The task is simple: I have about 1,800 .text data files downloaded from British Oceanographic Data Centre (BODC). They are the historical tidal data and are separated by year and by port. I need to combine all the data into one giant table in R, and save it later for modelling.

One sample data file looks like this:

Port:              P035
Site:              Wick
Latitude:          58.44097
Longitude:         -3.08631
Start Date:        01JAN1985-00.00.00
End Date:          03OCT1985-19.00.00
Contributor:       National Oceanography Centre, Liverpool
Datum information: The data refer to Admiralty Chart Datum (ACD)
Parameter code:    ASLVZZ01 = Surface elevation (unspecified datum) of the water body                      
  Cycle    Date      Time      ASLVZZ01     Residual  
 Number yyyy mm dd hh mi ssf           f            f 
     1) 1985/01/01 00:00:00      1.0300      -0.3845  
     2) 1985/01/01 01:00:00      1.0400      -0.3884  
     3) 1985/01/01 02:00:00      1.2000      -0.3666

The first 9 lines are the metadata, which describes the port ID, name and location of the port, and other information about the data. The line 10 and 11 are the headers of the data matrix.

After the glimpse of the data sample, my first thought was to skip the first 12 lines and treat the rest as a regular data files that has space as separator. It can be easily done by using read.table() with skip = 12 option.

read.table(data.file, skip = 12) ## error

It turned out this approach won't work for some files because when the way of measuring tidal were changed, the date and port were highlighted, leaving a second chunk of data matrix but again with metadata and few other characters. It looks like this:

;; end of first chunk 

########################################
 Difference in instrument
########################################

Port: P035
;; other metadata 

Although the first attempt isn't success, I've learnt a bit about the structure of the data files. And based on that, I came up with a second approach: read the data files into R as a vector of string, one element for a line, and then remove all the lines which are metadata. They start with Port:, Site: or Longitude: etc or the ### chunk. It can be done using grep function, which tells me exactly which element of the vector contains the metadata.

s <- readLines(data.file)
metainfo.list <- c("Port:", "Site:", "Latitude:", "Longitude:", "Start Date:", "End Date:", "Contributor:", "Datum information:", "Parameter code:")
meta.line.num <- sapply(metainfo.list, function(i) {
    grep(pattern = i, s)
})
res.2 <- s[-meta.line.num]

This approach works well as long as the metainfo.list contains all the lines I'd like to remove. The downside is that I won't able to know I've includes all of them until the whole process is finished. So when I was waiting for the program to finish, I came up with a third approach, a better one.

The above two approaches are to discard the unnecessary information, but I may be in the situation that there are other lines that should be discard but I haven't encounter yet, then the process becomes tedious try-error and takes quite long.

Equally, another approach is to select exactly what I am interested in by using regular expression. But first, I have to identify pattern. Each data point was recorded at a certain point, and therefore must be associated with a timestamp, for example, the first data point is recorded at 1926-01-01 00:00:00. They also has an ID values with an closing parentage's, for example 1.

1) 1985/01/01 00:00:00      1.0300      -0.3845  

So the content of my interests are have a common pattern that can be summarised as: the lines that start with a number of spaces, and also have

observation ID

few integers, and an ending parentheses,

observation date

few integers with forward slashes that means year, month and day, and then a space,

observation time

few integers with colons, means hour, minutes and seconds.

The patterns in RegExp can be formulated as the roi.pattern variable and the whole process can be implemented as:

roi.pattern <- "[[:space:]]+[[:digit:]]+\\) [[:digit:]]{4}/[[:digit:]]{2}/[[:digit:]]{2}"
roi.line.num <- grep(pattern = roi.pattern, s)
res.3 <- s[roi.line.num]

To me, there isn't an absolute winner between the second and third approach, but I prefer to use regular expression because it has more fun with it; I am a statistician and like to spot patterns.

Also, it is an direct approach and more flexible. Note I can continue to add components to the regular expression to increase the confidence in selecting the right data matrix. For example, there are spaces and then few integers at the timestamp. But it will presumably increase the run-time.

You can download the exmaple data and run the scripts listed below in R to reproduce all the results.

#### * Path
data.file <- "~/Downloads/1985WIC.txt" ## to the downloaded data file

#### * Approach 1
read.table(data.file, skip = 11) ## error

#### * Approach 2
s <- readLines(data.file)
metainfo.list <- c("Port:", "Site:", "Latitude:", "Longitude:", "Start Date:", "End Date:", "Contributor:", "Datum information:", "Parameter code:")
meta.line.num <- sapply(metainfo.list, function(i) {
    grep(pattern = i, s)
})
res.2 <- s[-meta.line.num]

#### * Approach 3
roi.pattern <- "[[:space:]]+[[:digit:]]+\\) [[:digit:]]{4}/[[:digit:]]{2}/[[:digit:]]{2}"
roi.line.num <- grep(pattern = roi.pattern, s)
res.3 <- s[roi.line.num]

A subtle change can lead to random results. For example, the above pattern can also be \\.RData$, which means file names ended with .RData. The dot (.) sign here literally means ".". Adding two backslashes \\ changes the meaning of the pattern completely, but both gives the same results. It gave me so much frustration when extrapolating a pattern that works in one case to a similar case but get random results.

The RegExp pattern above are reasonably easy to understand, if one spent 10 minutes reading the manual, but the following is just crazy.

m <- regexec(pattern = "^(([^:]+)://)?([^:/]+)(:([0-9]+))?(/.*)", x)

There are 12 parentheses, 6 square brackets and many other symbols. Even same symbol have different meanings, and it's hard to find out exactly what they means because

I've never seen an R book that mentioned regular expression. This topic is certainly not a teaching content in university courses or training workshops.

Even google fails to find any meaningful resource except for the Text Processing in Wiki, which is the best I could find.

Although there are related questions in StackOverflow, most of the answers were set in a very specific situation. It's hard make it applicable to other situations or learn this topic from the discrete Q&As.

It has created a mental barrier that statistician shouldn't teach nor learn RegExp at all, or at least for me. But my limited experience suggests that it is such a powerful feature that I've missed a lot.

I believe there will be more chances to process text files, for example, parse the log files of this blog. RegExp can improve the efficiency to a great extent. So I am considering to invest the time to learn it properly.

Are you a R user? What's your experience with regular expression? Do you have good learning materials to recommend? If so, please share your experience on the less-talked area.

I use a broad definition of project: as long as a task that requires a series of sub-tasks to be done, then it is a project. Normally I categories any tasks that relates to a project into three groups:

Project Tasks

the major tasks that must to been done in order to deliver the project product.

Tasks

administrative or miscellaneous tasks that keep the project goes on, like sent out the invoice.

Notes

anything that is important to the project and therefore worthy keeping a record, like meeting notes or decision made that that impacts the project progress.

Each category has a corresponding top level section or heading. Once this outline is setup, it is very convenient to view content under these categories, regardless of what tasks I was working on, either reading emails, coding, or writing report. Org-mode can scan all the .org files in a direcotry, and creates a tree-structure, with the file name being the root, and headings being the nodes.

An intuitive way to locate a any node is to start from the beginning, the process is same as finding a section in a text book. It can be summarised as:

1. first, find the right book by its name,
2. then find the right part,
3. then narrow down to the right section,

and continue to the section I am interested in. An more pleasure way is to use fuzzy match supported by Helm package - I can narrow down the selection by random nodes. For example, as the images below shows, to locate headline under this article among 40 org files, I only need to search "Small pro", because there are only three headlines has "Small" in its name, "small changes", "small talk", and "small project", and "pro" narrow down to the unique headline.

It saves me a lot of time in remembering where I saved one notes, and wandering around the files to find something. I only explain a bit of the features of Helm, if you want to try out, you can find my configuration here. I recommend a good tutorial if you want to know more.

We usually a couple of projects at the same time. Also, create a new tasks or notes is easy. org-capture-mode would create a temporary node and by default it will be saved as a subtree in refile.org, or I can directly re-locate the headline directly to this project using the locating mechanism above.

These two features are most enjoyable to use, and make me away from wandering in multiple directories, trying to find the right files, and therefore increase my productivity. Never under estimate how long you will spent in finding in one file.

Projects usually come with hard deadlines about the product delivery. Setting and change deadlines in org-mode is pleasurable with org-deadline C-c C-d.

It brings up a mini-calendar buffer (shown below), I can use shift+left and shfit+right to move forward and backward for a day, or shift-up and shift-down to move between weeks, and hit RET to select a deadline. Apart from navigating, I can also choose to type the exact date directly, like "2015-07-25" and hit RET.

Once the deadline is set it will show up in that day's calendar. I don't want to suddenly realise there is a deadline I had on that day. So it makes sense to have an early warning period to show the tasks if it is due in days. This behaviour is governed by the org-deadline-warning-days variable. In my Emacs configuration, I set to 30 days. It gives me plenty of time to do any tasks.

I also set deadlines for sub-tasks since it is quite easy to do in org-mode. But coming up with realistic deadlines is difficult. To me, it must give enough time to do the task properly, to the PM, it must be fit in the whole project plan and resource. Both are likely to have different opinion on how long to implement the new features with documentation. It is quite important skills to have: to me, it reflects my understand on the problem and also my own technical capability, to the manager, it is part of their project plan.

My initial estimation may be far from the actual effort, especially when the problem domain is new to me, or I haven't done similar tasks before. The more I do, the better I am good at estimating. At this stage, I practise this skill seriously, and like to have someone with more experienced to review my estimation.

To make this task easy for them, I'd present an overall view of the project time-lines, which clearly shows the period allocate to the specific tasks. org-timeline will generate a time-sorted view for all the tasks. The recent feedback I received is that I tend to overlook the time spent on documentation and tests. Someone with more than 10 years in software development says they usually takes about 3x times on these two tasks together than actually coding.

time-line view also provides benchmark to the progress and I check it frequently to make sure I am on track. It gives the PM a reference for swapping tasks if some becomes urgent.

Additional to have the early warning system to prevent sudden surprise, org-mode provides another way of monitoring the project in terms of resource - the actual time I spent on the project. This feature is quite useful when I am given a quite loose deadline but with limited resource, say 150 hours.

Since the sub-tasks are mostly defined in the early stage, whenever I start to do it, I clock in first by org-clock-in. The clocking will be stopped once I manually clock out, or clock in to another task, or the tasks is completed (marked as DONE.) For each clock entry, it shows start time, end time and duration.

Multiple clocking logs are accumulated, and each entry shows the start time, end time, and duration. The durations can be added up and tells me exactly how much time I spent on each tasks. The whole tasks under the project and aggregated across the whole project, by one single function org-clock-report (C-c C- C-r).

It is normal to underestimate the complexity of an tasks, and spent too much time in resolve them, and usually I can catch up the in the later stage, however if I had the feeling the overall progress has been affected, I need require more sources from the PM, and the quote I will give is extra hours I had based on my initial estimation. That's an quick reaction.

Also, the clock-report table tells me the different between my effort estimation and the actual time I spent on that tasks.