Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Plots #1186

Merged
merged 9 commits into from
May 7, 2020
Merged

Plots #1186

Show file tree
Hide file tree
Changes from 7 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions config/prismjs/dvc-commands.js
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,9 @@ module.exports = [
'metrics',
'params diff',
'params',
'plot show',
'plot diff',
'plot',
'lock',
'list',
'install',
Expand Down
130 changes: 130 additions & 0 deletions content/docs/command-reference/plot/diff.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,130 @@
# plot diff

Show multiple versions of
[continuous metrics](/doc/command-reference/plot#continous-metrics) by plotting
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wonder whether we want to name them continuous. This word applies to functions. What about, for example, confusion matrix? Data for that type of plot is not continuous.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed with @pared . Probably even plain explicit "non-scalar metric" would be better?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yeah. it will be changed. all the terminology around continuous will be removed in the next iteration.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • So let's leave "continuous" in this PR so we have a plot cmd ref for now, and update it in a following PR.

them in a single image.

## Synopsis

```usage
usage: dvc plot diff [-h] [-q | -v] [-t [TEMPLATE]] [-d [DATAFILE]] [-f FILE]
[-s SELECT] [-x X] [-y Y] [--stdout] [--no-csv-header]
[--no-html] [--title TITLE] [--xlab XLAB] [--ylab YLAB]

positional arguments:
revisions Git revisions to plot from
```

## Description

This command visualize difference between continuous metrics among experiments
in the repository history. Requires that Git is being used to version the
metrics files.

The metrics file needs to be specified through `-d`/`--datafile` option. Also, a
plot can be customized by [Vega](https://vega.github.io/) templates through
option `--template`. To learn more about the file formats and templates please
see `dvc plot`.

Run without any revision specified, this command compares metrics currently
presented in the workspace (uncommitted changes) with the latest committed
version. A single specified revision shows the difference between the revision
and the version in the workspace.

In contrast to many commands such as `git diff`, `dvc metrics diff` and
`dvc prams diff` the plot difference shows all the revisions in a single ouput
and does not limited by two versions. A user can specify as many revisions as
needed.
dmpetrov marked this conversation as resolved.
Show resolved Hide resolved

The files with metrics can be files commited in Git as well as data files under
DVC control. In the case of data files, the file revision is corresponded to Git
revision of [DVC-files](/doc/user-guide/dvc-file-format) that has this file as
an output.

## Options

- `-d [DATAFILE], --datafile [DATAFILE]` - Continuous metrics file to visualize.

- `-t [TEMPLATE], --template [TEMPLATE]` - File to be injected with data. The
default temlpate is `.dvc/plot/default.json`. See more details in `dvc plot`.

- `-f FILE, --file FILE` - Name of the generated file. By default, the output
file name is equal to the input filename with additional `.html` suffix or
`.json` suffix for `--no-html` mode.

- `--no-html` - Do not wrap output vega plot json with HTML.

- `-s SELECT, --select SELECT` - Select which fileds or jsonpath to put into
plot. All the fields will be included by default with DVC generated `index`
field - see `dvc plot`.

- `-x X` - Field name for x axis. `index` is the default field for X.

- `-y Y` - Field name for y axis. The dafult field is the last field found in
the input file: the last column in CSV file or the last field in the JSON
array object (the first object).

- `--xlab XLAB` - X axis title. The X column name is the default title.

- `--ylab YLAB` - Y axis title. The Y column name is the default title.

- `--title TITLE` - Plot title.

- `-o, --stdout` - Print plot content to stdout.

- `--no-csv-header` - Provided CSV or TSV datafile does not have a header.

- `-h`, `--help` - prints the usage/help message, and exit.

- `-q`, `--quiet` - do not write anything to standard output. Exit with 0 if no
problems arise, otherwise 1.

- `-v`, `--verbose` - displays detailed tracing information.

## Examples

The difference between a not commited version of the file and the last commited
one:
dmpetrov marked this conversation as resolved.
Show resolved Hide resolved

```dvc
$ dvc plot diff -d logs.csv
file:///Users/dmitry/src/plot/logs.csv.html
```

A new file `logs.csv.html` was generated. User can open it in a web browser.

![](/img/plot_diff_workspace.svg)

The difference between two specified commits (multiple commits, tag or branches
can be specified):

```dvc
$ dvc plot diff -d logs.csv HEAD 11c0bf1
file:///Users/dmitry/src/plot/logs.csv.html
```

![](/img/plot_diff.svg)

The predefined confusion matrix template shows how continuous metrics difference
can be faceted by separate plots:

```csv
actual,predicted
cat,cat
cat,cat
cat,cat
cat,dog
cat,dinosaur
cat,dinosaur
cat,bird
turtle,dog
turtle,cat
...
```

```dvc
$ dvc plot diff -d classes.csv -t confusion
file:///Users/dmitry/src/test/plot_old/classes.csv.html
```

![](/img/plot_diff_confusion.svg)
239 changes: 239 additions & 0 deletions content/docs/command-reference/plot/index.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,239 @@
# plot

Contains commands to visualize
[continuous metrics](/doc/command-reference/plot#continuous-metrics) in
structured files (JSON, CSV, or TSV): [show](/doc/command-reference/plot/show),
[diff](/doc/command-reference/plot/diff).

## Synopsis

```usage
usage: dvc plot [-h] [-q | -v] {show,diff} ...

positional arguments:
COMMAND
show Generate a plot image file from a continuous metrics file.
diff Plot continuous metrics differences between commits in the DVC
repository, or between the last commit and the workspace.
```

## Description

DVC provides a set of commands to visualize _continuous metrics_ of machine
learning experiments. Usual examples of plots are AUC curves, loss functions,
and confusion matrices.

Continuous metrics represent plots, and should be stored as data series in one
of the supported [file formats](#file-formats). These files are usually created
by users or generated by user modeling or data processing code.

The plot commands can work with these continuous metrics files that are commited
to a repository history, data files controlled by DVC or files from workspace.
For examlpe, the command `dvc plot diff` generates a plot with two versions of
the metrics:

```dvc
$ dvc plot diff -d logs.csv
file:///Users/dmitry/src/plot/logs.html
```

![](/img/plot_auc.svg)

### Difference between continuous and scolar metrics
jorgeorpinel marked this conversation as resolved.
Show resolved Hide resolved

DVC has two concepts for metrics for representing result of machine learning
training or data processing:

1. `dvc metrics` to represent scalar numbers such as AUC, true positive rate and
others.
2. `dvc plot` to visualize continuous metrics such as AUC curve, loss function,
confusion matrixes and others.
dmpetrov marked this conversation as resolved.
Show resolved Hide resolved

In contrast to continuous metrics, scalar metrics should be stored in a
hirarchical files such as JSON and YAML and `dvc metrics diff` command can
represent difference between the metrics in different experiments as a float
numbers. Like `AUC` metrics is `0.801807` and was increase by `+0.037826` from
the previous value:

```dvc
$ dvc metrics diff
Path Metric Value Change
summary.json AUC 0.801807 0.037826
```

### File formats

Supported file formats for continuous metrics are: JSON, CSV, TSV. DVC expects
to see an array (or multiple arrays) of objects (usually _float numbers_) in the
file.
Comment on lines +66 to +68
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Again here, maybe use the "data series" term and avoid complicated descriptions until you get into each format?

Suggested change
Supported file formats for continuous metrics are: JSON, CSV, TSV. DVC expects
to see an array (or multiple arrays) of objects (usually _float numbers_) in the
file.
Supported formats for continuous metrics are: JSON, CSV, and TSV. DVC expects
to find data series (usually containing _float numbers_) in the file.


In tabular file formats such as CSV and TSV the array is a column. Plot command
can generate visuals for a specified column or a set of columns. Like `AUC`
column:
Comment on lines +70 to +72
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
In tabular file formats such as CSV and TSV the array is a column. Plot command
can generate visuals for a specified column or a set of columns. Like `AUC`
column:
In tabular continuous metrics files (CSV and TSV formats), each column is a
series. `dvc plot show` can generate visuals for one, several, or all columns.
For example `AUC` below:


```
epoch, AUC, loss
34, 0.91935, 0.0317345
35, 0.91913, 0.0317829
36, 0.92256, 0.0304632
37, 0.92302, 0.0299015
```

In hierarchical file formats such as JSON an array of JSON-objects is expected.
Plot command can generate visuals for a specified field name or a set of fields
from the array's object. Like `val_loss` field in the `train` array in this
example:
Comment on lines +82 to +85
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
In hierarchical file formats such as JSON an array of JSON-objects is expected.
Plot command can generate visuals for a specified field name or a set of fields
from the array's object. Like `val_loss` field in the `train` array in this
example:
In JSON files (hierarchical format) a named array of matching JSON objects is
expected. `dvc plot show` can generate visuals for a one, several, or all field
names from the array's objects. For example `val_loss` in the `train` array below:


```
{
"train": [
{"val_accuracy": 0.9665, "val_loss": 0.10757},
{"val_accuracy": 0.9764, "val_loss": 0.07324},
{"val_accuracy": 0.8770, "val_loss": 0.08136},
{"val_accuracy": 0.8740, "val_loss": 0.09026},
{"val_accuracy": 0.8795, "val_loss": 0.07640},
{"val_accuracy": 0.8803, "val_loss": 0.07608},
{"val_accuracy": 0.8987, "val_loss": 0.08455}
]
}
```

### Plot templates

DVC generates plots as HTML files that a user can click and open in a web
browser. The HTML files contain plots as [Vega-Lite](https://vega.github.io/)
objects. The files can also be transformed to traditional PNG, JPEG, SVG image
formats using external tools.

Vega is a declarative, programming language agnostic format of defining plots as
JSON specification. DVC gives users the ability to change the specification and
generate plots in the format that fits the best to the users need. At the same
time, it does not make DVC dependent on user's visualization code or any
programming language or environment which allows DVC stay programming language
agnostic.

Plot templates are stored in `.dvc/plot/` directory as json files. A user can
define it's own templates or modify the existing ones. The default template is
`.dvc/plot/default.json`. User can change the temlpate by `--template` or `-t`
option of `dvc plot show` or `dvc plot diff` commands and specifying a file
name.

For temlpates in the templates directory the path and the json extension are not
required. User can specify only `--template scatter` instead of
`--template .dvc/plot/scatter.json`. Any custom template can be added to the
temlpate directory.

### Custom templates

User can define their own temlpate for specific plot types. Any temlpate file is
a JSON specification with predefined DVC anchors that help DVC to inject user's
data properly.

All input JSON files of `dvc plot show` and `dvc plot diff` commands are
combined together into a single array for the injection to a template file.

There are two important additional signals or fields that DVC adds:

- `rev` - specified revision, tag or branch of input file. This option helps to
destinguish between different revisions of the file in `dvc plot diff`
command.

- `index` - is a ordering number in the file. In many cases it corresponds to
mchine learning training epoch or step number.

DVC applies the same logic to all input CSV files but first transforms all CSV
data into JSON. DVC uses CSV files columns name from a header for JSON
conversion.

DVC temlpate anchors:

- `<DVC_METRIC_DATA>` - Plotting command input data from either CSV or JSON
files is converted to JSON array and injected instead of this anchor. Two
additional signal will be added `index` and `rev` - revision (See above).

- `<DVC_METRIC_TITLE>` - A plot title that can be defined by `--title` option.

- `<DVC_METRIC_Y>` - a field name for Y axis of the plot. It can be defined by
`-y` option of the commands. The dafult field is the last field found in the
input file: the last column in CSV file or the last field in the JSON array
object.

- `<DVC_METRIC_X>` - a field name for Y axes. It can be defined by `-x` option.
`index` is the default field for X.

- `<DVC_METRIC_Y_TITLE>` - a displayed field label for Y.

- `<DVC_METRIC_X_TITLE>` - a displayed field label for X.

## Options

- `-h`, `--help` - prints the usage/help message, and exit.

- `-q`, `--quiet` - do not write anything to standard output.

- `-v`, `--verbose` - displays detailed tracing information.

## Examples

Tabular file `logs.csv` visualization:

```
epoch,accuracy,loss,val_accuracy,val_loss
0,0.9418667,0.19958884770199656,0.9679,0.10217399864746257
1,0.9763333,0.07896138601688048,0.9768,0.07310650711813942
2,0.98375,0.05241111190887168,0.9788,0.06665669009438716
3,0.98801666,0.03681169906261687,0.9781,0.06697812260198989
4,0.99111664,0.027362171787042946,0.978,0.07385754839298315
5,0.9932333,0.02069501801203781,0.9771,0.08009233058886166
6,0.9945,0.017702101902437668,0.9803,0.07830339228538505
7,0.9954,0.01396906608727198,0.9802,0.07247738889862157
```

```dvc
$ dvc plot show logs.csv
file:///Users/dmitry/src/plot/logs.csv.html
```

![](/img/plot_show.svg)

Difference between the current file and the previous commited one:

```dvc
$ dvc plot diff -d logs.csv HEAD^
file:///Users/dmitry/src/plot/logs.csv.html
```

![](/img/plot_diff.svg)

Visualize a specific field:

```dvc
$ dvc plot show -y loss logs.csv
file:///Users/dmitry/src/plot/logs.html
```

![](/img/plot_show_field.svg)

Confusion matrix template is predefined in DVC (file
`.dvc/plot/confusion_matrix.json`):

```csv
actual,predicted
cat,cat
cat,cat
cat,cat
cat,dog
cat,dinosaur
cat,dinosaur
cat,bird
turtle,dog
turtle,cat
...
```

```dvc
$ dvc plot show classes.csv --template confusion -x actual -y predicted
file:///Users/dmitry/src/plot/classes.csv.html
```

![](/img/plot_show_confusion.svg)
Loading