FEA & FIX: allow training_neg_sample_num > 1 in Sequential && add clip_grad_norm to trainer.py

recbole/data/dataloader/sequential_dataloader.py

@@ -221,8 +221,9 @@ def _neg_sampling(self, data):
             return self.sampling_func(data, neg_iids)
 
     def _neg_sample_by_pair_wise_sampling(self, data, neg_iids):
-        data[self.neg_item_id] = neg_iids
-        return data
+        new_data = {key: np.concatenate([value] * self.times) for key, value in data.items()}
+        new_data[self.neg_item_id] = neg_iids
+        return new_data
 
     def _neg_sample_by_point_wise_sampling(self, data, neg_iids):
         new_data = {}

recbole/properties/overall.yaml

@@ -15,6 +15,7 @@ learning_rate: 0.001
 training_neg_sample_num: 1
 eval_step: 1
 stopping_step: 10
+clip_grad_norm: {'max_norm': 5, 'norm_type': 2}
 
 # evaluation settings
 eval_setting: RO_RS,full

recbole/trainer/trainer.py

@@ -16,6 +16,7 @@
 import itertools
 import torch
 import torch.optim as optim
+from torch.nn.utils.clip_grad import clip_grad_norm_
 import numpy as np
 import matplotlib.pyplot as plt
 
@@ -76,6 +77,7 @@ def __init__(self, config, model):
         self.epochs = config['epochs']
         self.eval_step = min(config['eval_step'], self.epochs)
         self.stopping_step = config['stopping_step']
+        self.clip_grad_norm = config['clip_grad_norm']
         self.valid_metric = config['valid_metric'].lower()
         self.valid_metric_bigger = config['valid_metric_bigger']
         self.test_batch_size = config['eval_batch_size']
@@ -149,6 +151,8 @@ def _train_epoch(self, train_data, epoch_idx, loss_func=None):
                 total_loss = losses.item() if total_loss is None else total_loss + losses.item()
             self._check_nan(loss)
             loss.backward()
+            if self.clip_grad_norm:
+                clip_grad_norm_(self.model.parameters(), **self.clip_grad_norm)
             self.optimizer.step()
         return total_loss