[fix] Check None value for Pytorch model gradients histogram

[kage08]

When tracking histograms of gradients, some tensors may not have gradients during specific training loops.
So check if grad is not None before converting to numpy.

This commit fixes #1925

[CLAassistant]

All committers have signed the CLA.

[gorarakelyan]

@kage08 thanks a lot for the contribution! 🙌
@alberttorosyan could you please take a look at this?

[SGevorg]

@kage08 thanks a lot for the contribution. 🙏
Could you open a corresponding issue (with some usage context) and connect to this PR?
We use issues to track the changes and by adding to milestones to which version the changes got shipped with 🙌

[alberttorosyan]

@kage08 thanks for contribution 🙌
may I ask you to update the CHANGELOG.md to include the fix in Unreleased section?
also I've changed the PR header to match the naming conventions.

[alberttorosyan]

looks good!
the remaining steps:

1. open a GH issue
2. update CHANGELOG.md to include this fix

[gorarakelyan]

Made a minor fix in the changelog :)

btw, it is worth considering getting the weight tensor only once, isn't it?

e.g:

weight = None
if hasattr(m, 'weight') and m.weight is not None:
    weight = getattr(m.weight, dt, None)

if weight is not None:
    layers[layer_name]['weight'] = get_pt_tensor(getattr(m.weight, dt)).numpy()

instead of:

if hasattr(m, 'weight') and m.weight is not None and getattr(m.weight, dt, None) is not None:
    layers[layer_name]['weight'] = get_pt_tensor(getattr(m.weight, dt)).numpy()

thoughts? @kage08 @alberttorosyan

[kage08]

Made a minor fix in the changelog :)

btw, it is worth considering getting the weight tensor only once, isn't it?

e.g:

weight = None
if hasattr(m, 'weight') and m.weight is not None:
    weight = getattr(m.weight, dt, None)

if weight is not None:
    layers[layer_name]['weight'] = get_pt_tensor(getattr(m.weight, dt)).numpy()

instead of:

if hasattr(m, 'weight') and m.weight is not None and getattr(m.weight, dt, None) is not None:
    layers[layer_name]['weight'] = get_pt_tensor(getattr(m.weight, dt)).numpy()

thoughts? @kage08 @alberttorosyan

This makes sense for me. It avoids getting value twice.
Personally I would use the walrus operator, but I think that may break compatibility. So I agree with proposed change.

[gorarakelyan]

@kage08 looks awesome 🙌

[mihran113]

@kage08 looks good 🙌 thanks a lot for contribution

[gorarakelyan]

Merged the PR! @kage08 congrats on the first merged contribution! 🚀