clean code

README.md

@@ -0,0 +1,13 @@
+# CSI-Code
+This project is a pytorch implementation of the paper [CSI: A Hybrid Deep Model for Fake News Detection"](https://dl.acm.org/citation.cfm?id=3132877)
+Several modifications have been made on the [original code](https://github.com/sungyongs/CSI-Code) including:
+    - minor bugs were fixed
+    - code is cleaner, standard, organized, and more understandable
+    - the main model is reimplemented using Pytorch and Pytorch_lightning
+    - training is integrated with Ray Tune for the sake of hyperparameter tuning
+    
+    
+## Getting started
+1. First, use the [data preprocessing notebook]() to prepare your data for the model. You should apply the code for different splits separately.
+
+2. Then, use the [train notebook]() to train your model and see the results.

src/data.py

@@ -0,0 +1,21 @@
+import os
+import numpy as np
+import pickle
+import torch
+from torch.utils.data import Dataset
+
+class CSIDataset(Dataset):
+    def __init__(self, pkl_dir):
+        super().__init__()
+        self.samples = []
+        for file in os.listdir(pkl_dir):
+            sample = pickle.load(open(pkl_dir + file, 'rb'))
+            if sample['x_score'].shape[0] > 0:
+                self.samples += [sample]
+        
+    def __getitem__(self, idx):
+        data = self.samples[idx]
+        return data['x_capture'], data['x_score'], data['label']
+    
+    def __len__(self):
+        return len(self.samples)

src/model.py

@@ -0,0 +1,96 @@
+import torch
+from torch import nn
+from torchmetrics.functional import confusion_matrix
+import pytorch_lightning as pl
+
+class CSIModel(pl.LightningModule):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.criterion = nn.BCELoss()
+        self.capture_rnn = nn.Sequential(
+            nn.Linear(config['capture_input_dim'], config['d_Wa']),
+            nn.Tanh(),
+            nn.Dropout(config['dropout']),
+            nn.LSTM(
+                input_size=config['d_Wa'], 
+                hidden_size=config['d_lstm'],
+                num_layers=1, 
+                batch_first=True)
+        )
+        self.capture_proj = nn.Sequential(
+            nn.Linear(config['d_lstm'], config['d_Wr']),
+            nn.Tanh(),
+            nn.Dropout(config['dropout'])
+        )
+        self.score = nn.Sequential(
+            nn.Linear(config['score_input_dim'], config['d_Wu']),
+            nn.Tanh(),
+            nn.Linear(config['d_Wu'], config['d_Ws']),
+            nn.Sigmoid()
+        )
+        self.cls = nn.Sequential(
+            nn.Linear(config['d_Ws'] + config['d_Wr'], 1),
+            nn.Sigmoid()
+        )
+        
+    def configure_optimizers(self):
+        all_params = dict(self.named_parameters())
+        wd_name = 'score.0.weight'
+        wd_params = all_params[wd_name]
+        del all_params[wd_name]
+        return torch.optim.Adam(
+            [
+                {'params':  wd_params, 'weight_decay': self.config['weight_decay']}, 
+                {'params': list(all_params.values())},
+            ], 
+            lr=self.config['lr']
+        )
+        
+
+    def count_parameters(self):
+        return sum(p.numel() for p in self.parameters() if p.requires_grad)    
+        
+    def forward(self, x_capture, x_score):
+        hc, (_, _) = self.capture_rnn(x_capture.float())
+        hc = self.capture_proj(hc[:, -1])
+        hs = self.score(x_score.float()).mean(dim=1)
+        h = torch.cat([hc, hs], dim=1)
+        return self.cls(h)
+    
+    def step(self, batch, mode='train'):
+        x_capture, x_score, labels = batch
+        labels = labels[:, None].float()
+        logits = self.forward(x_capture, x_score)
+        loss = self.criterion(logits, labels)
+        
+        preds = logits.clone()
+        preds[preds >=0.5] = 1
+        preds[preds < 0.5] = 0
+        acc = (preds == labels).sum() / labels.shape[0]
+        tn, fn, fp, tp = confusion_matrix(logits, labels.int(), num_classes=2, threshold=0.5).flatten()
+
+        self.log(f'{mode}_loss', loss.item())
+        self.log(f'{mode}_acc', acc.item())
+        self.log(f'{mode}_tn', tn.item())
+        self.log(f'{mode}_fn', fn.item())
+        self.log(f'{mode}_fp', fp.item())
+        self.log(f'{mode}_tp', tp.item())
+        return {
+            'loss':loss, 
+            'acc':acc, 
+            'tn':tn, 
+            'fn':fn, 
+            'fp':fp, 
+            'tp':tp
+        }
+    
+    def training_step(self, batch, batch_idx):
+        return self.step(batch)
+    
+    def test_step(self, batch, batch_idx):
+        return self.step(batch, mode='test')
+    
+    def validation_step(self, batch, batch_idx):
+        return self.step(batch, mode='val')
+    

train.ipynb

@@ -0,0 +1,214 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Global seed set to 42\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "42"
+      ]
+     },
+     "execution_count": 1,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import numpy as np\n",
+    "import os\n",
+    "from importlib import reload\n",
+    "from copy import deepcopy\n",
+    "import json\n",
+    "import pandas as pd\n",
+    "from tqdm.notebook import tqdm\n",
+    "\n",
+    "import ray\n",
+    "from ray import tune\n",
+    "from ray.tune.schedulers import ASHAScheduler\n",
+    "\n",
+    "import torch\n",
+    "from torch import nn\n",
+    "import pytorch_lightning as pl\n",
+    "\n",
+    "from pytorch_lightning import Trainer, seed_everything\n",
+    "from pytorch_lightning.loggers import TensorBoardLogger\n",
+    "from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor\n",
+    "from pytorch_lightning.callbacks.early_stopping import EarlyStopping\n",
+    "from ray.tune.integration.pytorch_lightning import TuneReportCallback, TuneReportCheckpointCallback\n",
+    "\n",
+    "seed_everything(42)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## data"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import src.data\n",
+    "reload(src.data)\n",
+    "from src.data import CSIDataset\n",
+    "\n",
+    "from torch.utils.data import DataLoader"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import src.model\n",
+    "reload(src.model)\n",
+    "from src.model import CSIModel\n",
+    "\n",
+    "\n",
+    "def experiment(args):\n",
+    "    dataset = 'weibo'\n",
+    "    path = f'/media/external_10TB/10TB/ramezani/Omranpour/assets/{dataset}/'\n",
+    "    \n",
+    "    train_set = CSIDataset(pkl_dir=path + 'train/pkls/')\n",
+    "    val_set = CSIDataset(pkl_dir=path + 'validation/pkls/')\n",
+    "    train_loader = DataLoader(train_set, batch_size=1, shuffle=False, num_workers=1)\n",
+    "    val_loader = DataLoader(val_set, batch_size=1, shuffle=False, num_workers=1)\n",
+    "    \n",
+    "    \n",
+    "    conf = {\n",
+    "        'capture_input_dim' : 112,\n",
+    "        'score_input_dim' : 50,\n",
+    "        'd_Wa': 100,\n",
+    "        'd_lstm' : 50,\n",
+    "        'd_Wr' : 100,\n",
+    "        'd_Wu' : 100,\n",
+    "        'd_Ws' : 1,\n",
+    "        'lr': args['lr'],\n",
+    "        'dropout' : args['dropout'],\n",
+    "        'weight_decay' : args['weight_decay']\n",
+    "    }\n",
+    "    model = CSIModel(conf)\n",
+    "\n",
+    "    name = f\"dataset={dataset}-do={args['dropout']}-lr={args['lr']}-wd={args['weight_decay']}\"\n",
+    "    save_dir = f'/media/external_10TB/10TB/ramezani/Omranpour/CSI/weights/{name}/'\n",
+    "    logger = TensorBoardLogger(save_dir='logs/', name=name)\n",
+    "    checkpoint = ModelCheckpoint(\n",
+    "        dirpath=save_dir, \n",
+    "        filename='{epoch}-{val_loss:.2f}', \n",
+    "        monitor='val_loss',\n",
+    "        mode='min',\n",
+    "        save_top_k=10, \n",
+    "        every_n_epochs = 5\n",
+    "    )\n",
+    "#     reporter = TuneReportCallback(\n",
+    "#         {\n",
+    "#             \"loss\": \"ptl/val_loss\",\n",
+    "#             \"mean_accuracy\": \"ptl/val_acc\"\n",
+    "#         },\n",
+    "#         on=\"validation_end\"\n",
+    "#     )\n",
+    "    os.makedirs(save_dir, exist_ok=True)\n",
+    "    json.dump(conf, open(save_dir + 'config.json', 'w'))\n",
+    "\n",
+    "    trainer = Trainer(\n",
+    "        benchmark=True, \n",
+    "        gpus=[1], \n",
+    "        accumulate_grad_batches=64,\n",
+    "        logger=logger, \n",
+    "        enable_progress_bar=False,\n",
+    "        max_epochs=10,\n",
+    "        callbacks=[checkpoint]\n",
+    "    )\n",
+    "    trainer.fit(model, train_loader, val_loader)\n",
+    "    res = trainer.validate(val_loader)[0]\n",
+    "    tune.report(**res)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "== Status ==<br>Current time: 2022-03-09 01:53:36 (running for 00:01:52.55)<br>Memory usage on this node: 14.1/125.6 GiB<br>Using FIFO scheduling algorithm.<br>Resources requested: 1.0/18 CPUs, 2.0/2 GPUs, 0.0/71.06 GiB heap, 0.0/34.44 GiB objects (0.0/1.0 accelerator_type:GTX)<br>Result logdir: /home/ramezani/ray_results/experiment_2022-03-09_01-51-43<br>Number of trials: 16/16 (15 PENDING, 1 RUNNING)<br><br>"
+      ],
+      "text/plain": [
+       "<IPython.core.display.HTML object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "analysis = tune.run(\n",
+    "    experiment,\n",
+    "    num_samples=4,\n",
+    "    resources_per_trial={\"cpu\": 1, \"gpu\": 2},\n",
+    "    verbose=1,\n",
+    "    config={\n",
+    "        \"weight_decay\": tune.grid_search([0., 0.1, 0.01, 0.001]),\n",
+    "        \"lr\": tune.loguniform(1e-5, 1e-1),\n",
+    "        \"dropout\": tune.uniform(0., 0.3)\n",
+    "    }\n",
+    ")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "df = analysis.results_df"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}