-
Notifications
You must be signed in to change notification settings - Fork 254
/
lrn.jl
192 lines (159 loc) · 6.29 KB
/
lrn.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
function test_lrn_layer(backend::Backend, mode::LRNModeType, tensor_dim, T, eps)
println("-- Testing LRN($(typeof(mode))) on $(typeof(backend)){$T}...")
dims = tuple((abs(rand(Int,tensor_dim)) % 6 + 6)...)
op_dim = max(abs(rand(Int)) % tensor_dim, 1)
println(" > Setup with dims $dims")
input = rand(T, dims)
input_blobs = Blob[make_blob(backend, input)]
diff_blobs = Blob[make_blob(backend, input)]
layer = LRNLayer(tops=[:output],bottoms=[:input], mode=mode, channel_dim=op_dim)
state = setup(backend, layer, input_blobs, diff_blobs)
@test size(state.blobs[1]) == size(input)
println(" > Forward")
forward(backend, state, input_blobs)
expected_output = lrn_forward(input, state, op_dim)
got_output = similar(input)
copy!(got_output, state.blobs[1])
@test all(abs(got_output - expected_output) .< eps)
println(" > Backward")
top_diff = rand(T, size(input))
copy!(state.blobs_diff[1], top_diff)
backward(backend, state, input_blobs, diff_blobs)
got_grad = zeros(T, size(input))
copy!(got_grad, diff_blobs[1])
expected_grad = lrn_backward(input, top_diff, state, op_dim)
@test all(abs(got_grad - expected_grad) .< eps)
shutdown(backend, state)
end
function lrn_forward_across_channel{T}(input::Array{T}, state, op_dim)
output = similar(input)
pre_dim, chann_dim, post_dim = split_dims(input, op_dim)
pre_pad = div(state.layer.kernel-1,2)
post_pad = state.layer.kernel - pre_pad - 1
canonical_input = reshape(input, (pre_dim, chann_dim, post_dim))
canonical_output = reshape(output, (pre_dim, chann_dim, post_dim))
for n = 1:post_dim
for c = 1:chann_dim
cstart = c-pre_pad
cend = min(c + post_pad, chann_dim)
cstart = max(1, cstart)
tmp = canonical_input[:,cstart:cend,n].^2 * (state.layer.scale / state.layer.kernel)
tmp = (sum(tmp, 2) + state.layer.shift) .^ state.layer.power
canonical_output[:,c,n] = canonical_input[:,c,n] ./ tmp
end
end
return output
end
function lrn_forward_within_channel{T}(input::Array{T}, state)
output = similar(input)
width, height, channels, num = size(input)
pooled_width = width; pooled_height = height
kernel_size = state.layer.kernel^2
pre_pad = div(state.layer.kernel-1,2)
for n = 1:num
for c = 1:channels
for ph = 1:pooled_height
for pw = 1:pooled_width
hstart = ph - pre_pad
wstart = pw - pre_pad
hend = min(hstart + state.layer.kernel - 1, height)
wend = min(wstart + state.layer.kernel - 1, width)
hstart = max(1, hstart)
wstart = max(1, wstart)
tmp = (input[wstart:wend,hstart:hend,c,n]).^2 * state.layer.scale / kernel_size
tmp = (sum(tmp) + state.layer.shift) .^ state.layer.power
output[pw,ph,c,n] = input[pw,ph,c,n] / tmp
end
end
end
end
return output
end
function lrn_forward{T}(input::Array{T}, state, op_dim)
if isa(state.layer.mode, LRNMode.AcrossChannel)
lrn_forward_across_channel(input, state, op_dim)
elseif isa(state.layer.mode, LRNMode.WithinChannel)
lrn_forward_within_channel(input, state)
else
error("Unknown LRN-mode $(state.layer.mode)")
end
end
function lrn_backward_across_channel{T}(input::Array{T}, top_diff::Array{T}, state, op_dim)
output = zeros(T, size(input))
pre_dim, chann_dim, post_dim = split_dims(input, op_dim)
pre_pad = div(state.layer.kernel-1,2)
post_pad = state.layer.kernel - pre_pad - 1
canonical_input = reshape(input, (pre_dim, chann_dim, post_dim))
canonical_output = reshape(output, (pre_dim, chann_dim, post_dim))
canonical_diff = reshape(top_diff, (pre_dim, chann_dim, post_dim))
for n = 1:post_dim
for c = 1:chann_dim
cstart = c-pre_pad
cend = min(c + post_pad, chann_dim)
cstart = max(1, cstart)
tmp = canonical_input[:,cstart:cend,n].^2 * (state.layer.scale / state.layer.kernel)
tmp = (sum(tmp, 2) + state.layer.shift)
canonical_output[:,c,n] += tmp .^ (-state.layer.power) .* canonical_diff[:,c,n]
tmp = -state.layer.power * tmp .^ (-state.layer.power - 1)
tmp = 2 * state.layer.scale / state.layer.kernel * tmp
canonical_output[:,cstart:cend,n] += tmp .* canonical_input[:,cstart:cend,n] .*
canonical_input[:,c,n] .* canonical_diff[:,c,n]
end
end
return output
end
function lrn_backward_within_channel{T}(input::Array{T}, top_diff::Array{T}, state)
output = zeros(T, size(input))
width, height, channels, num = size(input)
pooled_width = width; pooled_height = height
kernel_size = state.layer.kernel^2
pre_pad = div(state.layer.kernel-1,2)
for n = 1:num
for c = 1:channels
for ph = 1:pooled_height
for pw = 1:pooled_width
hstart = ph - pre_pad
wstart = pw - pre_pad
hend = min(hstart + state.layer.kernel - 1, height)
wend = min(wstart + state.layer.kernel - 1, width)
hstart = max(1, hstart)
wstart = max(1, wstart)
tmp = (input[wstart:wend,hstart:hend,c,n]).^2 * state.layer.scale / kernel_size
tmp = (sum(tmp) + state.layer.shift)
output[pw,ph,c,n] += tmp .^ (-state.layer.power) * top_diff[pw,ph,c,n]
tmp = -state.layer.power * tmp .^ (-state.layer.power-1)
tmp = 2 * state.layer.scale / kernel_size * tmp
output[wstart:wend, hstart:hend, c, n] +=
tmp * input[wstart:wend, hstart:hend, c, n] * input[pw,ph,c,n] * top_diff[pw,ph,c,n]
end
end
end
end
return output
end
function lrn_backward{T}(input::Array{T}, top_diff::Array{T}, state, op_dim)
if isa(state.layer.mode, LRNMode.AcrossChannel)
lrn_backward_across_channel(input, top_diff, state, op_dim)
elseif isa(state.layer.mode, LRNMode.WithinChannel)
lrn_backward_within_channel(input, top_diff, state)
else
error("Unknown LRN-mode $(state.layer.mode)")
end
end
function test_lrn_layer(backend::Backend, mode::LRNModeType, T, eps)
test_lrn_layer(backend, mode, 4, T, eps)
end
function test_lrn_layer(backend::Backend, T, eps)
test_lrn_layer(backend, LRNMode.AcrossChannel(), T, eps)
test_lrn_layer(backend, LRNMode.WithinChannel(), T, eps)
end
function test_lrn_layer(backend::Backend)
test_lrn_layer(backend, Float32, 1e-3)
test_lrn_layer(backend, Float64, 1e-9)
end
if test_cpu
test_lrn_layer(backend_cpu)
end
if test_gpu
test_lrn_layer(backend_gpu)
end