Exporting Data Point Reports
You can download a report of the expected and predicted results for every data point that was used in a specific test. The report is in CSV format.
This report is useful in many use cases. For example:
There are times when your training model does not pass a test. Perhaps the model's overall performance is better but simply knowing that 20% of cases have failed is not enough. The model might be performing worse on a specific critical group of data points - for instance on the pedestrians class in an autonomous cars use case.
You wish to compare tests for two models in a spreadsheet and then see samples where model A (let's say the model in production) was correct but model B (say, the newest best-trained model) was incorrect.
For this feature to be available with no additional integration, all of the following conditions must be fulfilled:
- Your code has been developed under Keras.
- Your data is stored in MissingLink data volumes.
- Your data has been streamed to your training machine using MissingLink iterators.
To download a data point report:
Access the experiment and the test for which you wish to download data. Tests for an experiment are listed in the Test tab.
Click Export Data Point Report.
Check your email after a few minutes. Look for a message with the subject: "Test report ready for you from MissingLink.ai."
Display the email message.
The message looks similar to this:
Click Download Report in the email message.
Display the file that has downloaded.
When displayed in a spreadsheet application, the content looks similar to this:
Data is arranged in the following columns:
Column name Description id ID of the file path Path of the file in the data volume url URL of file in user's storage (optional) version MissingLink version of the data in the data volume that were used for the query datetime Datestamp expected Expected outcome predicted Predicted outcome probability_0 Probability of the outcome being 0 probability_1 Probability of the outcome being 1 probability_i Probability of the outcome being i class_name_0 Class name of 0 class_name_1 Class name of 1 class_name_i Class name of i