# Convert existing model/weights fromData Parallel to generic?

**URL:** <https://discuss.pytorch.org/t/convert-existing-model-weights-fromdata-parallel-to-generic/128061>\
**Category:** deployment\
**Created:** [July 29, 2021, 2:42pm UTC](https://discuss.pytorch.org/t/convert-existing-model-weights-fromdata-parallel-to-generic/128061 "2021-07-29T14:42:00Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![DrEGPU](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/dregpu/32/13844_2.png) [@DrEGPU](https://discuss.pytorch.org/u/DrEGPU)\
**Post date:** [July 29, 2021, 2:42pm UTC](https://discuss.pytorch.org/t/convert-existing-model-weights-fromdata-parallel-to-generic/128061/1 "2021-07-29T14:42:00Z")

</div>

Sorry if this isn’t the correct topic. I did try to look for my problem, but I don’t I think I’m using the right search terms, which is why I’m here.

So, I trained a model using `nn.DataParallel` and it worked fine. I was able to inference fine on my machine, but only my machine. I finally figured out that I mistakenly did NOT save my model as generic as described here: [https://pytorch.org/tutorials/beginner/saving\_loading\_models.html#saving-torch-nn-dataparallel-models](https://pytorch.org/tutorials/beginner/saving_loading_models.html#saving-torch-nn-dataparallel-models)

Is there a way to convert my existing model to a generic form so it can be used by anyone? Hopefully in a automated way and not me sludging through the weights manually?

---

<div class="post-metadata">

**Author:** ![ayalaa2](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ayalaa2/32/21195_2.png) [@ayalaa2](https://discuss.pytorch.org/u/ayalaa2)\
**Post date:** [July 29, 2021, 7:38pm UTC](https://discuss.pytorch.org/t/convert-existing-model-weights-fromdata-parallel-to-generic/128061/2 "2021-07-29T19:38:37Z")

</div>

Are you still able to load the checkpoint? DataParallel just hides the actual model away under the attribute `module`. So you can do something like this:

```auto
my_model = ModelClass(...)
my_model = DataParallel(my_model, ...)

state_dict = torch.load(...)
my_model.load_state_dict(state_dict)

actual_model = my_model.module
new_state_dict = actual_model

```

Haven’t tested this, but I bet something like this would work as well:

```auto
state_dict = torch.load(...)

new_state_dict = {}
for k, v in state_dict.items():
    k = k.replace('module.', '') # Each weight k will be prefix with the word "module"
    new_state_dict[k] = v

model = SomeModel(...)
model.load_state_dict(new_state_dict)

```

---

<div class="post-metadata">

**Author:** ![DrEGPU](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/dregpu/32/13844_2.png) [@DrEGPU](https://discuss.pytorch.org/u/DrEGPU)\
**Post date:** [July 30, 2021, 2:45pm UTC](https://discuss.pytorch.org/t/convert-existing-model-weights-fromdata-parallel-to-generic/128061/3 "2021-07-30T14:45:22Z")

</div>

Thanks, we tried something similar to the second method and it “works”, but unsure if the model is accurately predicting or not. Still investigating… Thanks again!