# List comprehension in Zygote

**URL:** <https://discourse.julialang.org/t/list-comprehension-in-zygote/41881>\
**Category:** General Usage\
**Tags:** zygote\
**Created:** [June 22, 2020, 7:02pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881 "2020-06-22T19:02:34Z")\
**Posts on this page:** 9\
**Page:** 1

<div class="post-metadata">

**Author:** ![rlrs](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rlrs/32/18043_2.png) [@rlrs](https://discourse.julialang.org/u/rlrs)\
**Post date:** [June 22, 2020, 7:02pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/1 "2020-06-22T19:02:34Z")

</div>

I’m trying to do something like this, (I hope the snippet is enough although it’s incomplete)  
`prob_primitives = [(θ[i+1], p) for (i, p) in enumerate(primitives)]`  
where primitives is a Vector. However, using this code in a Zygote.gradient call results in an error **MethodError: no method matching iterate(::Nothing)**.

I’ve tried this code as an alternative,  
`prob_primitives = collect(zip(θ[2:end], primitives))`  
but `collect` constructs the array using mutation which is also unsupported.

---

<div class="post-metadata">

**Author:** ![rlrs](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rlrs/32/18043_2.png) [@rlrs](https://discourse.julialang.org/u/rlrs)\
**Post date:** [June 23, 2020, 8:49am UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/2 "2020-06-23T08:49:40Z")

</div>

Answering this myself: it’s possible by creating a `Zygote.Buffer` and instantiating this through mutation. I don’t know if it’s the best solution, but it works.

---

<div class="post-metadata">

**Author:** ![ianfiske](https://avatars.discourse-cdn.com/v4/letter/i/58f4c7/32.png) [@ianfiske](https://discourse.julialang.org/u/ianfiske)\
**Post date:** [June 23, 2020, 1:03pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/3 "2020-06-23T13:03:46Z")

</div>

Zygote works fine with array comprehensions. I use it frequently. Here’s a simple example:

```julia
using Zygote

x = rand(5)
y = rand(5)

f(y) = sum([x[i] * y[i] for i in eachindex(y)])

```

Then for my random x,y, I have:

```julia
julia> Zygote.gradient(f, y)
([0.20735224892728188, 0.023609460904363777, 0.8209142299297378, 0.5328550686359217, 0.11713217978556711],)

```

You would get better help here in the forum by posting a self contained minimal reproducible example that produces your error. See [Please read: make it easier to help you](https://discourse.julialang.org/t/psa-make-it-easier-to-help-you/14757) for more details on getting help.

---

<div class="post-metadata">

**Author:** ![rlrs](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rlrs/32/18043_2.png) [@rlrs](https://discourse.julialang.org/u/rlrs)\
**Post date:** [July 1, 2020, 7:08am UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/4 "2020-07-01T07:08:22Z")

</div>

While the Buffer “solution” does work, it complicates things and doesn’t make for very clean code, so I’d still like comprehension in general to work - but in most cases I still can’t make it.  
Here’s an example where I can’t make it work, although it also uses a dictionary which might add another issue to the mix:

```julia

using Zygote

w = randn(5)
grads = gradient(Params(w)) do
    dict = Dict{Int, Float64}(i => v for (i,v) in enumerate(w))
    m = maximum(i*v for (i,v) in dict)
    k = randn(5)
    sum(k .- m)
end

```

results in **MethodError: no method matching getindex(::Dict{Any,Any})**. Besides, you only get this far with the generator syntax since the array comprehension syntax allocates an array and copies data into it, which is unsupported mutation. I’ve tried a bunch of variations on the `maximum()` line to no avail.

---

<div class="post-metadata">

**Author:** ![ianfiske](https://avatars.discourse-cdn.com/v4/letter/i/58f4c7/32.png) [@ianfiske](https://discourse.julialang.org/u/ianfiske)\
**Post date:** [July 1, 2020, 1:10pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/5 "2020-07-01T13:10:57Z")

</div>

Indeed, learning Zygote has involved a bit of stepping on landmines of unsupported operations, but once you learn those bits, it’s quite amazing. Here, your problem is differentiating through indexing into a Dict. Here’s a minimal version that works (no dict)

```julia
using Zygote

w = randn(5)

grads = gradient(Params([w])) do
    m = maximum(i*v for (i,v) in enumerate(w))
    k = randn(5)
    sum(k .- m)
end

```

```julia
julia> grads[w]
5-element Array{Float64,1}:
   0.0
   0.0
   0.0
   0.0
 -50.0

```

Note a couple important changes from your version:

- No Dict. In fact, there is no intermediate data structure to hold the pairs at all.
- `Params(w)` should be `Params([w])` or `Params((w,))` because the argument to `Params` should be an iterable of objects you want to differentiate w.r.t.

---

<div class="post-metadata">

**Author:** ![rlrs](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rlrs/32/18043_2.png) [@rlrs](https://discourse.julialang.org/u/rlrs)\
**Post date:** [July 1, 2020, 5:53pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/6 "2020-07-01T17:53:32Z")

</div>

Is it doable with the dictionary, though? My example is a condensed version of some code I actually have, and the data structure I need to iterate over is indeed a dictionary. Unlike the toy case, I can’t easily replace the dict with something else.

---

<div class="post-metadata">

**Author:** ![ianfiske](https://avatars.discourse-cdn.com/v4/letter/i/58f4c7/32.png) [@ianfiske](https://discourse.julialang.org/u/ianfiske)\
**Post date:** [July 2, 2020, 3:33pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/7 "2020-07-02T15:33:32Z")

</div>

I haven’t been able to get any function that constructs a Dict and then accesses it to work with Zygote. But this hasn’t really come up in any applications that I’ve built. Usually, when I need to create key-value mappings and work with them inside of differentiable functions, I use NamedTuples and it works great.

---

<div class="post-metadata">

**Author:** ![jondeuce](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/jondeuce/32/16378_2.png) [@jondeuce](https://discourse.julialang.org/u/jondeuce)\
**Post date:** [August 30, 2020, 2:46am UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/8 "2020-08-30T02:46:14Z")

</div>

This can be made to work if instead of iterating over the dictionary one iterates over the keys explicitly:

```julia
w = randn(5)
grads = Zygote.gradient(Zygote.Params(w)) do
    dict = Dict{Int, Float64}(i => v for (i,v) in enumerate(w))
    # m = maximum(i*v for (i,v) in dict) # doesn't work
    m = maximum(i -> i * dict[i], eachindex(w)) # works
    k = randn(5)
    sum(k .- m)
end

```

This works because `getindex(d::Dict, ...)` has adjoints defined in Zygote, whereas I’m not sure the iteration protocol does. In general, I have found that iterating over set-like unordered containers like `Dict`s doesn’t tend to work unless you reformulate it into iteration over array-like containers (e.g. iterating over `eachindex(w)` here). I wouldn’t recommend using `Dict`s this way in performance critical areas, but if you are just returning e.g. `Dict`s of loss function values that should be fine.

---

<div class="post-metadata">

**Author:** ![rlrs](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rlrs/32/18043_2.png) [@rlrs](https://discourse.julialang.org/u/rlrs)\
**Post date:** [September 24, 2020, 2:25pm UTC](https://discourse.julialang.org/t/list-comprehension-in-zygote/41881/9 "2020-09-24T14:25:46Z")

</div>

Unfortunately, `eachindex` or `keys` on a dictionary returns a `Base.KeySet` which is also a dictionary-like container (e.g. `map` is undefined). And `collect` is seemingly unsupported since internally it mutates an array.