{"cells":[{"metadata":{},"cell_type":"markdown","source":"Single model","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"No K-flod","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"No external data except the translated data","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"training step from this notebook https://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta","execution_count":null},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 5GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import transformers\nfrom transformers import TFAutoModel, AutoTokenizer\nfrom tqdm.notebook import tqdm\nfrom tokenizers import Tokenizer, models, pre_tokenizers, decoders, processors\nimport time\nimport os\nimport re\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\nfrom tensorflow.keras.layers import Dense, Input,GRU\nfrom tensorflow.keras.optimizers import Adam\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras import optimizers\nfrom tensorflow.keras.callbacks import ModelCheckpoint\nfrom kaggle_datasets import KaggleDatasets\nimport transformers\nfrom transformers import TFAutoModel, AutoTokenizer\nfrom tqdm.notebook import tqdm\nfrom tokenizers import Tokenizer, models, pre_tokenizers, decoders, processors","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"first , we should align the translated data so that we can reduce duplication when randomly drawing.It is easy to do.We just need to align the 'id'. It takes 20s on my pc,but 2.5h in this kernal.","execution_count":null},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true},"cell_type":"code","source":"def align_data():\n    en=pd.read_csv('../input/jigsaw-multilingual-toxic-comment-classification/jigsaw-toxic-comment-train.csv')\n    f1=pd.read_csv('../input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-es-cleaned.csv')#translated data that we want to align\n\n    result=[]\n    last_cell=[]\n    for i in tqdm(range(len(en['id']))):\n        index=np.where(f1['id']==en['id'][i])[0]\n        l=len(index)\n        if l==1:\n            cell=[f1['comment_text'][index[0]],'es',f1['toxic'][index[0]]]\n            result.append(cell)\n            last_cell=cell\n        else:\n            cell=last_cell\n            result.append(cell)   \n    print(len(en['toxic']))\n    print(len(result))\n    column=['comment_text','lang','toxic']\n    toxic=pd.DataFrame(columns=column,data=result)\n    toxic.to_csv('/kaggle/working/train1_es_align.csv'%(LANG,LANG),index=False)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"After that , we make all translated data have the same amount : 223549 (named train1) . And unintended_bias amount : 1902194 (named train2_max).","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"Then we can mix them in a certain ratio . For every row , we choose one from 6 files randomly.","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def mix_data():\n    #ratio for every language \n    ES_FRAC=0.132232182034727\n    TR_FRAC=0.21939447125932426\n    RU_FRAC=0.17156647652479157\n    IT_FRAC=0.13310975991976431\n    FR_FRAC=0.17112768758227292\n    PT_FRAC=0.17256942267911993\n\n    train_tr=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_tr.csv.csv')\n    train_es=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_es.csv.csv')\n    train_it=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_it.csv.csv')\n    train_fr=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_fr.csv.csv')\n    train_ru=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_ru.csv.csv')\n    train_pt=pd.read_csv('../input/jigsaw_toxic_comment_train_align/train1_pt.csv.csv')\n\n    result=[]\n    label=[]\n    for i in tqdm(range(0, len(train_es['comment_text']))):\n    rf = random.random()\n    if rf < ES_FRAC:\n        cell = [train_es['comment_text'][i], train_es['lang'][i], train_es['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    elif rf < (ES_FRAC + FR_FRAC):\n        cell = [train_fr['comment_text'][i], train_fr['lang'][i], train_fr['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    elif rf < (ES_FRAC + FR_FRAC + TR_FRAC):\n        cell = [train_tr['comment_text'][i], train_tr['lang'][i], train_tr['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    elif rf < (ES_FRAC + FR_FRAC + TR_FRAC + IT_FRAC):\n        cell = [train_it['comment_text'][i], train_it['lang'][i], train_it['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    elif rf < (ES_FRAC + FR_FRAC + TR_FRAC + IT_FRAC + PT_FRAC):\n        cell = [train_pt['comment_text'][i], train_pt['lang'][i], train_pt['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    else:\n        cell = [train_ru['comment_text'][i], train_ru['lang'][i], train_ru['toxic'][i]]\n        label.append(cell[2])\n        result.append(cell)\n    print(len(result))\n    print(len(label))\n\n    label=np.array(label)\n    np.save('/kaggle/working/train1_label.npy')\n    column=['comment_text','lang','toxic']\n    toxic=pd.DataFrame(columns=column,data=result)\n    toxic.to_csv('/kaggle/working/train1_mix.csv')\n    x_train=regular_encode(toxic['comment_text'], tokenizer, maxlen=MAX_LEN)\n    np.save('/kaggle/working/train1_mix.npy')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"After encode we get a data with 6 laguages.   Because it is generated randomly.   we can do it again and again to get a different data.","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"omit the encoding step","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"Then we use the encoded data for training","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def build_model(transformer, max_len=512,learnig_rate=1e-5):\n    \"\"\"\n    https://www.kaggle.com/xhlulu/jigsaw-tpu-distilbert-with-huggingface-and-keras\n    \"\"\"\n    input_word_ids = Input(shape=(max_len,), dtype=tf.int32, name=\"input_word_ids\")\n    sequence_output = transformer(input_word_ids)[0]\n    cls_token = sequence_output[:, 0, :]\n    #x = tf.keras.layers.Dropout(0.2)(cls_token)\n    out = Dense(1, activation='sigmoid')(cls_token)\n    \n    model = Model(inputs=input_word_ids, outputs=out)\n    model.compile(Adam(lr=learnig_rate), loss='binary_crossentropy', metrics=['accuracy'])\n    \n    return model","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"AUTO = tf.data.experimental.AUTOTUNE\n\n# Data access\nGCS_DS_PATH = KaggleDatasets().get_gcs_path(\"jigsaw-multilingual-toxic-comment-classification\")\n\n# Configuration\nEPOCHS = 3\nBATCH_SIZE = 24 * strategy.num_replicas_in_sync\nMAX_LEN = 192\nMODEL = 'jplu/tf-xlm-roberta-large'","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"valid = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/validation.csv')\n#valid = pd.read_csv('/kaggle/input/trans-tr/validation_tr.csv')\ntest = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/test.csv')\nsub = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/sample_submission.csv')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x_valid=np.load('/kaggle/input/toxic-npy/x_valid_encode192.npy')\nprint(x_valid)\nprint(x_valid.shape)\n\ny_valid = valid.toxic.values\nprint(y_valid)\nprint(y_valid.shape)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x_test=np.load('/kaggle/input/toxic-npy/x_test_encode192.npy')\n#x_test=np.load('/kaggle/input/trans-tr/test_tr_data.npy')\nprint(x_test)\nprint(x_test.shape)\nprint(\"finish x_test_encode\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"valid_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices((x_valid, y_valid))\n    .batch(BATCH_SIZE)\n    .cache()\n    .prefetch(AUTO)\n)\n\ntest_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices(x_test)\n    .batch(BATCH_SIZE)\n)\n","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"train2max ( shape(1902194,192) ) first .","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def train_train2max():\n    print('train train2max')\n\n    # Detect hardware, return appropriate distribution strategy\n    try:\n        # TPU detection. No parameters necessary if TPU_NAME environment variable is\n        # set: this is always the case on Kaggle.\n        tpu = tf.distribute.cluster_resolver.TPUClusterResolver()\n        print('Running on TPU ', tpu.master())\n    except ValueError:\n        tpu = None\n\n    if tpu:\n        tf.config.experimental_connect_to_cluster(tpu)\n        tf.tpu.experimental.initialize_tpu_system(tpu)\n        strategy = tf.distribute.experimental.TPUStrategy(tpu)\n    else:\n        # Default distribution strategy in Tensorflow. Works on CPU and single GPU.\n        strategy = tf.distribute.get_strategy()\n\n    print(\"REPLICAS: \", strategy.num_replicas_in_sync)\n\n\n    x_train=np.load('/kaggle/input/mix-data2/train2max21.npy')\n\n\n    y_train=np.load('/kaggle/input/mix-data2/train2max_label21.npy')\n    y_train=y_train.round().astype(int)\n\n\n    train_dataset = (\n        tf.data.Dataset\n        .from_tensor_slices((x_train, y_train))\n        .repeat()\n        .batch(BATCH_SIZE)\n        .shuffle(2048)\n        .prefetch(AUTO)\n    )\n\n\n    with strategy.scope():\n        transformer_layer = TFAutoModel.from_pretrained(MODEL)\n        model = build_model(transformer_layer, max_len=MAX_LEN,learnig_rate=5e-6)\n    model.summary()\n\n\n\n    n_steps = x_train.shape[0] // BATCH_SIZE\n    #n_steps = 680\n    train_history = model.fit(\n        train_dataset,\n        steps_per_epoch=n_steps,\n        validation_data=valid_dataset,\n        epochs=1\n    )\n\n    model.save_weights('/kaggle/working/model_mix_1.h5')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Then train1(shape(223549,192))","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def train_train1():\n    print('train train1')\n    # Detect hardware, return appropriate distribution strategy\n    try:\n        # TPU detection. No parameters necessary if TPU_NAME environment variable is\n        # set: this is always the case on Kaggle.\n        tpu = tf.distribute.cluster_resolver.TPUClusterResolver()\n        print('Running on TPU ', tpu.master())\n    except ValueError:\n        tpu = None\n\n    if tpu:\n        tf.config.experimental_connect_to_cluster(tpu)\n        tf.tpu.experimental.initialize_tpu_system(tpu)\n        strategy = tf.distribute.experimental.TPUStrategy(tpu)\n    else:\n        # Default distribution strategy in Tensorflow. Works on CPU and single GPU.\n        strategy = tf.distribute.get_strategy()\n\n    print(\"REPLICAS: \", strategy.num_replicas_in_sync)\n\n    x_train=np.load('/kaggle/input/mix-data2/train1_mix_data21.npy')\n\n\n    y_train=np.load('/kaggle/input/mix-data2/train1_label21.npy')\n    print(y_train)\n    print(y_train.shape)\n\n    train_dataset = (\n        tf.data.Dataset\n        .from_tensor_slices((x_train, y_train))\n        .repeat()\n        .batch(BATCH_SIZE)\n        .shuffle(2048)\n        .prefetch(AUTO)\n    )\n\n    with strategy.scope():\n        transformer_layer = TFAutoModel.from_pretrained(MODEL)\n        model = build_model(transformer_layer, max_len=MAX_LEN,learnig_rate=5e-6)\n\n\n    model.load_weights('/kaggle/working/model_mix_1.h5')\n    print(\"learning rate:5e-6\")\n    n_steps = x_train.shape[0] // BATCH_SIZE\n    #n_steps = 680\n    train_history = model.fit(\n        train_dataset,\n        steps_per_epoch=n_steps,\n        validation_data=valid_dataset,\n        epochs=1\n    )\n\n\n\n    model.save_weights('/kaggle/working/model_mix_1.h5')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Finally train on valid dataset","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def train_valid():\n    # Detect hardware, return appropriate distribution strategy\n    try:\n        # TPU detection. No parameters necessary if TPU_NAME environment variable is\n        # set: this is always the case on Kaggle.\n        tpu = tf.distribute.cluster_resolver.TPUClusterResolver()\n        print('Running on TPU ', tpu.master())\n    except ValueError:\n        tpu = None\n\n    if tpu:\n        tf.config.experimental_connect_to_cluster(tpu)\n        tf.tpu.experimental.initialize_tpu_system(tpu)\n        strategy = tf.distribute.experimental.TPUStrategy(tpu)\n    else:\n        # Default distribution strategy in Tensorflow. Works on CPU and single GPU.\n        strategy = tf.distribute.get_strategy()\n\n    with strategy.scope():\n        transformer_layer = TFAutoModel.from_pretrained(MODEL)\n        model = build_model(transformer_layer, max_len=MAX_LEN,learnig_rate=4e-6)\n\n\n    model.load_weights('/kaggle/working/model_mix_1.h5')\n\n    n_steps = x_valid.shape[0] // BATCH_SIZE\n    train_history_2 = model.fit(\n        valid_dataset.repeat(),\n        steps_per_epoch=n_steps,\n        epochs=2\n    )","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def predict():\n    score = model.predict(test_dataset,verbose=1)\n    sub['toxic']=score[:sub['toxic'].shape[0]]\n    sub.to_csv('submission.csv', index=False)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"train_train2max() -> train_train1() -> train_valid() -> predict()  is a complete process. Every time we can get a submissiom. After repeating it 20 times. We can get 20 submissions","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"#train_train2max()\n#train_train1()\n#train_valid()\n#predict()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Then we can sum them up , and get the mean","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"submission_train2mix1 = pd.read_csv('/kaggle/input/haveatry/submission9450.csv')\nsubmission_train2mix2 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_2.csv')\nsubmission_train2mix3 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_3.csv')\nsubmission_train2mix4 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_4.csv')\nsubmission_train2mix5 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_5.csv')\nsubmission_train2mix6 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_6.csv')\nsubmission_train2mix7 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_7.csv')\nsubmission_train2mix8 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_8.csv')\nsubmission_train2mix9 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_9.csv')\nsubmission_train2mix10 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_10.csv')\nsubmission_train2mix11 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_11.csv')\nsubmission_train2mix12 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_12.csv')\nsubmission_train2mix13 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_13.csv')\nsubmission_train2mix14 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_14.csv')\nsubmission_train2mix15 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_15.csv')\nsubmission_train2mix16 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_16.csv')\nsubmission_train2mix17 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_17.csv')\nsubmission_train2mix18 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_18.csv')\nsubmission_train2mix19 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_19.csv')\nsubmission_train2mix20 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_20.csv')\nsubmission_train2mix21 = pd.read_csv('/kaggle/input/haveatry/submission_train2mix_21.csv')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"for i in range(len(submission_train2mix1['toxic'])):\n    sub['toxic'][i]=(submission_train2mix1['toxic'][i]+submission_train2mix2['toxic'][i]\n                     +submission_train2mix3['toxic'][i]+submission_train2mix4['toxic'][i]\n                     +submission_train2mix5['toxic'][i]+submission_train2mix6['toxic'][i]\n                     +submission_train2mix7['toxic'][i]+submission_train2mix8['toxic'][i]\n                     +submission_train2mix9['toxic'][i]+submission_train2mix10['toxic'][i]\n                     +submission_train2mix11['toxic'][i]+submission_train2mix12['toxic'][i]\n                     +submission_train2mix13['toxic'][i]+submission_train2mix14['toxic'][i]\n                    +submission_train2mix15['toxic'][i]+submission_train2mix16['toxic'][i]\n                    +submission_train2mix17['toxic'][i]+submission_train2mix18['toxic'][i]\n                    +submission_train2mix19['toxic'][i]+submission_train2mix20['toxic'][i]\n                    +submission_train2mix21['toxic'][i])/21\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sub.to_csv('submission.csv', index=False)#~.9471(Public LB)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"5times:~.9450(Public LB)","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"12times:~.9464(Public LB)","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"21times:~.9471(Public LB)","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"With easy ensemble,it can easily get a sliver medal.For me it's ~.9472(Private) ~.9488(Public)","execution_count":null}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}