You may wish to save models so that they can be loaded and run in a later session. The easiest way to do this is via BSON.jl.
Save a model:
julia> using Flux julia> model = Chain(Dense(10, 5, NNlib.relu), Dense(5, 2), NNlib.softmax) Chain( Dense(10 => 5, relu), # 55 parameters Dense(5 => 2), # 12 parameters NNlib.softmax, ) # Total: 4 arrays, 67 parameters, 524 bytes. julia> using BSON: @save julia> @save "mymodel.bson" model
Load it again:
julia> using Flux # Flux must be loaded before calling @load julia> using BSON: @load julia> @load "mymodel.bson" model julia> model Chain( Dense(10 => 5, relu), # 55 parameters Dense(5 => 2), # 12 parameters NNlib.softmax, ) # Total: 4 arrays, 67 parameters, 524 bytes.
Models are just normal Julia structs, so it's fine to use any Julia storage format for this purpose. BSON.jl is particularly well supported and most likely to be forwards compatible (that is, models saved now will load in future versions of Flux).
If a saved model's parameters are stored on the GPU, the model will not load later on if there is no GPU support available. It's best to move your model to the CPU with
cpu(model) before saving it.
Previous versions of Flux suggested saving only the model weights using
@save "mymodel.bson" params(model). This is no longer recommended and even strongly discouraged. Saving models this way will only store the trainable parameters which will result in incorrect behavior for layers like
julia> using Flux julia> model = Chain(Dense(10 => 5,relu),Dense(5 => 2),softmax) Chain( Dense(10 => 5, relu), # 55 parameters Dense(5 => 2), # 12 parameters NNlib.softmax, ) # Total: 4 arrays, 67 parameters, 524 bytes. julia> weights = Flux.params(model);
Loading the model as shown above will return a new model with the stored parameters. But sometimes you already have a model, and you want to load stored parameters into it. This can be done as
using Flux: loadmodel! using BSON: @load # some predefined model model = Chain(Dense(10 => 5, relu), Dense(5 => 2), softmax) # load one model into another model = loadmodel!(model, @load("mymodel.bson"))
This ensures that the model loaded from
"mymodel.bson" matches the structure of
Flux.loadmodel! is also convenient for copying parameters between models in memory.
Copy all the parameters (trainable and non-trainable) from
src together using
Functors.children, and calling
copyto! on parameter arrays or throwing an error when there is a mismatch. Non-array elements (such as activation functions) are not copied and need not match. Zero bias vectors and
bias=false are considered equivalent (see extended help for more details).
julia> dst = Chain(Dense(Flux.ones32(2, 5, tanh)), Dense(2 => 1; bias = [1f0])) Chain( Dense(5 => 2, tanh), # 12 parameters Dense(2 => 1), # 3 parameters ) # Total: 4 arrays, 15 parameters, 316 bytes. julia> dst.weight ≈ ones(2, 5) # by construction true julia> src = Chain(Dense(5 => 2, relu), Dense(2 => 1, bias=false)); julia> Flux.loadmodel!(dst, src); julia> dst.weight ≈ ones(2, 5) # values changed false julia> iszero(dst.bias) true
Throws an error when:
srcdo not share the same fields (at any level)
- the sizes of leaf nodes are mismatched between
- copying non-array values to/from an array parameter (except inactive parameters described below)
dstis a "tied" parameter (i.e. refers to another parameter) and loaded into multiple times with mismatched source values
Inactive parameters can be encoded by using the boolean value
false instead of an array. If
dst == false and
src is an all-zero array, no error will be raised (and no values copied); however, attempting to copy a non-zero array to an inactive parameter will throw an error. Likewise, copying a
src value of
false to any
dst array is valid, but copying a
src value of
true will error.
In longer training runs it's a good idea to periodically save your model, so that you can resume if training is interrupted (for example, if there's a power cut). You can do this by saving the model in the callback provided to
julia> using Flux: throttle julia> using BSON: @save julia> m = Chain(Dense(10 => 5, relu), Dense(5 => 2), softmax) Chain( Dense(10 => 5, relu), # 55 parameters Dense(5 => 2), # 12 parameters NNlib.softmax, ) # Total: 4 arrays, 67 parameters, 524 bytes. julia> evalcb = throttle(30) do # Show loss @save "model-checkpoint.bson" model end;
This will update the
"model-checkpoint.bson" file every thirty seconds.
You can get more advanced by saving a series of models throughout training, for example
@save "model-$(now()).bson" model
will produce a series of models like
"model-2018-03-06T02:57:10.41.bson". You could also store the current test set loss, so that it's easy to (for example) revert to an older copy of the model if it starts to overfit.
@save "model-$(now()).bson" model loss = testloss()
Note that to resume a model's training, you might need to restore other stateful parts of your training loop. Possible examples are stateful optimizers (which usually utilize an
IdDict to store their state), and the randomness used to partition the original data into the training and validation sets.
You can store the optimiser state alongside the model, to resume training exactly where you left off. BSON is smart enough to cache values and insert links when saving, but only if it knows everything to be saved up front. Thus models and optimizers must be saved together to have the latter work after restoring.
opt = Adam() @save "model-$(now()).bson" model opt