{"cells":[{"metadata":{},"cell_type":"markdown","source":"## Imports","execution_count":null},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import os\nimport re\nimport numpy as np\nimport pandas as pd\n\n\nimport tensorflow as tf\nfrom tensorflow.keras.layers import Dense, Input, Embedding, Conv1D, MaxPooling1D, Flatten\nfrom tensorflow.keras.optimizers import Adam\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.callbacks import ModelCheckpoint\n\nimport transformers\nfrom transformers import TFAutoModel, AutoTokenizer, RobertaTokenizerFast\n\nfrom tqdm.notebook import tqdm\nfrom tokenizers import Tokenizer, models, pre_tokenizers, decoders, processors\n\n","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Helper Functions","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"def clean_sentence(sentence):\n    ''' \n    Author: louilghada@gmail.com | kaggle.com/swannnn\n    Removes punctuation, digits, special characters, stopwords and words of 1 character\n    '''\n    clean_sent = ''''''\n    for word in sentence.split():\n        if len(word) > 1 and not re.match(r'.*[0-9]+', word):\n            clean_sent = \"{} {}\".format(clean_sent, word)\n    clean_sent = re.sub(r'[.,:;?!/\\|@#$%^&-_(){}]', '', clean_sent)\n    clean_sent = text = re.sub(r'(http|www)\\S*', '', clean_sent)\n    return clean_sent.strip()\n\ndef clean_df(df, column):\n    '''\n    louilghada@gmail.com | kaggle.com/swannnn\n    Cleans text in specified column of dataframe df\n    '''\n    df = df.apply(lambda x: x.astype(str).str.lower())\n    df[column] = df[column].apply(lambda x: clean_sentence(x))\n\ndef regular_encode(texts, tokenizer, maxlen=512):\n    enc_di = tokenizer.batch_encode_plus(\n        texts, \n        return_attention_masks=False, \n        return_token_type_ids=False,\n        pad_to_max_length=True,\n        max_length=maxlen\n    )\n    \n    return np.array(enc_di['input_ids'])\n\ndef build_model(transformer, max_len=512):\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    out = Dense(1, activation='sigmoid')(cls_token)\n    \n    model = Model(inputs=input_word_ids, outputs=out)\n    model.compile(Adam(lr=1e-5), loss='binary_crossentropy', metrics=['accuracy'])\n    \n    return model\n\ndef build_CNN_model():\n    input_layer = Input(shape=(MAX_LEN,), dtype=tf.int32, name=\"input_words_ids\")\n    embedding = Embedding(VOCAB_SIZE[0], VOCAB_SIZE[1], input_length=MAX_LEN, name='embed')(input_layer)\n    conv_1 = Conv1D(256, (100), activation='relu')(embedding)\n    max_pool = MaxPooling1D()(conv_1)\n    conv_1 = Conv1D(128, (5), activation='relu')(max_pool)\n    max_pool = MaxPooling1D()(conv_1)\n    dense = Dense(128, activation='relu')(max_pool)\n    dense = Dense(128, activation='relu')(dense)\n    flatten = Flatten()(dense)\n    out = Dense(1, activation='sigmoid')(flatten)\n    \n    \n    model = Model(inputs = input_layer, outputs=out)\n    model.compile(Adam(lr=1e-5), loss='binary_crossentropy', metrics=['accuracy'])\n    return model","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## TPU Configs","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"# Detect hardware, return appropriate distribution strategy\ntry:\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())\nexcept ValueError:\n    tpu = None\n\nif tpu:\n    tf.config.experimental_connect_to_cluster(tpu)\n    tf.tpu.experimental.initialize_tpu_system(tpu)\n    strategy = tf.distribute.experimental.TPUStrategy(tpu)\nelse:\n    # Default distribution strategy in Tensorflow. Works on CPU and single GPU.\n    strategy = tf.distribute.get_strategy()\n\nprint(\"REPLICAS: \", strategy.num_replicas_in_sync)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"AUTO = tf.data.experimental.AUTOTUNE\n\n# Configuration\nEPOCHS = 1\nBATCH_SIZE = 16 * strategy.num_replicas_in_sync\nMAX_LEN = 192\nMODEL = 'jplu/tf-xlm-roberta-large'","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Create fast tokenizer","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"# First load the real tokenizer\ntokenizer = AutoTokenizer.from_pretrained(MODEL)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Load text data into memory","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"train1 = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/jigsaw-toxic-comment-train.csv\")\ntrain2 = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/jigsaw-unintended-bias-train.csv\")\ntrain3 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-es-cleaned.csv\")\ntrain4 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-fr-cleaned.csv\")\ntrain5 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-it-cleaned.csv\")\ntrain6 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-pt-cleaned.csv\")\ntrain7 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-ru-cleaned.csv\")\ntrain8 = pd.read_csv(\"/kaggle/input/jigsaw-train-multilingual-coments-google-api/jigsaw-toxic-comment-train-google-tr-cleaned.csv\")\n\nvalid1 = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/validation.csv')\nvalid2 = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-test-translated/jigsaw_miltilingual_valid_translated.csv')\ntest1 = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/test.csv')\ntest2 = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-test-translated/jigsaw_miltilingual_test_translated.csv')\nsub = pd.read_csv('/kaggle/input/jigsaw-multilingual-toxic-comment-classification/sample_submission.csv')\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Combine train1 with a subset of train2\ntrain = pd.concat([\n    train1[['comment_text', 'toxic']].query('toxic==1'),\n    train1[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train2[['comment_text', 'toxic']].query('toxic==1'),\n    train2[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train3[['comment_text', 'toxic']].query('toxic==1'),\n    train3[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train4[['comment_text', 'toxic']].query('toxic==1'),\n    train4[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train5[['comment_text', 'toxic']].query('toxic==1'),\n    train5[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train6[['comment_text', 'toxic']].query('toxic==1'),\n    train6[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train7[['comment_text', 'toxic']].query('toxic==1'),\n    train7[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    train8[['comment_text', 'toxic']].query('toxic==1'),\n    train8[['comment_text', 'toxic']].query('toxic==0').sample(n=50000, random_state=0),\n    \n])\n\nvalid = pd.concat([valid1, valid2])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"clean_df(train, 'comment_text')\ntrain.toxic = train.toxic.round().astype(int)\ntrain = train.sample(frac = 1)\n\nclean_df(valid, 'comment_text')\n\nclean_df(test1, 'content')\nclean_df(test2, 'content')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x_train = regular_encode(train.comment_text.values, tokenizer, maxlen=MAX_LEN)\nx_valid = regular_encode(valid.comment_text.values, tokenizer, maxlen=MAX_LEN)\n\nx_test1 = regular_encode(test1.content.values, tokenizer, maxlen=MAX_LEN)\nx_test2 = regular_encode(test2.content.values, tokenizer, maxlen=MAX_LEN)\n\ny_train = train.toxic.values\ny_valid = valid.toxic.values\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"x_train1 = regular_encode(train1.comment_text.values, tokenizer, maxlen=MAX_LEN)\n\ntrain_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices((x_train, y_train))\n    .repeat()\n    .shuffle(2048)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTO)\n)\n","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Build datasets objects","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"train_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices((x_train, y_train))\n    .repeat()\n    .shuffle(2048)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTO)\n)\n\nvalid_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices((x_valid, y_valid))\n    .batch(BATCH_SIZE)\n    .cache()\n    .prefetch(AUTO)\n)\n\ntest1_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices(x_test1)\n    .batch(BATCH_SIZE)\n)\n\ntest2_dataset = (\n    tf.data.Dataset\n    .from_tensor_slices(x_test2)\n    .batch(BATCH_SIZE)\n)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Load models into the TPU","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"VOCAB_SIZE = x_train.shape","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"with strategy.scope():\n    transformer_layer = TFAutoModel.from_pretrained(MODEL)\n    roberta_model = build_model(transformer_layer, max_len=MAX_LEN)\nroberta_model.summary()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"with strategy.scope():\n    cnn_model = build_CNN_model()\ncnn_model.summary()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Train Models","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"First, we train the XLM-Roberta Model.","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"n_steps = x_train.shape[0] // BATCH_SIZE\n\ntrain_history = roberta_model.fit(\n    train_dataset,\n    steps_per_epoch=n_steps,\n    validation_data=valid_dataset,\n    epochs=EPOCHS\n)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"n_steps = x_valid.shape[0] // BATCH_SIZE\n\ntrain_history_2 = roberta_model.fit(\n    valid_dataset.repeat(),\n    steps_per_epoch=n_steps,\n    epochs=EPOCHS\n)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Then we train the CNN model","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"n_steps = x_train.shape[0] // BATCH_SIZE\n\ntrain_history = cnn_model.fit(\n    train_dataset,\n    steps_per_epoch=n_steps,\n    validation_data=valid_dataset,\n    epochs=EPOCHS*5\n)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"n_steps = x_valid.shape[0] // BATCH_SIZE\n\ntrain_history_2 = cnn_model.fit(\n    valid_dataset.repeat(),\n    steps_per_epoch=n_steps,\n    epochs=EPOCHS\n)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## blending","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"multi_ling_sub = roberta_model.predict(test1_dataset, verbose=1)\neng_sub = roberta_model.predict(test2_dataset, verbose=1)\n\nsub['toxic'] = multi_ling_sub*0.5 + eng_sub*0.5","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Submission","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"sub.to_csv('submission.csv', index=False)","execution_count":null,"outputs":[]}],"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}