y_valid),\n    callbacks = callbacks\n)\nhists.append(hist)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"05767778d356bc63b7cded355159fd4082eee1a5","_kg_hide-input":true},"cell_type":"code","source":"hist_df = pd.DataFrame(hist.history)\nhist_df.to_csv('gs_mobile_history.csv', index=False)\nfig, axs = plt.subplots(nrows=2, sharex=True, figsize=(16, 10))\naxs[0].plot(hist_df.val_categorical_accuracy, lw=5, label='Validation Accuracy')\naxs[0].plot(hist_df.categorical_accuracy, lw=5, label='Training Accuracy')\naxs[0].set_ylabel('Accuracy')\naxs[0].set_xlabel('Epoch')\naxs[0].grid()\naxs[0].legend(loc=0)\naxs[1].plot(hist_df.val_categorical_crossentropy, lw=5, label='Validation MLogLoss')\naxs[1].plot(hist_df.categorical_crossentropy, lw=5, label='Training MLogLoss')\naxs[1].set_ylabel('MLogLoss')\naxs[1].set_xlabel('Epoch')\naxs[1].grid()\naxs[1].legend(loc=0)\nfig.savefig('hist.png', dpi=300)\nplt.show();","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"8c1927f22d3c45cba0bdee7d6f4b6c858d82d614"},"cell_type":"code","source":"valid_predictions = model.predict(x_valid, batch_size=128, verbose=1)\nmap3 = mapk(valid_df[['y']].values, preds2catids(valid_predictions).values)\nprint('Map3: {:.3f}'.format(map3))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"b418f4c06c4e4453aa1b5ab16dde344eb8b735c5"},"cell_type":"code","source":"end = dt.datetime.now()\nprint('Latest run {}.\\nTotal time {}s'.format(end, (end - start).seconds))","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"name":"python","version":"3.6.6","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat":4,"nbformat_minor":1}