{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\nfrom sklearn import metrics\n\nimport os\n\n# Any results you write to the current directory are saved as output.","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true},"cell_type":"markdown","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"},{"metadata":{"trusted":true},"cell_type":"code","source":"import torch\n\n# import torch_xla.distributed.xla_multiprocessing as xmp\n\nimport torch.nn as nn\n\nimport transformers\nfrom transformers import AdamW, get_linear_schedule_with_warmup, get_constant_schedule\n\n# import torch_xla\n# import torch_xla.debug.metrics as met\n# import torch_xla.distributed.data_parallel as dp\n# import torch_xla.distributed.parallel_loader as pl\n# import torch_xla.utils.utils as xu\n# import torch_xla.core.xla_model as xm\n# import torch_xla.distributed.xla_multiprocessing as xmp\n# import torch_xla.test.test_utils as test_utils\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"markdown","source":"class JigsawDataset:\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    def __len__(self):\n        return len(self.comment_text)\n\n    def __getitem__(self, item):\n        comment_text = str(self.comment_text[item])\n        comment_text = \" \".join(comment_text.split())\n\n        inputs = self.tokenizer.encode_plus(\n            comment_text,\n            None,\n            add_special_tokens=True,\n            max_length=self.max_length,\n        )\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 {\n            'ids': torch.tensor(ids, dtype=torch.long),\n            'mask': torch.tensor(mask, dtype=torch.long),\n            'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),\n            'targets': torch.tensor(self.targets[item], dtype=torch.float)\n        }"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"df_train1 = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/jigsaw-toxic-comment-train.csv\", usecols=[\"comment_text\", \"toxic\"]).fillna(\"none\")\ndf_train2 = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/jigsaw-unintended-bias-train.csv\", usecols=[\"comment_text\", \"toxic\"]).fillna(\"none\")"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"df_train= pd.concat([df_train1, df_train2], axis=0).reset_index(drop=True)\ndf_train = df_train.sample(frac=1).reset_index(drop=True).head(200000)"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"df_valid = pd.read_csv('../input/jigsaw-multilingual-toxic-comment-classification/validation.csv', \n                       usecols=[\"comment_text\", \"toxic\"])"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"MAX_LEN = 192\nTRAIN_BATCH_SIZE = 64*2\nEPOCHS = 2 #2\n\ntokenizer = transformers.BertTokenizer.from_pretrained(\"../input/bert-base-multilingual-uncased/\", do_lower_case=True)"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"    train_targets = df_train.toxic.values\n    valid_targets = df_valid.toxic.values\n\n    train_dataset = JigsawDataset(\n        comment_text=df_train.comment_text.values,\n        targets=train_targets,\n        tokenizer=tokenizer,\n        max_length=MAX_LEN\n    )\n\n    train_sampler = torch.utils.data.distributed.DistributedSampler(\n          train_dataset,\n          num_replicas=xm.xrt_world_size(),\n          rank=xm.get_ordinal(),\n          shuffle=True)\n\n    train_data_loader = torch.utils.data.DataLoader(\n        train_dataset,\n        batch_size=TRAIN_BATCH_SIZE,\n        sampler=train_sampler,\n        drop_last=True,\n        num_workers=1\n    )\n\n    valid_dataset = JigsawDataset(\n        comment_text=df_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    )"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"class BERTBaseMultiUncased(nn.Module):\n    def __init__(self, bert_path):\n        super(BERTBaseMultiUncased, self).__init__()\n#         super(self).__init__()\n        self.bert_path = bert_path\n        self.bert = transformers.BertModel.from_pretrained(self.bert_path)\n        self.bert_drop = nn.Dropout(0.3)\n        self.out = nn.Linear(768 * 2, 1)\n\n    def forward(\n            self,\n            ids,\n            mask,\n            token_type_ids\n    ):\n        o1, o2 = self.bert(\n            ids,\n            attention_mask=mask,\n            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        p2 = self.out(bo)\n        return p2"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"mx = BERTBaseMultiUncased(bert_path=\"../input/bert-base-multilingual-uncased/\")"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"xm.xla_device()"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"device = xm.xla_device()\nmodel = mx.to(device)\n\nparam_optimizer = list(model.named_parameters())\nno_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']\noptimizer_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\nlr = 0.4 * 1e-5 * xm.xrt_world_size()\nnum_train_steps = int(len(train_dataset) / TRAIN_BATCH_SIZE / xm.xrt_world_size() * EPOCHS)\nxm.master_print(f'num_train_steps = {num_train_steps}, world_size={xm.xrt_world_size()}')\n\noptimizer = AdamW(optimizer_grouped_parameters, lr=lr)\nscheduler = get_linear_schedule_with_warmup(\n    optimizer,\n    num_warmup_steps=0,\n    num_training_steps=num_train_steps\n)"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"\ndef _run():\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(\n                ids=ids,\n                mask=mask,\n                token_type_ids=token_type_ids\n            )\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            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    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_multi.bin\")\n        torch.save(model.to('cpu').state_dict(), 'model_multi.bin')\n        auc = metrics.roc_auc_score(np.array(t) >= 0.5, o)\n        xm.master_print(f'AUC = {auc}')"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"# Start training processes\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=1, start_method='fork')"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"print(\"done\")"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"\nfrom IPython.display import FileLink\nFileLink(r'model_multi.bin')"},{"metadata":{"trusted":true},"cell_type":"markdown","source":"from sklearn import metrics"},{"metadata":{"trusted":true},"cell_type":"code","source":"    \nclass JigsawDataset:\n    def __init__(self, comment_text, tokenizer, max_length):\n        self.comment_text = comment_text\n        self.tokenizer = tokenizer\n        self.max_length = max_length\n\n    def __len__(self):\n        return len(self.comment_text)\n\n    def __getitem__(self, item):\n        comment_text = str(self.comment_text[item])\n        comment_text = \" \".join(comment_text.split())\n\n        inputs = self.tokenizer.encode_plus(\n            comment_text,\n            None,\n            add_special_tokens=True,\n            max_length=self.max_length,\n        )\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 {\n            'ids': torch.tensor(ids, dtype=torch.long),\n            'mask': torch.tensor(mask, dtype=torch.long),\n            'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long)\n        }\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"    \nclass BERTBaseMultiUncased(nn.Module):\n    def __init__(self, bert_path):\n        super(BERTBaseMultiUncased, self).__init__()\n        self.bert_path = bert_path\n        self.bert = transformers.BertModel.from_pretrained(self.bert_path)\n        self.bert_drop = nn.Dropout(0.3)\n        self.out = nn.Linear(768 * 2, 1)\n\n    def forward(\n            self,\n            ids,\n            mask,\n            token_type_ids\n    ):\n        o1, o2 = self.bert(\n            ids,\n            attention_mask=mask,\n            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        p2 = self.out(bo)\n        return p2","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"df = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/test.csv\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tokenizer = transformers.BertTokenizer.from_pretrained(\"../input/bert-base-multilingual-uncased/\", do_lower_case=True)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"device = \"cuda\"\nmodel = BERTBaseMultiUncased(bert_path=\"../input/bert-base-multilingual-uncased/\").to(device)\nmodel.load_state_dict(torch.load(\"../input/jigsawmulti-bertmulti/model_multi.bin\"))\nmodel.eval()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"valid_dataset = JigsawDataset(\n        comment_text=df.content.values,\n        tokenizer=tokenizer,\n        max_length=192\n)\n\nvalid_data_loader = torch.utils.data.DataLoader(\n    valid_dataset,\n    batch_size=64,\n    drop_last=False,\n    num_workers=1,\n    shuffle=False\n)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from tqdm import tqdm","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"with torch.no_grad():\n    fin_outputs = []\n    for bi, d in tqdm(enumerate(valid_data_loader)):\n        ids = d[\"ids\"]\n        mask = d[\"mask\"]\n        token_type_ids = d[\"token_type_ids\"]\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\n        outputs = model(\n            ids=ids,\n            mask=mask,\n            token_type_ids=token_type_ids\n        )\n\n        outputs_np = outputs.cpu().detach().numpy().tolist()\n        fin_outputs.extend(outputs_np)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sample = pd.read_csv(\"../input/jigsaw-multilingual-toxic-comment-classification/sample_submission.csv\")\nsample.loc[:, \"toxic\"] = fin_outputs\nsample.to_csv(\"submission.csv\", index=False)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sample.head()","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"name":"python","version":"3.6.4","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat":4,"nbformat_minor":4}