Namespace

File Namespace
We have reserved some names for your files so special actions and emphasize can be done automatically on the platform when you upload them.
Here is the full list :
All the names are in lowercase
config
This is the config file used for training with our python training packages for Tensorflow 1 and Tensorflow 2.
checkpoint-data-latest
This is the data file of checkpoints when training with Tensorflow.
checkpoint-index-latest
This is the index file of checkpoints when training with Tensorflow.
model-latest
This is the exported trained model file, it must be called that way to be deployed using Picsell.ia
Data Namespace
We have reserved some names for what you log to Picsell.ia, it is used to emphasize the most important information in the 'Summary' section of your experiment or so we can automatically compare your experiments according to those values.
Here is the full list : 
All the names are in lowercase
accuracy
If you log your accuracy to Picsell.ia under this name, we will automatically display the last (or the only) value to the summary (see above).
loss
If you log your loss to Picsell.ia under this name, we will automatically display the last (or the only) value to the summary (see above).
train-split
It will frequently happen that you split your data between a training and a test/validation batch. If you want to visualize the repartition of the data among all your classes, you will likely log a bar chart looking like this :
With the following command:
data = {
    'x': ['car', 'person', 'bird'],
    'y': [10, 25, 12],
    'image_list': [...]
}
experiment.log(name='train-split', type='bar', data=data)
By naming your data train-split, you will have access to a brand new tab in your experiment that allows you to dive deep in the batches from your splits and check if there appears to be no issues with your data.
The image_list key allows you to match each of your image with a particular set so you can explore it later in the platform. The value is a list containing the internal_id of each picture.
test-split
See train-split​
eval-split
See train-split​
labelmap
If you train a neural network within your experiment, you will need a label map to teach the network with your labels. To save it, we encourage you to log it that way :
data = {
    '0': 'car',
    '1': 'person',
    '2': 'bus'
}
experiment.log(name='labelmap', data=data, type='labelmap')
Remember to set the type as 'labelmap' so we will not try to display it in your dashboard, but we will be able to use it in the playground.
confusion-matrix
If you log a heatmap with the name 'confusion-matrix', it will obviously be displayed in your logs tab, but will also be used in the eval tab for the interactive evaluation visualization.
See heatmap for more information on the format.
evaluation
After you train your model, you will surely perform an evaluation step. To make it easier for you to find edge-cases, compare predictions to ground-truth ... We created a dedicated tab in the experiment dashboard called eval.
To see how to correctly use the evaluation name, please refer to this page of the documentation.

Publish your model

Hyperparameter tuning

Last modified 1yr ago

Copy link

Outline

File Namespace

config

checkpoint-data-latest

checkpoint-index-latest