Something wrong with the loss #10
Comments
I cut the dataset to use 500 sentences only. you might want to use the whole dataset for better experiments. |
Actually, I used the whole dataset but the perplexity shows too low (lower means better). |
Oh, I believe thats good? how about test some BLEU or word position accuracy on dev / test split? I implemented word position accuracy in the notebooks, can you paste some logs here? |
Epoch 0 Batch 50/225 - Loss: 3.2396 - perplextity: 25.5224 |
According to the results showed above, the test perplexity achieves 6.3277 after 20 epochs, which is too much lower than usual results from other published papers. I will try some other metrics. |
skeptical much? haha. Maybe you read old research papers? maybe they use old seq2seq APIs? |
Hello, I have run your codes of 'chatbot' on a conversation dataset. But the loss seems unnormally low. The results on dailydialog dataset from paper 'DailyDialog: AManuallyLabelledMulti-turnDialogueDataset' show that perplexity is more than 30 and loss is more than 3. But the perplexity obtained by your codes is lower 3 which is absolutely wrong. Could you provide some advice? Thank you.
'
The text was updated successfully, but these errors were encountered: