{"cells":[{"metadata":{"trusted":true},"cell_type":"code","source":"!curl https://raw.githubusercontent.com/pytorch/xla/master/contrib/scripts/env-setup.py -o pytorch-xla-env-setup.py\n!python pytorch-xla-env-setup.py --apt-packages libomp5 libopenblas-dev","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import os\nimport torch\nimport pandas as pd\nfrom scipy import stats\nimport numpy as np\n\nfrom tqdm import tqdm\nfrom collections import OrderedDict, namedtuple\nimport torch.nn as nn\nfrom torch.optim import lr_scheduler\nimport joblib\n\nimport logging\nimport transformers\nfrom transformers import AdamW, get_linear_schedule_with_warmup, get_constant_schedule\nimport sys\nfrom sklearn import metrics, model_selection\n\nimport warnings\nimport torch_xla\nimport torch_xla.debug.metrics as met\nimport torch_xla.distributed.data_parallel as dp\nimport torch_xla.distributed.parallel_loader as pl\nimport torch_xla.utils.utils as xu\nimport torch_xla.core.xla_model as xm\nimport torch_xla.distributed.xla_multiprocessing as xmp\nimport torch_xla.test.test_utils as test_utils\n\nclass AverageMeter:\n    \"\"\"\n    Computes and stores the average and current value\n    \"\"\"\n    def __init__(self):\n        self.reset()\n        \n    #resets the values to zero\n    def reset(self):\n        self.val = 0\n        self.avg = 0\n        self.sum = 0\n        self.count = 0\n    \n    #updates the values of average\n    def update(self, val, n=1):\n        self.val = val\n        self.sum += val * n\n        self.count += n\n        self.avg = self.sum / self.count","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Making the Bert Model"},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true},"cell_type":"code","source":"# a class to convert the dataset into lowercase\nclass BERTBaseUncased(nn.Module):\n    #initialization function taking the path as a parameter\n        def __init__(self, path):\n            super(BERTBaseUncased, self).__init__()\n            self.bert_path = path\n            #Instantiate a pretrained pytorch model from a pre-trained model configuration.\n            self.bert = transformers.BertModel.from_pretrained(self.bert_path)\n            #randomly sets elements to zero to prevent overfitting.\n            self.bert_drop = nn.Dropout(0.3) \n            #linear/outer layer as bert base model has 768*2 output features (bert base and multilingual) and 1 for binary classification\n            self.out = nn.Linear(768 * 2, 1)\n        \n        def forward(self,ids,mask,token_type_ids):\n            #o1 ,is the last hidden and we neglect pooler outputs of the bert \n            o1,_ = self.bert(ids,attention_mask=mask,token_type_ids=token_type_ids)\n            \n            apool = torch.mean(o1, 1)\n            mpool, _ = torch.max(o1, 1)\n            cat = torch.cat((apool, mpool), 1)\n\n            bo = self.bert_drop(cat)\n            output = self.out(bo)\n            return output","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## data loader"},{"metadata":{"trusted":true},"cell_type":"code","source":"#a class to accomodate the dataset to bert model\nclass BERTDatasetTrain:\n    def __init__(self, comment_text,targets, tokenizer, max_length):\n        self.comment_text = comment_text\n        self.tokenizer = tokenizer\n        self.max_length = max_length\n        self.targets=targets\n\n    #getting the total length\n    def __len__(self):\n        return len(self.comment_text)\n    \n    #returns the token type ids from the dataset of the comment_text\n    def __getitem__(self, item):\n        #checking for the digits \n        comment_text = str(self.comment_text[item])\n        #removing all the unwanted spaces\n        comment_text = \" \".join(comment_text.split())\n\n        #encode 2 strings at a time hence 2nd string is none and add the CLS token\n        inputs = self.tokenizer.encode_plus(comment_text,None,add_special_tokens=True,max_length=self.max_length,)\n        ids = inputs[\"input_ids\"]\n        token_type_ids = inputs[\"token_type_ids\"]\n        mask = inputs[\"attention_mask\"]\n        \n        padding_length = self.max_length - len(ids)\n        \n        ids = ids + ([0] * padding_length)\n        mask = mask + ([0] * padding_length)\n        token_type_ids = token_type_ids + ([0] * padding_length)\n        \n        return {'ids': torch.tensor(ids, dtype=torch.long),'mask': torch.tensor(mask, dtype=torch.long),'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),'targets': torch.tensor(self.targets[item], dtype=torch.float)}","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"mx = BERTBaseUncased(path=\"../input/bert-base-multilingual-uncased/\")\n\ntrain1 = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/jigsaw-toxic-comment-train.csv\", usecols=[\"comment_text\", \"toxic\"]).fillna(\"none\")\ntrain2 = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/jigsaw-unintended-bias-train.csv\", usecols=[\"comment_text\", \"toxic\"]).fillna(\"none\")\ntrain_combine = pd.concat([train1, train2], axis=0).reset_index(drop=True)\ntrain = train_combine.sample(frac=1).reset_index(drop=True).head(200000)\n\nvalid = pd.read_csv('../input/jigsaw-multilingual-toxic-comment-classification/validation.csv', usecols=[\"comment_text\", \"toxic\"])\n\n#combining the valid dataset with the train dataset\ntrain = pd.concat([train, valid], axis=0).reset_index(drop=True)\ntrain = train.sample(frac=1).reset_index(drop=True)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def run():\n    \n    def loss_fn(outputs, targets):\n        return nn.BCEWithLogitsLoss()(outputs, targets.view(-1, 1))\n\n    def train_loop_fn(data_loader, model, optimizer, device, scheduler=None):\n        model.train()\n        for bi, d in enumerate(data_loader):\n            ids = d[\"ids\"]\n            mask = d[\"mask\"]\n            token_type_ids = d[\"token_type_ids\"]\n            targets = d[\"targets\"]\n\n            ids = ids.to(device, dtype=torch.long)\n            mask = mask.to(device, dtype=torch.long)\n            token_type_ids = token_type_ids.to(device, dtype=torch.long)\n            targets = targets.to(device, dtype=torch.float)\n\n            optimizer.zero_grad()\n            outputs = model(ids=ids,mask=mask,token_type_ids=token_type_ids)\n\n            loss = loss_fn(outputs, targets)\n            if bi % 10 == 0:\n                xm.master_print(f'bi={bi}, loss={loss}')\n\n            loss.backward()\n            #Xla optimizer\n            xm.optimizer_step(optimizer)\n            if scheduler is not None:\n                scheduler.step()\n\n    def eval_loop_fn(data_loader, model, device):\n        model.eval()\n        fin_targets = []\n        fin_outputs = []\n        for bi, d in enumerate(data_loader):\n            ids = d[\"ids\"]\n            mask = d[\"mask\"]\n            token_type_ids = d[\"token_type_ids\"]\n            targets = d[\"targets\"]\n\n            ids = ids.to(device, dtype=torch.long)\n            mask = mask.to(device, dtype=torch.long)\n            token_type_ids = token_type_ids.to(device, dtype=torch.long)\n            targets = targets.to(device, dtype=torch.float)\n\n            outputs = model(\n                ids=ids,\n                mask=mask,\n                token_type_ids=token_type_ids\n            )\n\n            targets_np = targets.cpu().detach().numpy().tolist()\n            outputs_np = outputs.cpu().detach().numpy().tolist()\n            fin_targets.extend(targets_np)\n            fin_outputs.extend(outputs_np)    \n\n        return fin_outputs, fin_targets\n\n    \n    MAX_LEN = 192\n    TRAIN_BATCH_SIZE = 64\n    EPOCHS = 15\n\n    tokenizer = transformers.BertTokenizer.from_pretrained(\"../input/bert-base-multilingual-uncased/\", do_lower_case=True)\n\n    train_targets = train.toxic.values\n    valid_targets = valid.toxic.values\n\n    train_dataset = BERTDatasetTrain(comment_text=train.comment_text.values,targets=train_targets,tokenizer=tokenizer,max_length=MAX_LEN)\n\n    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset,num_replicas=xm.xrt_world_size(),rank=xm.get_ordinal(),shuffle=True)\n\n    train_data_loader = torch.utils.data.DataLoader(train_dataset,batch_size=TRAIN_BATCH_SIZE,sampler=train_sampler,drop_last=True,num_workers=1)\n\n    valid_dataset = BERTDatasetTrain(\n        comment_text=valid.comment_text.values,\n        targets=valid_targets,\n        tokenizer=tokenizer,\n        max_length=MAX_LEN\n    )\n\n    valid_sampler = torch.utils.data.distributed.DistributedSampler(\n          valid_dataset,\n          num_replicas=xm.xrt_world_size(),\n          rank=xm.get_ordinal(),\n          shuffle=False)\n\n    valid_data_loader = torch.utils.data.DataLoader(\n        valid_dataset,\n        batch_size=16,\n        sampler=valid_sampler,\n        drop_last=False,\n        num_workers=1\n    )\n\n    device = xm.xla_device()\n    model = mx.to(device)\n\n    param_optimizer = list(model.named_parameters())\n    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']\n    optimizer_grouped_parameters = [\n        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.001},\n        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]\n\n    lr = 0.4 * 1e-5 * xm.xrt_world_size()\n    num_train_steps = int(len(train_dataset) / TRAIN_BATCH_SIZE / xm.xrt_world_size() * EPOCHS)\n    xm.master_print(f'num_train_steps = {num_train_steps}, world_size={xm.xrt_world_size()}')\n\n    #Adam Optimizer\n    optimizer = AdamW(optimizer_grouped_parameters, lr=lr)\n    scheduler = get_linear_schedule_with_warmup(\n        optimizer,\n        num_warmup_steps=0,\n        num_training_steps=num_train_steps\n    )\n\n    for epoch in range(EPOCHS):\n        para_loader = pl.ParallelLoader(train_data_loader, [device])\n        train_loop_fn(para_loader.per_device_loader(device), model, optimizer, device, scheduler=scheduler)\n\n        para_loader = pl.ParallelLoader(valid_data_loader, [device])\n        o, t = eval_loop_fn(para_loader.per_device_loader(device), model, device)\n        xm.save(model.state_dict(), \"model.bin\")\n        \n        # AUC tells how much model is capable of distinguishing between classes. \n        auc = metrics.roc_auc_score(np.array(t) >= 0.5, o)\n        xm.master_print(f'AUC = {auc}')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Start the training process\ndef _mp_fn(rank, flags):\n    torch.set_default_tensor_type('torch.FloatTensor')\n    a = run()\n\nFLAGS={}\nxmp.spawn(_mp_fn, args=(FLAGS,), nprocs=8, start_method='fork')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","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}