code

[sleepyzh]

        with torch.no_grad():
            val_metrics = corpus.get_validation_pred(model, 'test')
        if val_metrics['Mean Reciprocal Rank'] > best_test_metrics['Mean Reciprocal Rank']:
            best_test_metrics['Mean Reciprocal Rank'] = val_metrics['Mean Reciprocal Rank']
        if val_metrics['Mean Rank'] < best_test_metrics['Mean Rank']:
            best_test_metrics['Mean Rank'] = val_metrics['Mean Rank']
        if val_metrics['Hits@1'] > best_test_metrics['Hits@1']:
            best_test_metrics['Hits@1'] = val_metrics['Hits@1']
        if val_metrics['Hits@3'] > best_test_metrics['Hits@3']:
            best_test_metrics['Hits@3'] = val_metrics['Hits@3']
        if val_metrics['Hits@10'] > best_test_metrics['Hits@10']:
            best_test_metrics['Hits@10'] = val_metrics['Hits@10']
        if val_metrics['Hits@100'] > best_test_metrics['Hits@100']:
            best_test_metrics['Hits@100'] = val_metrics['Hits@100']

print(' '.join(['Test set results:',
                model.format_metrics(best_test_metrics, 'test')]))

In the code, the result obtained from this evaluation is not the best epoch result, but the best value of all indicators. They do not appear in the same epoch.

[sleepyzh]

GAT( (spgat): SpGAT( (dropout_layer): Dropout(p=0.3, inplace=False) (attention_0): SpGraphAttentionLayer (256 -> 256) (attention_1): SpGraphAttentionLayer (256 -> 256) (out_att): SpGraphAttentionLayer (512 -> 512) ) ) In GATencoder, with 2 heads and a dim of 256, the final output dimension of the model is 512, which does not match the embedded dim.

[sleepyzh]

The context model you proposed in the paper was not found in the code,Comparative learning（def contrastive_loss(self, s_embed, v_embed, t_embed):） is not used in the default code, how do you use it. Looking forward to your reply

[sun8982]

The context model you proposed in the paper was not found in the code,Comparative learning（def contrastive_loss(self, s_embed, v_embed, t_embed):） is not used in the default code, how do you use it. Looking forward to your reply

May I ask what version of the environment you are running the code in? I used a 1.7 torch and an 11.0 cuda to run on the 3090 graphics card, but they couldn't run and reported the error ："Traceback (most recent call last):" File "main. py", line 234, in Train_encoder (args) File "main. py", line 119, in train_encoder Entity_embed, relation_embed=model. forward (corpus. train.adj_matrix, train_indices) File "/media/xxx/4962f605-dcf4-4c7f-95a6-36b784b7ace0/gpuserver/sh/paper_code/IMF Pytorch/models/model. py", line 78, in forward Self. entity_embeddings. data=F.normalize (self. entity_embeddings. data, dim=1) File "/home/xxx/anaconda3/envs/imf/lib/python3.7/site packages/torch/nn/functional. py", line 3788, in normalize Denom=input. norm (p, dim, keepdim=True). slap_min (eps). expand.as (input) File "/home/xxx/anaconda3/envs/imf/lib/python3.7/site packages/torch/sensor. py", line 389, in norm Return torch. norm (self, p, dim, keepdim, dtype=dtype) File "/home/xxx/anaconda3/envs/imf/lib/python3.7/site packages/torch/functional. py", line 1337, in norm Return VVF. norm (input, p, _dim, keepdim=keepdim) # type: ignore "RuntimeError: CUDA error: no kernel image is available for execution on the device"