Metrics chart wrongly connects points on X axis

[zegerhoogeboom]

System information

• Have I written custom code (as opposed to using a stock example script provided in MLflow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• MLflow installed from (source or binary): source
• MLflow version (run python -c "from mlflow import version; print(version.VERSION)"): 0.1.0
• Python version: 3.6.5
• npm version (if running the dev UI): 6.0.1
• Exact command to reproduce: NA

Describe the problem

In the UI at the metrics graph view, it looks like this for me:

The data points are logged at the end of an epoch using this Keras callback:

class LossHistory(Callback):
    def on_epoch_end(self, epoch, logs=None):
        log_metric("loss", logs.get('loss'))
        # in reality more metrics here

logs.get('loss') just returns a number. The connections of the dots in the interface should be connecting X=0 to X=1 to X=2 and so on. Why they are currently connected in this strange way, I have no clue.

[mateiz]

Is there any chance you could share your MLflow run folder that caused this (mlruns/0/<runID>) or the file for the metric inside there? Also, which browser are you using?

[zegerhoogeboom]

Terribly sorry for the late reply. cat mlruns/0/2924aa6b449944bc849d4381e0fa3570/metrics/loss:

0 4.54306590325
0 3.50579357207
0 3.63842143944
0 3.63074730358
0 3.62895112526
0 3.60019535707
0 3.34938469377
0 3.69437173072
0 3.4442458695
0 5.05011896209

cat mlruns/0/2924aa6b449944bc849d4381e0fa3570/meta.yml:

artifact_uri: /app/mlruns/0/2924aa6b449944bc849d4381e0fa3570/artifacts
end_time: null
entry_point_name: !!python/unicode ''
experiment_id: !!python/long '0'
name: Run 5
run_uuid: 2924aa6b449944bc849d4381e0fa3570
source_name: !!python/unicode '/home/{my source code location}.py'
source_type: 4
source_version: !!python/unicode '94de7c2a7562525db2ead17fe8d020a7ca1cf32a'
start_time: !!python/long '1529310274917'
status: 1
tags: []
user_id: !!python/unicode '{my user name}'

It's being run from a Docker container using continuumio/miniconda, the host is running CentOS 7. I'm using Chrome (version 67) as my browser.

Note that the experiment in my screenshot and this one are not the same but the same happens for any experiment.

[mateiz]

Interesting, I don't see this on Chrome 67 with the same file. Do you have any extensions that might be blocking JavaScript? It's also a bit weird that your metrics are being logged with 0 as the timestamp (it should be the time since the epoch), but I tried it with either 0 or a recent timestamp and it seems to work OK.