iakarshu/docformer_for_document_classification

1import torch.nn as nn2from modeling import DocFormerEncoder,ResNetFeatureExtractor,DocFormerEmbeddings,LanguageFeatureExtractor3 4class DocFormerForClassification(nn.Module):5  6    def __init__(self, config):7      super(DocFormerForClassification, self).__init__()8 9      self.resnet = ResNetFeatureExtractor(hidden_dim = config['max_position_embeddings'])10      self.embeddings = DocFormerEmbeddings(config)11      self.lang_emb = LanguageFeatureExtractor()12      self.config = config13      self.dropout = nn.Dropout(config['hidden_dropout_prob'])14      self.linear_layer = nn.Linear(in_features = config['hidden_size'], out_features = 16)  ## Number of Classes15      self.encoder = DocFormerEncoder(config)16 17    def forward(self, batch_dict):18 19      x_feat = batch_dict['x_features']20      y_feat = batch_dict['y_features']21 22      token = batch_dict['input_ids']23      img = batch_dict['resized_scaled_img']24 25      v_bar_s, t_bar_s = self.embeddings(x_feat,y_feat)26      v_bar = self.resnet(img)27      t_bar = self.lang_emb(token)28      out = self.encoder(t_bar,v_bar,t_bar_s,v_bar_s)29      out = self.linear_layer(out)30      out = out[:, 0, :]31      return out32 33 34## Defining pytorch lightning model35import pytorch_lightning as pl36from sklearn.metrics import accuracy_score, confusion_matrix37import pandas as pd38import matplotlib.pyplot as plt39import seaborn as sns40import numpy as np41import torchmetrics42import wandb43import torch44 45class DocFormer(pl.LightningModule):46 47  def __init__(self, config , lr = 5e-5):48    super(DocFormer, self).__init__()49    50    self.save_hyperparameters()51    self.config = config52    self.docformer = DocFormerForClassification(config)53    54    self.num_classes = 1655    self.train_accuracy_metric = torchmetrics.Accuracy()56    self.val_accuracy_metric = torchmetrics.Accuracy()57    self.f1_metric = torchmetrics.F1Score(num_classes=self.num_classes)58    self.precision_macro_metric = torchmetrics.Precision(59            average="macro", num_classes=self.num_classes60        )61    self.recall_macro_metric = torchmetrics.Recall(62            average="macro", num_classes=self.num_classes63        )64    self.precision_micro_metric = torchmetrics.Precision(average="micro")65    self.recall_micro_metric = torchmetrics.Recall(average="micro")66 67  def forward(self, batch_dict):68    logits = self.docformer(batch_dict)69    return logits70 71  def training_step(self, batch, batch_idx):72    logits = self.forward(batch)73 74    loss = nn.CrossEntropyLoss()(logits, batch['label'])75    preds = torch.argmax(logits, 1)76 77    ## Calculating the accuracy score78    train_acc = self.train_accuracy_metric(preds, batch["label"])79 80    ## Logging81    self.log('train/loss', loss,prog_bar = True, on_epoch=True, logger=True, on_step=True)82    self.log('train/acc', train_acc, prog_bar = True, on_epoch=True, logger=True, on_step=True)83 84    return loss85  86  def validation_step(self, batch, batch_idx):87    logits = self.forward(batch)88    loss = nn.CrossEntropyLoss()(logits, batch['label'])89    preds = torch.argmax(logits, 1)90    91    labels = batch['label']92    # Metrics93    valid_acc = self.val_accuracy_metric(preds, labels)94    precision_macro = self.precision_macro_metric(preds, labels)95    recall_macro = self.recall_macro_metric(preds, labels)96    precision_micro = self.precision_micro_metric(preds, labels)97    recall_micro = self.recall_micro_metric(preds, labels)98    f1 = self.f1_metric(preds, labels)99 100    # Logging metrics101    self.log("valid/loss", loss, prog_bar=True, on_step=True, logger=True)102    self.log("valid/acc", valid_acc, prog_bar=True, on_epoch=True, logger=True, on_step=True)103    self.log("valid/precision_macro", precision_macro, prog_bar=True, on_epoch=True, logger=True, on_step=True)104    self.log("valid/recall_macro", recall_macro, prog_bar=True, on_epoch=True, logger=True, on_step=True)105    self.log("valid/precision_micro", precision_micro, prog_bar=True, on_epoch=True, logger=True, on_step=True)106    self.log("valid/recall_micro", recall_micro, prog_bar=True, on_epoch=True, logger=True, on_step=True)107    self.log("valid/f1", f1, prog_bar=True, on_epoch=True)108    109    return {"label": batch['label'], "logits": logits}110 111  def validation_epoch_end(self, outputs):112        labels = torch.cat([x["label"] for x in outputs])113        logits = torch.cat([x["logits"] for x in outputs])114        preds = torch.argmax(logits, 1)115 116        wandb.log({"cm": wandb.sklearn.plot_confusion_matrix(labels.cpu().numpy(), preds.cpu().numpy())})117        self.logger.experiment.log(118            {"roc": wandb.plot.roc_curve(labels.cpu().numpy(), logits.cpu().numpy())}119        )120        121  def configure_optimizers(self):122    return torch.optim.AdamW(self.parameters(), lr = self.hparams['lr'])