This repository has been archived by the owner on Aug 16, 2020. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMovieLens_Autodiff.jl
181 lines (123 loc) · 5.85 KB
/
MovieLens_Autodiff.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
# using ZipFile,
using CSV, DataFrames
using StatsBase, Random, Distributions
using Flux
# Import ratings file
cd(raw"/home/emmanuel/Development/Learning/capstone-movielens/")
# z = ZipFile.Reader("datasets/ratings.zip")
# ratings = CSV.read(z.files[1];
# delim="::",
# header=["userId", "movieId", "rating", "timestamp"],
# types = [Int, Int, Float64, Int])
edx = CSV.read("datasets/edx.csv";
delim=",",
copycols=true,
datarow=2,
header=["rowN", "userId", "movieId", "rating", "timestamp", "title", "genres"],
types = [Int, Int, Int, Float64, Int, String, String])
validation = CSV.read("datasets/validation.csv";
delim=",",
copycols=true,
datarow=2,
header=["rowN", "userId", "movieId", "rating", "timestamp", "title", "genres"],
types = [Int, Int, Int, Float64, Int, String, String])
function stochastic_grad_descent_simple(P::Array{Float64, 2}, Q::Array{Float64, 2};
times = 1, batch_size = 10000, λ = 0.1, α = 0.01, verbose = true)
for i = 1:times
spl_training = view(training, sample(axes(training, 1), batch_size; replace = false, ordered = true), :)
spl_P = P[spl_training[:,:userN], :]
spl_Q = Q[spl_training[:,:movieN], :]
err = spl_training.rating_z - sum(spl_P .* spl_Q; dims = 2)
ΔP = - 2 * err .* spl_Q + λ * spl_P
ΔQ = - 2 * err .* spl_P + λ * spl_Q
P[spl_training[:,:userN], :] = spl_P - α * ΔP
Q[spl_training[:,:movieN], :] = spl_Q - α * ΔQ
end # for loop
return P, Q
end
r_m, r_sd = mean_and_std(edx.rating)
edx.rating_z = (edx.rating .- r_m) ./ r_sd
validation.rating_z = (validation.rating .- r_m) ./ r_sd
movieIndex = DataFrame(movieId = unique(edx, :movieId).movieId)
nMovie = size(movieIndex)[1]
movieIndex.movieN = 1:nMovie
userIndex = DataFrame(userId = unique(edx, :userId).userId)
nUser = size(userIndex)[1]
userIndex.userN = 1:nUser
movieMean = by(edx, :movieId, movie_mean = :rating_z => mean)
userMean = by(edx, :userId, user_mean = :rating_z => mean)
training = join(edx, movieIndex, on = :movieId )
training = join(training, userIndex, on = :userId )
training = training[:, [:userN, :movieN, :rating_z]]
nSamples = size(training)[1]
test = join(validation, movieIndex, on = :movieId )
test = join(test, userIndex, on = :userId )
test = test[:, [:userN, :movieN, :rating, :rating_z]]
nTest = size(test)[1]
nLF = 3
P = zeros(Float64, nUser, nLF)
Q = zeros(Float64, nMovie, nLF)
P[:,1] .= 1
P[:,2] = sort(join(userIndex, userMean, on = :userId, kind = :left), :userN)[:, :user_mean]
Q[:,1] = sort(join(movieIndex, movieMean, on = :movieId, kind = :left), :movieN)[:, :movie_mean]
Q[:,2] .= 1
Random.seed!(42)
P[:, 3] = rand(Normal(), nUser) / 1000
Q[:, 3] = rand(Normal(), nMovie) / 1000
# Cost function
function J(P::Array{Float64, 2}, Q::Array{Float64, 2})
error = sum( P[test[:,:userN], :] .* Q[test[:,:movieN], :]; dims = 2) .* r_sd .+ r_m
error = sqrt( sum((test.rating - error).^2) / nTest )
return error
end
function stochastic_grad_descent(P::Array{Float64, 2}, Q::Array{Float64, 2}, grad;
times = 1, batch_size = 10000, λ = 0.1, α = 0.01, verbose = true)
for i = 1:times
spl_training = view(training, sample(axes(training, 1), batch_size; replace = false, ordered = true), :)
spl_P = P[spl_training[:,:userN], :]
spl_Q = Q[spl_training[:,:movieN], :]
spl_P = P[spl_training[:,:userN], :]
spl_Q = Q[spl_training[:,:movieN], :]
# Gradient
(ΔP, ΔQ) = gradient((p, q) -> J(p, q), P, Q)[spl_P, spl_Q]
P[spl_training[:,:userN], :] = spl_P - α * ΔP
Q[spl_training[:,:movieN], :] = spl_Q - α * ΔQ
end # for loop
return P, Q
end
float_t = sum( P[test[:,:userN], :] .* Q[test[:,:movieN], :]; dims = 2) .* r_sd .+ r_m
float_t = sqrt( sum((test.rating - float_t).^2) / nTest )
println("Initial floating point RMSE test = ",float_t)
round_t = round.(P[test[:,:userN], :] .* Q[test[:,:movieN], :])
round_t = sum(round_t ; dims = 2) .* r_sd .+ r_m
round_t = sqrt( sum((test.rating - round_t).^2) / nTest )
println("Initial rounded RMSE test = ", round_t)
starting_α = 0.01
validation_results =[0 float_t round_t]
for n = 1:500
batch_size = 10000
α = starting_α
nFeatures = size(P)[2]
λ = 0.1 * (nUser + nMovie) * nFeatures / 2_000_000
RMSE = sqrt(sum((training.rating_z -
sum( P[training[:,:userN], :] .* Q[training[:,:movieN], :]; dims = 2)).^2) / nSamples)
for i = 1:250
RMSE_tmp = RMSE
(newP, newQ) = stochastic_grad_descent(P, Q; times = 100, batch_size = 1000 * nFeatures, λ = λ, α = α)
global P = newP
global Q = newQ
RMSE = sqrt(sum((training.rating_z -
sum( P[training[:,:userN], :] .* Q[training[:,:movieN], :]; dims = 2)).^2) / nSamples)
println("Step: ", i, " RMSE of z-score training = ", RMSE)
if (RMSE > RMSE_tmp)
α /= 2
println("α decreased to ", α)
end
if (starting_α / α > 10000) | (abs((RMSE - RMSE_tmp) / RMSE_tmp) < 1e-6)
break
end
end
test_RMSE = P[test[:,:userN], :] .* Q[test[:,:movieN], :]
# floating point ratings float_RMSE = sum(test_RMSE ; dims = 2).* r_sd .+ r_m float_RMSE = sqrt( sum((test.rating - float_RMSE).^2) / nTest ) println("Step: ", n, " float RMSE test = ",float_RMSE, " with number features = ", nFeatures) # Round to only obtain legal ratings round_RMSE = sum( round.(test_RMSE) ; dims = 2).* r_sd .+ r_m round_RMSE = sqrt( sum((test.rating - round_RMSE).^2) / nTest ) println("Step: ", n, " round RMSE test = ",round_RMSE, " with number features = ", nFeatures) global validation_results = [validation_results; nFeatures float_RMSE round_RMSE] # Add 1 features global P = [P rand(Normal(), nUser)/1000] global Q = [Q rand(Normal(), nMovie)/1000]
nFeatures += 1
end