# Loss for multi output model

**URL:** <https://discuss.pytorch.org/t/loss-for-multi-output-model/107959>\
**Category:** Uncategorized\
**Created:** [January 5, 2021, 7:42am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959 "2021-01-05T07:42:20Z")\
**Posts on this page:** 7\
**Page:** 1

<div class="post-metadata">

**Author:** ![di\_19](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/di_19/32/33056_2.png) [@di\_19](https://discuss.pytorch.org/u/di_19)\
**Post date:** [January 5, 2021, 7:42am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/1 "2021-01-05T07:42:20Z")

</div>

```auto
 issue_loss_fn = nn.CrossEntropyLoss()
 product_loss_fn = nn.CrossEntropyLoss()
 issue_loss = issue_loss_fn(issue_logits, issue_y)
 product_loss = product_loss_fn(pro_logits, issue_y)
 loss = issue_loss + product_loss
 loss.backward()

```

is this the correct way to do it? Or am I doing anything wrong?  
when one label result improves other degrades

```auto
epoch: 0 | loss: 1.705373126316846 | issue acc: 0.45827687443541104 | product acc: 0.5626646717856068 | time elapsed: 143.8780710697174
------------------------------------------------------------------------------------------
epoch: 0 | loss: 1.223887874321504 | issue acc: 0.630754662004662 | product acc: 0.7407852564102565 | time elapsed: 152.7476155757904
------------------------------------------------------------------------------------------

epoch: 1 | loss: 1.1055800391406547 | issue acc: 0.6752625338753387 | product acc: 0.6483316395663957 | time elapsed: 146.28524684906006
------------------------------------------------------------------------------------------
epoch: 1 | loss: 1.1996180794455789 | issue acc: 0.663224796037296 | product acc: 0.6807255244755245 | time elapsed: 155.0542027950287
------------------------------------------------------------------------------------------

epoch: 2 | loss: 0.8235846665816579 | issue acc: 0.7612635501355014 | product acc: 0.6567487202649804 | time elapsed: 146.46119809150696
------------------------------------------------------------------------------------------
epoch: 2 | loss: 1.1156758611852473 | issue acc: 0.7048186188811189 | product acc: 0.6636800699300699 | time elapsed: 155.3083529472351
------------------------------------------------------------------------------------------

epoch: 3 | loss: 0.6182250726998337 | issue acc: 0.8257772508280637 | product acc: 0.6497760463715748 | time elapsed: 146.60942959785461
------------------------------------------------------------------------------------------
epoch: 3 | loss: 1.4946378902955488 | issue acc: 0.700247668997669 | product acc: 0.638567162004662 | time elapsed: 155.38528728485107
------------------------------------------------------------------------------------------

```

any suggestions  
second output for every epoch is for validation set

---

<div class="post-metadata">

**Author:** ![caonv](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/caonv/32/31818_2.png) [@caonv](https://discuss.pytorch.org/u/caonv)\
**Post date:** [January 6, 2021, 4:18am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/2 "2021-01-06T04:18:52Z")

</div>

You should have to put weights on partial losses to avoid the circumstance that one dominates the sum.

---

<div class="post-metadata">

**Author:** ![di\_19](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/di_19/32/33056_2.png) [@di\_19](https://discuss.pytorch.org/u/di_19)\
**Post date:** [January 6, 2021, 4:39am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/3 "2021-01-06T04:39:31Z")

</div>

Hi @caonv Thanks for the answer.  
I have wrote some code based on your suggestion. Could you check once. Is this the correct way

```auto
 w1 = torch.randn(1)
 w2 = torch.randn(1)
 issue_loss_fn = nn.CrossEntropyLoss()
 product_loss_fn = nn.CrossEntropyLoss()
 issue_loss = w1 * issue_loss_fn(issue_logits, issue_y)
 product_loss = w2 * product_loss_fn(pro_logits, issue_y)
 loss = issue_loss + product_loss
 loss.backward()

```

---

<div class="post-metadata">

**Author:** ![caonv](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/caonv/32/31818_2.png) [@caonv](https://discuss.pytorch.org/u/caonv)\
**Post date:** [January 6, 2021, 5:18am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/4 "2021-01-06T05:18:14Z")

</div>

yes, these weights are hyper-parameters, so that you have to search for proper values. I’ve read somewhere, you could start from values that scale the losses into the same range.

---

<div class="post-metadata">

**Author:** ![di\_19](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/di_19/32/33056_2.png) [@di\_19](https://discuss.pytorch.org/u/di_19)\
**Post date:** [January 6, 2021, 6:00am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/5 "2021-01-06T06:00:07Z")

</div>

Okay @caonv  
Thanks for information. I’ll try that

---

<div class="post-metadata">

**Author:** ![di\_19](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/di_19/32/33056_2.png) [@di\_19](https://discuss.pytorch.org/u/di_19)\
**Post date:** [January 6, 2021, 7:19am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/6 "2021-01-06T07:19:58Z")

</div>

unfortunately, That trick doesn’t work in my case @caonv

---

<div class="post-metadata">

**Author:** ![caonv](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/caonv/32/31818_2.png) [@caonv](https://discuss.pytorch.org/u/caonv)\
**Post date:** [January 6, 2021, 8:16am UTC](https://discuss.pytorch.org/t/loss-for-multi-output-model/107959/7 "2021-01-06T08:16:53Z")

</div>

you should search for the proper values 😃