search

chengchingwen / Transformers.jl

Julia Implementation of Transformer models
MIT License
523 stars 74 forks source link

MethodError: no method matching AbstractFloat(::Type{Any}) with HuggingFace.GPT2 #65

Closed AdarshKumar712 closed 3 years ago

AdarshKumar712 commented 3 years ago

I was working with HuggingFace GPT2. I am trying to differentiate some loss on logits of GPT2 with respect to past_key_values pre-calculated. While the forward pass works correctly, I am getting the following error in backward pass:

MethodError: no method matching AbstractFloat(::Type{Any})
Closest candidates are:
  (::Type{T})(::AbstractChar) where T<:Union{AbstractChar, Number} at char.jl:50
  (::Type{T})(::Base.TwicePrecision) where T<:Number at twiceprecision.jl:243
  (::Type{T})(::Flux.NilNumber.Nil) where T<:Number at /home/adarshkumar712/.julia/packages/Flux/0c9kI/src/outputsize.jl:17
  ...

Stacktrace:
  [1] float(x::Type)
    @ Base ./float.jl:206
  [2] (::ComposedFunction{typeof(float), typeof(eltype)})(x::Function)
    @ Base ./operators.jl:938
  [3] softmax(x::Function; dims::Int64)
    @ NNlib ~/.julia/packages/NNlib/3MZcC/src/softmax.jl:48
  [4] rrule(::typeof(softmax), xs::Function; dims::Int64)
    @ NNlib ~/.julia/packages/NNlib/3MZcC/src/softmax.jl:80
  [5] chain_rrule_kw
    @ ~/.julia/packages/Zygote/zowrf/src/compiler/chainrules.jl:101 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0 [inlined]
  [7] _pullback(::Zygote.Context, ::NNlib.var"#softmax##kw", ::NamedTuple{(:dims,), Tuple{Int64}}, ::typeof(softmax), ::typeof(Transformers.HuggingFace.apply_shift_mask))
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:9
  [8] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:124 [inlined]
  [9] _pullback(::Zygote.Context, ::typeof(Transformers.HuggingFace._attn), ::Array{Float32, 4}, ::Array{Float32, 4}, ::Array{Float32, 4}, ::Transformers.HuggingFace.ShiftAttentionMask{Array{Float32, 4}})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [10] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:138 [inlined]
 [11] _pullback(::Zygote.Context, ::Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, ::Array{Float32, 4}, ::Array{Float32, 4}, ::Array{Float32, 4}, ::Transformers.HuggingFace.ShiftAttentionMask{Array{Float32, 4}}, ::Val{false}, ::Val{true})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [12] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:97 [inlined]
 [13] _pullback(::Zygote.Context, ::Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, ::Array{Float32, 3}, ::Tuple{Array{Float32, 4}, Array{Float32, 4}}, ::Transformers.HuggingFace.ShiftAttentionMask{Array{Float32, 4}}, ::Val{false}, ::Val{true})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [14] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:224 [inlined]
 [15] _pullback(::Zygote.Context, ::Transformers.HuggingFace.HGFGPT2Block{Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.HGFGPT2MLP{typeof(gelu), Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}}, ::Array{Float32, 3}, ::Tuple{Array{Float32, 4}, Array{Float32, 4}}, ::Transformers.HuggingFace.ShiftAttentionMask{Array{Float32, 4}}, ::Val{false}, ::Val{true})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [16] macro expansion
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:413 [inlined]
 [17] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:326 [inlined]
 [18] _pullback(::Zygote.Context, ::Transformers.HuggingFace.HGFGPT2Model{12, Transformers.HuggingFace.FakeTHModuleList{12, NTuple{12, Transformers.HuggingFace.HGFGPT2Block{Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.HGFGPT2MLP{typeof(gelu), Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}}}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}}, ::Matrix{Int64}, ::Nothing, ::Nothing, ::NTuple{12, Tuple{Array{Float32, 4}, Array{Float32, 4}}}, ::Nothing, ::Val{false}, ::Val{true}, ::Val{true})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [19] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:473 [inlined]
 [20] _pullback(::Zygote.Context, ::Transformers.HuggingFace.HGFGPT2LMHeadModel{Transformers.HuggingFace.HGFGPT2Model{12, Transformers.HuggingFace.FakeTHModuleList{12, NTuple{12, Transformers.HuggingFace.HGFGPT2Block{Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.HGFGPT2MLP{typeof(gelu), Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}}}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}}, Transformers.HuggingFace.FakeTHLinear{LinearAlgebra.Transpose{Float32, Matrix{Float32}}, Nothing}}, ::Matrix{Int64}, ::Nothing, ::Nothing, ::NTuple{12, Tuple{Array{Float32, 4}, Array{Float32, 4}}}, ::Nothing, ::Val{false}, ::Val{true}, ::Val{true})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [21] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:449 [inlined]
 [22] _pullback(::Zygote.Context, ::Transformers.HuggingFace.var"##_#201", ::Nothing, ::Nothing, ::NTuple{12, Tuple{Array{Float32, 4}, Array{Float32, 4}}}, ::Nothing, ::Bool, ::Bool, ::Bool, ::Transformers.HuggingFace.HGFGPT2LMHeadModel{Transformers.HuggingFace.HGFGPT2Model{12, Transformers.HuggingFace.FakeTHModuleList{12, NTuple{12, Transformers.HuggingFace.HGFGPT2Block{Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.HGFGPT2MLP{typeof(gelu), Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}}}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}}, Transformers.HuggingFace.FakeTHLinear{LinearAlgebra.Transpose{Float32, Matrix{Float32}}, Nothing}}, ::Matrix{Int64})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [23] _pullback
    @ ~/.julia/packages/Transformers/rCnGb/src/huggingface/models/gpt2.jl:449 [inlined]
 [24] _pullback(::Zygote.Context, ::Core.var"#Any##kw", ::NamedTuple{(:past_key_values, :output_attentions, :output_hidden_states, :use_cache), Tuple{NTuple{12, Tuple{Array{Float32, 4}, Array{Float32, 4}}}, Bool, Bool, Bool}}, ::Transformers.HuggingFace.HGFGPT2LMHeadModel{Transformers.HuggingFace.HGFGPT2Model{12, Transformers.HuggingFace.FakeTHModuleList{12, NTuple{12, Transformers.HuggingFace.HGFGPT2Block{Transformers.HuggingFace.HGFGPT2Attention{Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}, Transformers.HuggingFace.HGFGPT2MLP{typeof(gelu), Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}, Transformers.HuggingFace.FakeHGFConv1D{Matrix{Float32}, Vector{Float32}}}}}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHEmbedding{Matrix{Float32}}, Transformers.HuggingFace.FakeTHLayerNorm{Vector{Float32}}}, Transformers.HuggingFace.FakeTHLinear{LinearAlgebra.Transpose{Float32, Matrix{Float32}}, Nothing}}, ::Matrix{Int64})
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [25] _pullback
    @ ./In[8]:3 [inlined]
 [26] _pullback(::Zygote.Context, ::var"#1#2")
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface2.jl:0
 [27] pullback(f::Function, ps::Params)
    @ Zygote ~/.julia/packages/Zygote/zowrf/src/compiler/interface.jl:250
 [28] top-level scope
    @ In[8]:2
 [29] eval
    @ ./boot.jl:360 [inlined]
 [30] include_string(mapexpr::typeof(REPL.softscope), mod::Module, code::String, filename::String)
    @ Base ./loading.jl:1094

I believe that this is something related to the ShiftAttentionMask. Here is the MWE:

using Transformers.HuggingFace
using Zygote, Flux
model = hgf"gpt2:lmheadmodel"
tokens = reshape(Array(1:10),(:,1));
outputs = model(tokens[1:end-1, :]; position_ids=nothing, token_type_ids=nothing,
                                    past_key_values=nothing,
                                    attention_mask=nothing,
                                    output_attentions=true,
                                    output_hidden_states=true,
                                    use_cache=true);
past = outputs.past_key_values
prev = tokens[end:end, :];
ps = params(past)
_, back = Zygote.pullback(ps) do
    output_1 = model(prev; past_key_values=past,
                        output_attentions=false,
                        output_hidden_states=true,
                        use_cache=true);
    hidden = output_1.hidden_states[end]
    logits = model.lm_head(hidden)[:, end, :]
    logits[1]
end

Please let me know, if I am doing something wrong in this code. cc @chengchingwen

chengchingwen commented 3 years ago

Should be fixed with the new release (v0.1.13). Let me know if the problem remains.