{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","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"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":10200,"databundleVersionId":868375,"sourceType":"competition"},{"sourceId":7371743,"sourceType":"kernelVersion"}],"dockerImageVersionId":25160,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"#CNN usando Keras ¡Desde cero!\n \n* Este es un problema de reconocimiento/clasificación de imágenes complejas dibujadas por personas.\n* Para ello se utilizó una CNN especializada en reconocimiento de imágenes.\n* En lugar de una red preentrenada, el modelo básico se creó utilizando Keras.\n* Para el preprocesamiento de datos y la entrada/salida de datos, consulte el kernel ['Image-Based CNN'] (https://www.kaggle.com/jpmiller/image-based-cnn).\n* Debido a limitaciones en el preprocesamiento de datos, se utilizaron 2000 conjuntos de datos por clase. Sería bueno desarrollar un modelo utilizando un generador para obtener más datos.\n* En el caso del modelo CNN, se utilizaron 3 capas de convolución-Pooling.\n​","metadata":{}},{"cell_type":"code","source":"%matplotlib inline\nfrom IPython.core.interactiveshell import InteractiveShell\nInteractiveShell.ast_node_interactivity = \"all\"\nimport os\nimport json\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()","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2024-04-16T17:14:21.529805Z","iopub.execute_input":"2024-04-16T17:14:21.530263Z","iopub.status.idle":"2024-04-16T17:14:21.542025Z","shell.execute_reply.started":"2024-04-16T17:14:21.53019Z","shell.execute_reply":"2024-04-16T17:14:21.54128Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"","metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a"}},{"cell_type":"code","source":"DP_DIR = '../input/shuffle-csvs/'\nINPUT_DIR = '../input/quickdraw-doodle-recognition/'\n\nBASE_SIZE = 256\nNCSVS = 100\nNCATS = 340\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)","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:21.543311Z","iopub.execute_input":"2024-04-16T17:14:21.543584Z","iopub.status.idle":"2024-04-16T17:14:21.5578Z","shell.execute_reply.started":"2024-04-16T17:14:21.543531Z","shell.execute_reply":"2024-04-16T17:14:21.557044Z"},"trusted":true},"execution_count":null,"outputs":[]},{"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)","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:21.559304Z","iopub.execute_input":"2024-04-16T17:14:21.559527Z","iopub.status.idle":"2024-04-16T17:14:21.569287Z","shell.execute_reply.started":"2024-04-16T17:14:21.559489Z","shell.execute_reply":"2024-04-16T17:14:21.5684Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#CNN 모델 설계\n\n* INPUT -> [[CONV -> RELU]-> POOL]*3-> [FC-> RELU]-> FC\n\n* Para reconocer la imagen, se realizó la convolución tres veces para crear una estructura jerárquica para cada parte.\n* En convolución, la activación se realizó con Relu y la activación de clase final usó softmax para 340 etiquetas.","metadata":{}},{"cell_type":"code","source":"STEPS = 800\nEPOCHS = 70\nsize = 64\nbatchsize = 680","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:21.570397Z","iopub.execute_input":"2024-04-16T17:14:21.570614Z","iopub.status.idle":"2024-04-16T17:14:21.582229Z","shell.execute_reply.started":"2024-04-16T17:14:21.570578Z","shell.execute_reply":"2024-04-16T17:14:21.581256Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from tensorflow.keras.layers import BatchNormalization\n\n\nmodel =Sequential()\nmodel.add(Conv2D(32,kernel_size=(3,3),padding='same',activation='relu',input_shape=(64,64,1)))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(32,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(64,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(64,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(128,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(128,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\n\nmodel.add(Conv2D(256,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\nmodel.add(Conv2D(256,kernel_size=(3,3),padding='same',activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2,2)))\nmodel.add(BatchNormalization())\nmodel.add(Dropout(0.1))\n\n\n\n\nmodel.add(Flatten())\n\nmodel.add(Dense(4096, activation='relu'))  # Capa Dense añadida\nmodel.add(Dropout(0.1))\n\nmodel.add(Dense(512, activation='relu'))  # Capa Dense añadida\nmodel.add(Dropout(0.1))\n\nmodel.add(Dense(340, activation='softmax'))\nmodel.summary()\n\n\n\n\nmodel.compile(optimizer=Adam(lr=0.002), loss='categorical_crossentropy',\n              metrics=[categorical_crossentropy, categorical_accuracy, top_3_accuracy])","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:21.583457Z","iopub.execute_input":"2024-04-16T17:14:21.583743Z","iopub.status.idle":"2024-04-16T17:14:22.851969Z","shell.execute_reply.started":"2024-04-16T17:14:21.58369Z","shell.execute_reply":"2024-04-16T17:14:22.851285Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**IMAGE GENERATOR**","metadata":{}},{"cell_type":"code","source":"def draw_cv2(raw_strokes, size=256, lw=6, time_color=False):\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=False):\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(json.loads)\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                x = ((x / 127.5) - 1).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=False):\n    df['drawing'] = df['drawing'].apply(json.loads)\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    x = ((x / 127.5) - 1).astype(np.float32)\n    return x","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:22.853448Z","iopub.execute_input":"2024-04-16T17:14:22.853721Z","iopub.status.idle":"2024-04-16T17:14:22.865198Z","shell.execute_reply.started":"2024-04-16T17:14:22.853658Z","shell.execute_reply":"2024-04-16T17:14:22.864362Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"valid_df = pd.read_csv(os.path.join(DP_DIR, 'train_k{}.csv.gz'.format(NCSVS - 1)), nrows=34000)\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 ))","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:22.866513Z","iopub.execute_input":"2024-04-16T17:14:22.866809Z","iopub.status.idle":"2024-04-16T17:14:30.877741Z","shell.execute_reply.started":"2024-04-16T17:14:22.866758Z","shell.execute_reply":"2024-04-16T17:14:30.876863Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_datagen = image_generator_xd(size=size, batchsize=batchsize, ks=range(NCSVS - 1))\n\n#x, y = next(train_datagen)\n#n = 8\n#fig, axs = plt.subplots(nrows=n, ncols=n, sharex=True, sharey=True, figsize=(12, 12))\n#for 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')\n#plt.tight_layout()\n#fig.savefig('gs.png', dpi=300)\n#plt.show();","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:30.879581Z","iopub.execute_input":"2024-04-16T17:14:30.879924Z","iopub.status.idle":"2024-04-16T17:14:30.885081Z","shell.execute_reply.started":"2024-04-16T17:14:30.87986Z","shell.execute_reply":"2024-04-16T17:14:30.884004Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"* En el caso de la categoría de etiqueta, se utilizó 'sparse_categorical_crossentropy' porque no se utilizó la codificación One-Hot por motivos de memoria.\n* Reduzca la tasa de aprendizaje cada vez que se llama a la devolución de llamada con reduceLROnPlat.\n* val_acc se utilizó como condición para la parada anticipada.\n* El optimizador utiliza  'adam'","metadata":{}},{"cell_type":"code","source":"#%%timeit\n#x, y = next(train_datagen)\n","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:30.886078Z","iopub.execute_input":"2024-04-16T17:14:30.886424Z","iopub.status.idle":"2024-04-16T17:14:30.897033Z","shell.execute_reply.started":"2024-04-16T17:14:30.886345Z","shell.execute_reply":"2024-04-16T17:14:30.896278Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"callbacks = [\n    ReduceLROnPlateau(monitor='val_top_3_accuracy', factor=0.75, patience=3, min_delta=0.001,\n                          mode='max', min_lr=1e-5, verbose=1),\n    ModelCheckpoint('model.h5', monitor='val_top_3_accuracy', mode='max', save_best_only=True,\n                    save_weights_only=True),\n]\nhists = []\nhist = model.fit_generator(\n    train_datagen, steps_per_epoch=STEPS, epochs=100, verbose=1,\n    validation_data=(x_valid, y_valid),\n    callbacks = callbacks\n)\nhists.append(hist)","metadata":{"execution":{"iopub.status.busy":"2024-04-16T17:14:30.898691Z","iopub.execute_input":"2024-04-16T17:14:30.898934Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Loss & Accuracy Graph\n* Entrenamiento, precisión de validación y gráfico de pérdidas.\n","metadata":{}},{"cell_type":"code","source":"hist_df = pd.concat([pd.DataFrame(hist.history) for hist in hists], sort=True)\nhist_df.index = np.arange(1, len(hist_df)+1)\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();\n\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"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))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"\ntest = pd.read_csv(os.path.join(INPUT_DIR, 'test_simplified.csv'))\ntest.head()\nx_test = df_to_image_array_xd(test, size)\nprint(test.shape, x_test.shape)\nprint('Test array memory {:.2f} GB'.format(x_test.nbytes / 1024.**3 ))\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.save('keras.h5')","metadata":{"trusted":true},"execution_count":null,"outputs":[]}]}