-
-
Notifications
You must be signed in to change notification settings - Fork 604
/
layers.jl
292 lines (250 loc) · 10.2 KB
/
layers.jl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
# Test layers and data/model movements on and off the GPU
# Add tests for layers and their gradients on the GPU
# Most of the forward passes should be fine being applied
# to bitstype objects, but this gives higher coverage for our use-cases
# Check that getting the gradients does not throw
# generic movement tests
@testset "Basic GPU Movement" begin
@test gradient(x -> sum(gpu(x)), rand(3,3)) isa Tuple
@test gradient(x -> sum(cpu(x)), gpu(rand(3,3))) isa Tuple
end
# TODO: These layers get into scalar indexing issues.
const BROKEN_LAYERS = Union{DepthwiseConv}
const ACTIVATIONS = [identity, relu, tanh,
sigmoid, exp, softplus,
elu, selu]
function gpu_gradtest(name::String, layers::Vector, x_cpu = nothing, args...; test_cpu = true)
isnothing(x_cpu) && error("Missing input to test the layers against.")
@testset "$name GPU grad tests" begin
for layer in layers
@testset "$layer Layer GPU grad test" begin
# compute output and grad of parameters
l_cpu = layer(args...)
ps_cpu = Flux.params(l_cpu)
y_cpu, back_cpu = pullback(() -> sum(l_cpu(x_cpu)), ps_cpu)
gs_cpu = back_cpu(1f0)
x_gpu = gpu(x_cpu)
l_gpu = l_cpu |> gpu
ps_gpu = Flux.params(l_gpu)
if typeof(l_gpu) <: BROKEN_LAYERS
@test_broken gradient(() -> sum(l_gpu(x_gpu)), ps_gpu) isa Flux.Zygote.Grads
else
y_gpu, back_gpu = pullback(() -> sum(l_gpu(x_gpu)), ps_gpu)
gs_gpu = back_gpu(1f0) # TODO many layers error out when backprop int 1, should fix
# compute grad of input
xg_cpu = gradient(x -> sum(l_cpu(x)), x_cpu)[1]
xg_gpu = gradient(x -> sum(l_gpu(x)), x_gpu)[1]
# test
if test_cpu
@test y_gpu ≈ y_cpu rtol=1f-3 atol=1f-3
if isnothing(xg_cpu)
@test isnothing(xg_gpu)
else
if layer === GroupedConvTranspose
@test Array(xg_gpu) ≈ xg_cpu rtol = 2f-2 atol = 1f-3
else
@test Array(xg_gpu) ≈ xg_cpu rtol = 1f-3 atol = 1f-3
end
end
end
@test gs_gpu isa Flux.Zygote.Grads
for (p_cpu, p_gpu) in zip(ps_cpu, ps_gpu)
if isnothing(gs_cpu[p_cpu])
@test isnothing(gs_gpu[p_gpu])
else
@test gs_gpu[p_gpu] isa Flux.CUDA.CuArray
if test_cpu
@test Array(gs_gpu[p_gpu]) ≈ gs_cpu[p_cpu] rtol=1f-3 atol=1f-3
end
end
end
end
end
end
end
end
# Just to give testset in gpu_gradtest meaningful labels
ConvNoBias(args...) = Conv(args...; bias = false)
ConvTransposeNoBias(args...) = ConvTranspose(args...; bias = false)
CrossCorNoBias(args...) = CrossCor(args...; bias = false)
DepthwiseConvNoBias(args...) = DepthwiseConv(args...; bias = false)
GroupedConv(args...) = Conv(args..., groups = 5)
GroupedConvTranspose(args...) = ConvTranspose(args..., groups = 5)
for act in ACTIVATIONS
r = rand(Float32, 28, 28, 1, 1)
conv_layers = [Conv, ConvNoBias,
ConvTranspose, ConvTransposeNoBias,
CrossCor, CrossCorNoBias,
DepthwiseConv, DepthwiseConvNoBias]
gpu_gradtest("Convolution with $act", conv_layers, r, (2,2), 1=>3, act, test_cpu = false)
groupedconv = [GroupedConv, GroupedConvTranspose]
gpu_gradtest("GroupedConvolution with $act", groupedconv, rand(Float32, 28, 28, 100, 2), (3,3), 100 => 25, act, test_cpu = true)
batch_norm = [BatchNorm]
gpu_gradtest("BatchNorm 1 with $act", batch_norm, rand(Float32, 28,28,3,4), 3, act, test_cpu = false) #TODO fix errors
gpu_gradtest("BatchNorm 2 with $act", batch_norm, rand(Float32, 5,4), 5, act, test_cpu = false)
instancenorm = [InstanceNorm]
gpu_gradtest("InstanceNorm with $act", instancenorm, r, 1, act, test_cpu = false)
groupnorm = [GroupNorm]
gpu_gradtest("GroupNorm with $act", groupnorm, rand(Float32, 28,28,3,1), 3, 1, act, test_cpu = false)
end
r = rand(Float32, 28, 28, 1, 1)
pooling_layers = [MaxPool, MeanPool]
gpu_gradtest("Pooling", pooling_layers, r, (2,2))
adaptive_pooling_layers = [AdaptiveMaxPool, AdaptiveMeanPool]
gpu_gradtest("AdaptivePooling", adaptive_pooling_layers, r, (7,7), test_cpu = false)
dropout_layers = [Dropout, AlphaDropout]
gpu_gradtest("Dropout", dropout_layers, r, 0.5f0; test_cpu = false) # dropout is not deterministic
layer_norm = [LayerNorm]
gpu_gradtest("LayerNorm 1", layer_norm, rand(Float32, 28,28,3,4), 1, test_cpu = false) #TODO fix errors
gpu_gradtest("LayerNorm 2", layer_norm, rand(Float32, 5,4), 5)
upsample = [x -> Upsample(scale=x)]
gpu_gradtest("Upsample 2d", upsample, rand(Float32, 3, 4, 2, 3), (2,2))
gpu_gradtest("Upsample 1d", upsample, rand(Float32, 3, 4, 2, 3), (2,))
pixelshuffle = [PixelShuffle]
gpu_gradtest("PixelShuffle 2d", pixelshuffle, rand(Float32, 3, 4, 18, 3), 3)
gpu_gradtest("PixelShuffle 1d", pixelshuffle, rand(Float32, 3, 18, 3), 3)
embedding = [Flux.Embedding]
gpu_gradtest("Embedding", embedding, [1,3,5], 5, 2)
gpu_gradtest("Embedding repeated indices", embedding, [1,3,5,3], 5, 2)
gpu_gradtest("Embedding integer index", embedding, 1, 5, 2)
gpu_gradtest("Embedding 2d index", embedding, [1 2; 3 4], 5, 2)
gpu_gradtest("Embedding OneHotVec index", embedding, OneHotVector(1, 5), 5, 2)
gpu_gradtest("Embedding OneHotMatrix index", embedding, OneHotMatrix([1,2,3], 5), 5, 2)
gpu_gradtest("Embedding OneHotMatrix repeated indices", embedding, OneHotMatrix([1,2,2], 5), 5, 2)
@testset "function layers" begin
x = rand(Float32, 3,3)
gpu_autodiff_test(x -> sum(Flux.normalise(x; dims=1)), x)
gpu_autodiff_test(x -> sum(Flux.normalise(x; dims=2)), x)
gpu_autodiff_test(x -> sum(Flux.normalise(x)), x)
end
@testset "Zeros mapped for $cl" for cl in (Conv, ConvTranspose, CrossCor, DepthwiseConv)
l = cl((2,2), 1=>3, bias = false) |> gpu
ip = zeros(Float32, 28,28,1,1) |> gpu
if typeof(l) <: BROKEN_LAYERS
@test_broken sum(l(ip)) ≈ 0.f0
@test_broken gradient(() -> sum(l(ip)), Flux.params(l)) isa Flux.Zygote.Grads
else
@test sum(l(ip)) ≈ 0.f0
gs = gradient(() -> sum(l(ip)), Flux.params(l))
@test l.bias ∉ gs.params
end
end
@testset "Dense with Zeros bias" begin
l = Dense(ones(Float32, 4, 3), Flux.Zeros()) |> gpu
ip = zeros(Float32, 3, 7) |> gpu
@test sum(l(ip)) ≈ 0.f0
gs = gradient(() -> sum(l(ip)), Flux.params(l))
@test l.bias ∉ gs.params
end
@testset "Extended BatchNorm" begin
m_cpu = BatchNorm(2)
m_gpu = m_cpu |> gpu
x_cpu = rand(Float32, 3, 2, 2)
x_gpu = x_cpu |> gpu
## In :auto mode, track statistics only in gradient contest
μ_cpu = copy(m_cpu.μ)
m_cpu(x_cpu)
@test m_cpu.μ ≈ μ_cpu
gradient(() -> sum(m_cpu(x_cpu)), Flux.params(m_cpu))
@test !(m_cpu.μ ≈ μ_cpu)
μ_gpu = copy(m_gpu.μ)
m_gpu(x_gpu)
@test m_gpu.μ ≈ μ_gpu
gradient(() -> sum(m_gpu(x_gpu)), Flux.params(m_gpu))
@test !(m_gpu.μ ≈ μ_gpu)
@test Array(m_gpu.μ) ≈ m_cpu.μ
## In testmode, never track statistics
testmode!(m_cpu)
μ_cpu = copy(m_cpu.μ)
m_cpu(x_cpu)
@test m_cpu.μ ≈ μ_cpu
gradient(() -> sum(m_cpu(x_cpu)), Flux.params(m_cpu))
@test m_cpu.μ ≈ μ_cpu
testmode!(m_gpu)
μ_gpu = copy(m_gpu.μ)
m_gpu(x_gpu)
@test m_gpu.μ ≈ μ_gpu
gradient(() -> sum(m_gpu(x_gpu)), Flux.params(m_gpu))
@test m_gpu.μ ≈ μ_gpu
## In trainmode, always track statistics
trainmode!(m_cpu)
μ_cpu = copy(m_cpu.μ)
m_cpu(x_cpu)
@test !(m_cpu.μ ≈ μ_cpu)
μ_cpu = copy(m_cpu.μ)
gradient(() -> sum(m_cpu(x_cpu)), Flux.params(m_cpu))
@test !(m_cpu.μ ≈ μ_cpu)
trainmode!(m_gpu)
μ_gpu = copy(m_gpu.μ)
m_gpu(x_gpu)
@test !(m_gpu.μ ≈ μ_gpu)
μ_gpu = copy(m_gpu.μ)
gradient(() -> sum(m_gpu(x_gpu)), Flux.params(m_gpu))
@test !(m_gpu.μ ≈ μ_gpu)
## No errors if input type mistmatch
# x_cpu = rand(Float64, 3, 2, 2)
# x_gpu = x_cpu |> gpu
# m_cpu(x_cpu)
# gradient(() -> sum(m_cpu(x_cpu)), Flux.params(m_cpu))
# m_gpu(x_gpu)
# gradient(() -> sum(m_gpu(x_gpu)), Flux.params(m_gpu))
end
@testset "Two-streams Bilinear" begin
x = zeros(Float32,10,9) |> gpu
y = zeros(Float32,2,9) |> gpu
b = Flux.Bilinear(10, 2, 3) |> gpu
@test size(b(x,y)) == (3,9)
@test sum(abs2, b(x,y)) ≈ 0f0
gs_gpu = gradient(() -> sum(abs2.(b(x, y))), params(b))
b_cpu, x_cpu, y_cpu = b |> cpu, x |> cpu, y |> cpu
gs_cpu = gradient(() -> sum(abs2.(b_cpu(x_cpu, y_cpu))), params(b_cpu))
for (pgpu, pcpu) in zip(params(b), params(b_cpu))
@test gs_cpu[pcpu] ≈ Array(gs_gpu[pgpu])
end
end
@testset "Two-streams Bilinear" begin
x = zeros(Float32,10,9) |> gpu
y = zeros(Float32,2,9) |> gpu
b = Flux.Bilinear(10, 2, 3) |> gpu
@test size(b(x,y)) == (3,9)
@test sum(abs2, b(x,y)) ≈ 0f0
gs_gpu = gradient(() -> sum(abs2.(b(x, y))), params(b))
b_cpu, x_cpu, y_cpu = b |> cpu, x |> cpu, y |> cpu
gs_cpu = gradient(() -> sum(abs2.(b_cpu(x_cpu, y_cpu))), params(b_cpu))
for (pgpu, pcpu) in zip(params(b), params(b_cpu))
@test gs_cpu[pcpu] ≈ Array(gs_gpu[pgpu])
end
end
@testset "Parallel" begin
@testset "zero sum" begin
input = randn(10, 10, 10, 10) |> gpu
layer_gpu = Parallel(+, zero, identity) |> gpu
@test layer_gpu(input) == input
@test layer_gpu(input) isa Flux.CUDA.CuArray
end
@testset "vararg input" begin
inputs = (randn(10), randn(5), randn(4)) .|> gpu
layer = Parallel(+, Dense(10, 2), Dense(5, 2), Dense(4, 2)) |> gpu
@test size(layer(inputs)) == (2,)
end
@testset "gradient" begin
input_cpu = randn(10, 10, 10, 10)
input_gpu = input_cpu |> gpu
layer_cpu = Parallel(+, x -> zero(x), identity)
layer_gpu = layer_cpu |> gpu
gs_cpu = gradient(() -> sum(abs2.(layer_cpu(input_cpu))), params(layer_cpu))
gs_gpu = gradient(() -> sum(abs2.(layer_gpu(input_gpu))), params(layer_gpu))
for (pgpu, pcpu) in zip(params(layer_cpu), params(layer_gpu))
@test gs_cpu[pcpu] ≈ gs_gpu[pgpu]
end
end
end
@testset "Dropout RNGs" begin
@test_throws ArgumentError Flux.dropout(MersenneTwister(), CUDA.rand(Float32, 2, 3), 0.1)
@testset for layer in (Dropout, AlphaDropout)
m = layer(0.1; rng = MersenneTwister(123))
@test_throws ErrorException gpu(m)
m = layer(0.1; rng = CUDA.default_rng())
@test gpu(m).rng isa CUDA.RNG
end
end