Reading event records

library(tfevents)

tfevents stores events in files inside a logdir, there might be many files tfevents records files in the log directory and each file can contain many events.

tfevents provides functionality to read records writen to the log directory and to extract their value in a convenient representation, useful if you want to analyse the results from R.

Let’s write a few records a temporary directory, then we’ll use tfevents functionality to read the data into R.

temp <- tempfile()
set_default_logdir(temp)

log_event(text = "hello world")
for (i in 1:10) {
  log_event(hello = i* runif(1))
}

We can collect all events from the logdir with:

events <- collect_events(temp)
events
#> # A tibble: 12 × 4
#>         event run    step    summary
#>    <tfvnts_v> <chr> <int> <tfvnts_s>
#>  1     <./ 0> .         0           
#>  2     <./ 0> .         0        [1]
#>  3     <./ 1> .         1        [1]
#>  4     <./ 2> .         2        [1]
#>  5     <./ 3> .         3        [1]
#>  6     <./ 4> .         4        [1]
#>  7     <./ 5> .         5        [1]
#>  8     <./ 6> .         6        [1]
#>  9     <./ 7> .         7        [1]
#> 10     <./ 8> .         8        [1]
#> 11     <./ 9> .         9        [1]
#> 12     <./10> .        10        [1]

collect_events() returns a tibble with each row repesenting an event writen to that logdir. Note that the tibble has 12 rows, 1 more than the number of our calls to log_event. That’s because every event record file includes an event storing some metadata, like the time it was created and etc. The run column indicates the directory within the logdir that the events were collected from. Although tfevents only supports writing summary events, tfevents record files can contain other kind of events like log message, TensorFlow graphs and metadata information (as we see in this file), in those cases the summary column will have a NA value. In summary events it will contain additional information on the summary.

You might want to collect only the summary events, like those that were created with log_event, in this case you can pass the type argument to collect_events specifying the kind of events that you want to collect.

summaries <- collect_events(temp, type = "summary")
summaries
#> # A tibble: 11 × 6
#>         event run    step    summary tag   plugin 
#>    <tfvnts_v> <chr> <int> <tfvnts__> <chr> <chr>  
#>  1     <./ 0> .         0     <text> text  text   
#>  2     <./ 1> .         1    <hello> hello scalars
#>  3     <./ 2> .         2    <hello> hello scalars
#>  4     <./ 3> .         3    <hello> hello scalars
#>  5     <./ 4> .         4    <hello> hello scalars
#>  6     <./ 5> .         5    <hello> hello scalars
#>  7     <./ 6> .         6    <hello> hello scalars
#>  8     <./ 7> .         7    <hello> hello scalars
#>  9     <./ 8> .         8    <hello> hello scalars
#> 10     <./ 9> .         9    <hello> hello scalars
#> 11     <./10> .        10    <hello> hello scalars

Since the above asked for summary events, the returned data frame can include additional information like the name of the plugin that was used to created the summary (eg. scalars, images, audio, etc) and the tag name for the summary. This data is already included in the objects in the summary column, but is extracted as columns to make analyses easier.

You can extract the value out of a summary using the value function.

value(summaries$summary[1])
#> [1] "hello world"

Notice that value extracts values of a single summary and errors if you pass more summary values. To query all values you can pass the as_list = TRUE argument. This ensures that value will always return a list, making it type stable no matter the size of the summary_values vector that you pass.

# we remove the first summary, as it's a text summary
value(summaries$summary[-1], as_list = TRUE)
#> [[1]]
#> [1] 0.08075014
#> 
#> [[2]]
#> [1] 1.668666
#> 
#> [[3]]
#> [1] 1.802283
#> 
#> [[4]]
#> [1] 0.6288338
#> 
#> [[5]]
#> [1] 0.03699721
#> 
#> [[6]]
#> [1] 2.798361
#> 
#> [[7]]
#> [1] 3.484442
#> 
#> [[8]]
#> [1] 2.318138
#> 
#> [[9]]
#> [1] 6.595938
#> 
#> [[10]]
#> [1] 7.725215

If you are only interested in scalar summaries, you can use type="scalar" in collect_events:

scalars <- collect_events(temp, type = "scalar")
scalars
#> # A tibble: 10 × 7
#>         event run    step    summary tag   plugin   value
#>    <tfvnts_v> <chr> <int> <tfvnts__> <chr> <chr>    <dbl>
#>  1     <./ 1> .         1    <hello> hello scalars 0.0808
#>  2     <./ 2> .         2    <hello> hello scalars 1.67  
#>  3     <./ 3> .         3    <hello> hello scalars 1.80  
#>  4     <./ 4> .         4    <hello> hello scalars 0.629 
#>  5     <./ 5> .         5    <hello> hello scalars 0.0370
#>  6     <./ 6> .         6    <hello> hello scalars 2.80  
#>  7     <./ 7> .         7    <hello> hello scalars 3.48  
#>  8     <./ 8> .         8    <hello> hello scalars 2.32  
#>  9     <./ 9> .         9    <hello> hello scalars 6.60  
#> 10     <./10> .        10    <hello> hello scalars 7.73

Now, values can be expanded in the data frame and you get a ready to use data frame, for example:

library(ggplot2)
ggplot(scalars, aes(x = step, y = value)) +
  geom_line()

Iterating over a logdir

Passing a directory path to collect_events by default collects all events in that directory, but you might not want to collect them all at once because of memory constraints or even because they are not yet written, in this case you can use events_logdir to create an object, similar to a file connection that will allow you to load events in smaller batches.

con <- events_logdir(temp)
collect_events(con, n = 1, type = "scalar")
#> # A tibble: 1 × 7
#>        event run    step    summary tag   plugin   value
#>   <tfvnts_v> <chr> <int> <tfvnts__> <chr> <chr>    <dbl>
#> 1      <./1> .         1    <hello> hello scalars 0.0808

Notice that the first call to collect_events collected the first scalar summary event in the directory because of n = 1. The next call will collect the remaining ones as by default n = NULL.

collect_events(con, type = "scalar")
#> # A tibble: 9 × 7
#>        event run    step    summary tag   plugin   value
#>   <tfvnts_v> <chr> <int> <tfvnts__> <chr> <chr>    <dbl>
#> 1     <./ 2> .         2    <hello> hello scalars 1.67  
#> 2     <./ 3> .         3    <hello> hello scalars 1.80  
#> 3     <./ 4> .         4    <hello> hello scalars 0.629 
#> 4     <./ 5> .         5    <hello> hello scalars 0.0370
#> 5     <./ 6> .         6    <hello> hello scalars 2.80  
#> 6     <./ 7> .         7    <hello> hello scalars 3.48  
#> 7     <./ 8> .         8    <hello> hello scalars 2.32  
#> 8     <./ 9> .         9    <hello> hello scalars 6.60  
#> 9     <./10> .        10    <hello> hello scalars 7.73

We can now log some more scalars and recollect, you will see that the events that were just written are now collected.

for (i in 1:3) {
  log_event(hello = i* runif(1))
}
collect_events(con, type = "scalar")
#> # A tibble: 3 × 7
#>        event run    step    summary tag   plugin  value
#>   <tfvnts_v> <chr> <int> <tfvnts__> <chr> <chr>   <dbl>
#> 1     <./11> .        11    <hello> hello scalars 0.875
#> 2     <./12> .        12    <hello> hello scalars 0.350
#> 3     <./13> .        13    <hello> hello scalars 0.103

This interface allows you to efficiently read tfevents records without having them all on RAM at once, or work in a streaming way in a sense that a process might be writing tfevents (for example in a training loop) and another one is used to display intermediate results - for example in a Shiny app.