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
|
---
title: "XGBoost from JSON"
output:
rmarkdown::html_vignette:
number_sections: yes
toc: yes
author: Roland Stevenson
vignette: >
%\VignetteIndexEntry{XGBoost from JSON}
%\VignetteEngine{knitr::rmarkdown}
\usepackage[utf8]{inputenc}
---
XGBoost from JSON
=================
## Introduction
The purpose of this Vignette is to show you how to correctly load and work with an **Xgboost** model that has been dumped to JSON. **Xgboost** internally converts all data to [32-bit floats](https://en.wikipedia.org/wiki/Single-precision_floating-point_format), and the values dumped to JSON are decimal representations of these values. When working with a model that has been parsed from a JSON file, care must be taken to correctly treat:
- the input data, which should be converted to 32-bit floats
- any 32-bit floats that were stored in JSON as decimal representations
- any calculations must be done with 32-bit mathematical operators
## Setup
For the purpose of this tutorial we will load the xgboost, jsonlite, and float packages. We'll also set `digits=22` in our options in case we want to inspect many digits of our results.
```{r}
require(xgboost)
require(jsonlite)
require(float)
options(digits=22)
```
We will create a toy binary logistic model based on the example first provided [here](https://github.com/dmlc/xgboost/issues/3960), so that we can easily understand the structure of the dumped JSON model object. This will allow us to understand where discrepancies can occur and how they should be handled.
```{r}
dates <- c(20180130, 20180130, 20180130,
20180130, 20180130, 20180130,
20180131, 20180131, 20180131,
20180131, 20180131, 20180131,
20180131, 20180131, 20180131,
20180134, 20180134, 20180134)
labels <- c(1, 1, 1,
1, 1, 1,
0, 0, 0,
0, 0, 0,
0, 0, 0,
0, 0, 0)
data <- data.frame(dates = dates, labels=labels)
bst <- xgboost(
data = as.matrix(data$dates),
label = labels,
nthread = 2,
nrounds = 1,
objective = "binary:logistic",
missing = NA,
max_depth = 1
)
```
## Comparing results
We will now dump the model to JSON and attempt to illustrate a variety of issues that can arise, and how to properly deal with them.
First let's dump the model to JSON:
```{r}
bst_json <- xgb.dump(bst, with_stats = FALSE, dump_format='json')
bst_from_json <- fromJSON(bst_json, simplifyDataFrame = FALSE)
node <- bst_from_json[[1]]
cat(bst_json)
```
The tree JSON shown by the above code-chunk tells us that if the data is less than 20180132, the tree will output the value in the first leaf. Otherwise it will output the value in the second leaf. Let's try to reproduce this manually with the data we have and confirm that it matches the model predictions we've already calculated.
```{r}
bst_preds_logodds <- predict(bst,as.matrix(data$dates), outputmargin = TRUE)
# calculate the logodds values using the JSON representation
bst_from_json_logodds <- ifelse(data$dates<node$split_condition,
node$children[[1]]$leaf,
node$children[[2]]$leaf)
bst_preds_logodds
bst_from_json_logodds
# test that values are equal
bst_preds_logodds == bst_from_json_logodds
```
None are equal. What happened?
At this stage two things happened:
- input data was not converted to 32-bit floats
- the JSON variables were not converted to 32-bit floats
### Lesson 1: All data is 32-bit floats
> When working with imported JSON, all data must be converted to 32-bit floats
To explain this, let's repeat the comparison and round to two decimals:
```{r}
round(bst_preds_logodds,2) == round(bst_from_json_logodds,2)
```
If we round to two decimals, we see that only the elements related to data values of `20180131` don't agree. If we convert the data to floats, they agree:
```{r}
# now convert the dates to floats first
bst_from_json_logodds <- ifelse(fl(data$dates)<node$split_condition,
node$children[[1]]$leaf,
node$children[[2]]$leaf)
# test that values are equal
round(bst_preds_logodds,2) == round(bst_from_json_logodds,2)
```
What's the lesson? If we are going to work with an imported JSON model, any data must be converted to floats first. In this case, since '20180131' cannot be represented as a 32-bit float, it is rounded up to 20180132, as shown here:
```{r}
fl(20180131)
```
### Lesson 2: JSON parameters are 32-bit floats
> All JSON parameters stored as floats must be converted to floats.
Let's now say we do care about numbers past the first two decimals.
```{r}
# test that values are equal
bst_preds_logodds == bst_from_json_logodds
```
None are exactly equal. What happened? Although we've converted the data to 32-bit floats, we also need to convert the JSON parameters to 32-bit floats. Let's do this:
```{r}
# now convert the dates to floats first
bst_from_json_logodds <- ifelse(fl(data$dates)<fl(node$split_condition),
as.numeric(fl(node$children[[1]]$leaf)),
as.numeric(fl(node$children[[2]]$leaf)))
# test that values are equal
bst_preds_logodds == bst_from_json_logodds
```
All equal. What's the lesson? If we are going to work with an imported JSON model, any JSON parameters that were stored as floats must also be converted to floats first.
### Lesson 3: Use 32-bit math
> Always use 32-bit numbers and operators
We were able to get the log-odds to agree, so now let's manually calculate the sigmoid of the log-odds. This should agree with the xgboost predictions.
```{r}
bst_preds <- predict(bst,as.matrix(data$dates))
# calculate the predictions casting doubles to floats
bst_from_json_preds <- ifelse(fl(data$dates)<fl(node$split_condition),
as.numeric(1/(1+exp(-1*fl(node$children[[1]]$leaf)))),
as.numeric(1/(1+exp(-1*fl(node$children[[2]]$leaf))))
)
# test that values are equal
bst_preds == bst_from_json_preds
```
None are exactly equal again. What is going on here? Well, since we are using the value `1` in the calcuations, we have introduced a double into the calculation. Because of this, all float values are promoted to 64-bit doubles and the 64-bit version of the exponentiation operator `exp` is also used. On the other hand, xgboost uses the 32-bit version of the exponentation operator in its [sigmoid function](https://github.com/dmlc/xgboost/blob/54980b8959680a0da06a3fc0ec776e47c8cbb0a1/src/common/math.h#L25-L27).
How do we fix this? We have to ensure we use the correct datatypes everywhere and the correct operators. If we use only floats, the float library that we have loaded will ensure the 32-bit float exponention operator is applied.
```{r}
# calculate the predictions casting doubles to floats
bst_from_json_preds <- ifelse(fl(data$dates)<fl(node$split_condition),
as.numeric(fl(1)/(fl(1)+exp(fl(-1)*fl(node$children[[1]]$leaf)))),
as.numeric(fl(1)/(fl(1)+exp(fl(-1)*fl(node$children[[2]]$leaf))))
)
# test that values are equal
bst_preds == bst_from_json_preds
```
All equal. What's the lesson? We have to ensure that all calculations are done with 32-bit floating point operators if we want to reproduce the results that we see with xgboost.
|
