Peter

I would be interested. I actually have a [function for computing the gradient of a layer norm directly](https://github.com/chengchingwen/NeuralAttentionlib.jl/blob/master/src/functional/layernorm.jl#L30) in NAlib. This is the best (in terms of both performance and...

It should definitely be possible. They are wrapped in `Defer` because it cannot found the corresponding method to reconstruct the object. There is a method lookup mechanism behind it, though...

Yes, I think we can use some rule to determine whether the `Array` is compatible with numpy array, and if not, fallback to python list.

numpy array also have python object dtype, we probably can use it for `Array{Any, N}`, but I haven't try it.

> I tried to dig into the source code to check if an ordered Dict is used. I.e. Python at some point changed from unordered to ordered, so when you...

I kinda agree the situation of python dict and it is problematic. But instead of changing the default type, I am thinking of some ways to change the behavior of...

``` Try setting the `proto` keyword argument when loading, e.g. `load(file; proto = 5)`. ``` What is the error log of `load("MV-Budget.pkl"; proto = 5)`?

> That will lead to this: That is a result indicating that there are some stuff unknown to Pickle.jl. Pandas.jl don't have issues because they call python directly. In order...

> build the DataFrame from the Defer obj? It's definitely doable. This is how we support a new python object with Pickle.jl, but we need someone to actually implement that.

You would need to implement a julia function that perform `keras.src.saving.pickle_utils.deserialize_model_from_bytecode` on `UInt8[0x50, 0x4b, 0x03, 0x04, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00 … 0xb2, 0x00, 0x00, 0x00, 0xc6, 0xa5, 0xee,...