|
| 1 | +# Choosing differentiable/gpu parts of the model |
| 2 | +!!! note |
| 3 | + This tutorial features somewhat disconnected topics about customizing your |
| 4 | + models even further. It is advised to be familiar with |
| 5 | + [`Flux.@layer`](@ref), [`Flux.@functor`](@ref), [`freeze!`](@ref |
| 6 | + Flux.freeze!) and other basics of Flux. |
| 7 | + |
| 8 | +Flux provides several ways of freezing, excluding from backprop entirely and |
| 9 | +marking custom struct fields not to be moved to the GPU |
| 10 | +([Functors.@functor](@ref)) hence excluded from being trained. The following |
| 11 | +subsections should make it clear which one suits your needs the best. |
| 12 | + |
| 13 | +## On-the-fly freezing per model instance |
| 14 | +Perhaps you'd like to freeze some of the weights of the model (even at |
| 15 | +mid-training), and Flux accomplishes this through [`freeze!`](@ref Flux.freeze!) and `thaw!`. |
| 16 | + |
| 17 | +```julia |
| 18 | +m = Chain( |
| 19 | + Dense(784 => 64, relu), # freeze this one |
| 20 | + Dense(64 => 64, relu), |
| 21 | + Dense(32 => 10) |
| 22 | + ) |
| 23 | +opt_state = Flux.setup(Momentum(), m); |
| 24 | + |
| 25 | +# Freeze some layers right away |
| 26 | +Flux.freeze!(opt_state.layers[1]) |
| 27 | + |
| 28 | +for data in train_set |
| 29 | + input, label = data |
| 30 | + |
| 31 | + # Some params could be frozen during the training: |
| 32 | + Flux.freeze!(opt_state.layers[2]) |
| 33 | + |
| 34 | + grads = Flux.gradient(m) do m |
| 35 | + result = m(input) |
| 36 | + loss(result, label) |
| 37 | + end |
| 38 | + Flux.update!(opt_state, m, grads[1]) |
| 39 | + |
| 40 | + # Optionally unfreeze the params later |
| 41 | + Flux.thaw!(opt_state.layers[1]) |
| 42 | +end |
| 43 | +``` |
| 44 | + |
| 45 | +## Static freezing per model definition |
| 46 | +Sometimes some parts of the model ([`Flux.@layer`](@ref)) needn't to be trained at all but these params |
| 47 | +still need to reside on the GPU (these params are still needed in the forward |
| 48 | +and/or backward pass). |
| 49 | +```julia |
| 50 | +struct MaskedLayer{T} |
| 51 | + chain::Chain |
| 52 | + mask::T |
| 53 | +end |
| 54 | +Flux.@layer MyLayer trainable=(chain,) |
| 55 | +# mask field will not be updated in the training loop |
| 56 | + |
| 57 | +function (m::MaskedLayer)(x) |
| 58 | + # mask field will still move to to gpu for efficient operations: |
| 59 | + return m.chain(x) + x + m.mask |
| 60 | +end |
| 61 | + |
| 62 | +model = MaskedLayer(...) # this model will not have the `mask` field trained |
| 63 | +``` |
| 64 | +Note how this method permanently sets some model fields to be excluded from |
| 65 | +training without on-the-fly changing. |
| 66 | + |
| 67 | +## Excluding from model definition |
| 68 | +Sometimes some parameters aren't just "not trainable" but they shouldn't even |
| 69 | +transfer to the GPU (or be part of the functor). All scalar fields are like this |
| 70 | +by default, so things like learning rate multipliers are not trainable nor |
| 71 | +transferred to the GPU by default. |
| 72 | +```julia |
| 73 | +struct CustomLayer{T, F} |
| 74 | + chain::Chain |
| 75 | + activation_results::Vector{F} |
| 76 | + lr_multiplier::Float32 |
| 77 | +end |
| 78 | +Flux.@functor CustomLayer (chain, ) # Explicitly leaving out `activation_results` |
| 79 | + |
| 80 | +function (m::CustomLayer)(x) |
| 81 | + result = m.chain(x) + x |
| 82 | + |
| 83 | + # `activation_results` are not part of the GPU loop, hence we could do |
| 84 | + # things like `push!` |
| 85 | + push!(m.activation_results, mean(result)) |
| 86 | + return result |
| 87 | +end |
| 88 | +``` |
| 89 | +See more about this in [`Flux.@functor`](@ref) |
| 90 | + |
| 91 | + |
| 92 | +## Freezing Layer Parameters (deprecated) |
| 93 | + |
| 94 | +When it is desired to not include all the model parameters (for e.g. transfer learning), we can simply not pass in those layers into our call to `params`. |
| 95 | + |
| 96 | +!!! compat "Flux ≤ 0.14" |
| 97 | + The mechanism described here is for Flux's old "implicit" training style. |
| 98 | + When upgrading for Flux 0.15, it should be replaced by [`freeze!`](@ref Flux.freeze!) and `thaw!`. |
| 99 | + |
| 100 | +Consider a simple multi-layer perceptron model where we want to avoid optimising the first two `Dense` layers. We can obtain |
| 101 | +this using the slicing features `Chain` provides: |
| 102 | + |
| 103 | +```julia |
| 104 | +m = Chain( |
| 105 | + Dense(784 => 64, relu), |
| 106 | + Dense(64 => 64, relu), |
| 107 | + Dense(32 => 10) |
| 108 | + ); |
| 109 | + |
| 110 | +ps = Flux.params(m[3:end]) |
| 111 | +``` |
| 112 | + |
| 113 | +The `Zygote.Params` object `ps` now holds a reference to only the parameters of the layers passed to it. |
| 114 | + |
| 115 | +During training, the gradients will only be computed for (and applied to) the last `Dense` layer, therefore only that would have its parameters changed. |
| 116 | + |
| 117 | +`Flux.params` also takes multiple inputs to make it easy to collect parameters from heterogenous models with a single call. A simple demonstration would be if we wanted to omit optimising the second `Dense` layer in the previous example. It would look something like this: |
| 118 | + |
| 119 | +```julia |
| 120 | +Flux.params(m[1], m[3:end]) |
| 121 | +``` |
| 122 | + |
| 123 | +Sometimes, a more fine-tuned control is needed. |
| 124 | +We can freeze a specific parameter of a specific layer which already entered a `Params` object `ps`, |
| 125 | +by simply deleting it from `ps`: |
| 126 | + |
| 127 | +```julia |
| 128 | +ps = Flux.params(m) |
| 129 | +delete!(ps, m[2].bias) |
| 130 | +``` |
| 131 | + |
0 commit comments