add hyperparam to control frobenius

.vscode/launch.json

@@ -0,0 +1,36 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Python: Current File",
+            "type": "python",
+            "request": "launch",
+            "program": "${file}",
+            "console": "integratedTerminal",
+            "justMyCode": true,
+            "args": [
+                "--gpus", "1,", 
+                "--max_epochs=16",  
+                "--num_workers=32", 
+                "--model_name_or_path",  "bert-base-uncased",
+                "--accumulate_grad_batches", "1", 
+                "--model_class", "BertKGC",
+                "--batch_size", "32",
+                "--checkpoint", "/root/kg_374/Relphormer/pretrain/output/FB15k-237/epoch=15-step=38899-Eval/hits10=0.96.ckpt",
+                "--pretrain", "0",
+                "--bce", "0",
+                "--check_val_every_n_epoch", "1",
+                "--data_dir", "dataset/FB15k-237", 
+                "--eval_batch_size", "128",
+                "--max_seq_length", "128",
+                "--lr", "3e-5",
+                "--max_triplet", "64",
+                "--add_attn_bias", "True",
+                "--use_global_node", "True",
+            ]
+        }
+    ]
+}

lit_models/transformer.py

@@ -48,7 +48,11 @@ class TransformerLitModel(BaseLitModel):
         if args.bce:
             self.loss_fn = torch.nn.BCEWithLogitsLoss()
         elif args.label_smoothing != 0.0:
-            self.loss_fn = LabelSmoothSoftmaxCEV1(lb_smooth=args.label_smoothing)
+            self.cross_entropy_loss = nn.CrossEntropyLoss()
+            self.smoothing = args.label_smoothing
+            self.loss_fn = self.label_smoothed_cross_entropy
+            self.frobenius_reg = args.weight_decay
+            # self.loss_fn = LabelSmoothSoftmaxCEV1(lb_smooth=args.label_smoothing)
         else:
             self.loss_fn = nn.CrossEntropyLoss()
         self.best_acc = 0
@@ -69,13 +73,71 @@ class TransformerLitModel(BaseLitModel):
         self.spatial_pos_encoder = nn.Embedding(5, self.num_heads, padding_idx=0)
         self.graph_token_virtual_distance = nn.Embedding(1, self.num_heads)
         
+    def label_smoothed_cross_entropy(self, logits, labels):
+        num_classes = logits.size(1)
+        one_hot = torch.zeros_like(logits).scatter(1, labels.unsqueeze(1), 1)
+        one_hot = one_hot * (1 - self.smoothing) + (1 - one_hot) * self.smoothing / (num_classes - 1)
+        loss = self.cross_entropy_loss(logits, labels)
+        return loss
+
+    def frobenius_norm_loss(self):
+        frobenius_norm = 0.0
+        for name, param in self.model.named_parameters():
+            if 'bias' not in name:
+                frobenius_norm += torch.norm(param, p='fro')
+        return frobenius_norm
 
     def forward(self, x):
         return self.model(x)
 
+    def create_negatives(self, batch):
+        negativeBatches = []
+        label = batch['label']
+
+        for i in range(label.shape[0]):
+            newBatch = {}
+            newBatch['attention_mask'] = None
+            newBatch['input_ids'] = torch.clone(batch['input_ids'])
+            newBatch['label'] = torch.zeros_like(batch['label'])
+            negativeBatches.append(newBatch)
+
+        entity_idx = []
+        self_label = []
+        for idx, l in enumerate(label):
+            decoded = self.decode([batch['input_ids'][idx]])[0].split(' ')
+            for j in range(1, len(decoded)):
+                if (decoded[j].startswith("[ENTITY_")):
+                    entity_idx.append(j)
+                    self_label.append(batch['input_ids'][idx][j])   
+                    break
+        
+        for idx, lbl in enumerate(label):
+            for i in range(label.shape[0]):
+                if (negativeBatches[idx]['input_ids'][i][entity_idx[i]] != lbl):
+                    negativeBatches[idx]['input_ids'][i][entity_idx[i]] = lbl
+                else:
+                    negativeBatches[idx]['input_ids'][i][entity_idx[i]] = self_label[i]
+                
+        return negativeBatches
+
     def training_step(self, batch, batch_idx):  # pylint: disable=unused-argument
         # embed();exit()
         # print(self.optimizers().param_groups[1]['lr'])
+
+        negativeBatches = self.create_negatives(batch)
+
+        loss = 0
+
+        for negativeBatch in negativeBatches:
+            label = negativeBatch.pop("label")
+            input_ids = batch['input_ids']
+            
+            logits = self.model(**negativeBatch, return_dict=True, distance_attention=None).logits
+            _, mask_idx = (input_ids == self.tokenizer.mask_token_id).nonzero(as_tuple=True)
+            bs = input_ids.shape[0]
+            mask_logits = logits[torch.arange(bs), mask_idx][:, self.entity_id_st:self.entity_id_ed]
+            loss += self.loss_fn(mask_logits, label)
+
         labels = batch.pop("labels")
         label = batch.pop("label")
         pos = batch.pop("pos")
@@ -110,9 +172,12 @@ class TransformerLitModel(BaseLitModel):
 
         assert mask_idx.shape[0] == bs, "only one mask in sequence!"
         if self.args.bce:
-            loss = self.loss_fn(mask_logits, labels)
+            loss += self.loss_fn(mask_logits, labels)
         else:
-            loss = self.loss_fn(mask_logits, label)
+            loss += self.loss_fn(mask_logits, label)
+
+        if self.smoothing is not None and self.smoothing != 0.0:
+            loss += self.frobenius_reg * self.frobenius_norm_loss()
 
         if batch_idx == 0:
             print('\n'.join(self.decode(batch['input_ids'][:4])))

main.py

@@ -120,7 +120,7 @@ def main():
     callbacks = [early_callback, model_checkpoint]
 
     # args.weights_summary = "full"  # Print full summary of the model
-    trainer = pl.Trainer.from_argparse_args(args, callbacks=callbacks, logger=logger, default_root_dir="training/logs", accelerator="ddp")
+    trainer = pl.Trainer.from_argparse_args(args, callbacks=callbacks, logger=logger, default_root_dir="training/logs")
     
     if "EntityEmbedding" not in lit_model.__class__.__name__:
         trainer.fit(lit_model, datamodule=data)

pretrain/scripts/pretrain_fb15k-237.sh

@@ -1,13 +1,13 @@
-nohup python -u main.py --gpus "1" --max_epochs=16  --num_workers=32 \
+nohup python -u main.py --gpus "1," --max_epochs=16  --num_workers=32 \
    --model_name_or_path  bert-base-uncased \
    --accumulate_grad_batches 1 \
    --model_class BertKGC \
-   --batch_size 128 \
+   --batch_size 64 \
    --pretrain 1 \
    --bce 0 \
    --check_val_every_n_epoch 1 \
    --overwrite_cache \
-   --data_dir /kg_374/Relphormer/dataset/FB15k-237 \
+   --data_dir /root/kg_374/Relphormer/dataset/FB15k-237 \
    --eval_batch_size 256 \
    --max_seq_length 64 \
    --lr 1e-4 \

scripts/fb15k-237/fb15k-237.sh

@@ -1,13 +1,12 @@
-nohup python -u main.py --gpus "1" --max_epochs=16  --num_workers=32 \
+nohup python -u main.py --gpus "3," --max_epochs=16  --num_workers=32 \
    --model_name_or_path  bert-base-uncased \
    --accumulate_grad_batches 1 \
    --model_class BertKGC \
-   --batch_size 64 \
-   --checkpoint /kg_374/Relphormer/pretrain/output/FB15k-237/epoch=15-step=19299-Eval/hits10=0.96.ckpt \
+   --batch_size 16 \
+   --checkpoint /root/kg_374/Relphormer_instance_1/pretrain/output/FB15k-237/epoch\=15-step\=38899-Eval/hits10=0.96.ckpt \
    --pretrain 0 \
    --bce 0 \
    --check_val_every_n_epoch 1 \
-   --overwrite_cache \
    --data_dir dataset/FB15k-237 \
    --eval_batch_size 128 \
    --max_seq_length 128 \