DL Python Network Learner

This node allows custom training and fine-tuning/transfer learning of a (Python compatible) deep learning network in a local Python installation via a user-defined script. The path to the Python executable has to be configured in Preferences → KNIME → Python. It also allows to import Jupyter notebooks as Python modules via the knime_jupyter module that is part of the Python workspace.

Options

Script

Editor
The script in the editor panel in the center will be executed to generate this node's output. "Execute script" will execute the full script while "Execute selected lines" will only execute the currently selected part of the script.

The editor optionally provides autocompletion (CTRL + Space) if the Python module jedi is installed.
Flow variables
The pane on the left side of the editor panel lists the available flow variables. A double click on a flow variable will insert the corresponding identifier at the current cursor position in the editor panel.
Workspace
The pane on the right side of the editor panel is the workspace and lists the currently defined identifiers in the running Python instance. "Reset workspace" will remove the output of all past script executions.
Console
The console under the editor panel displays standard output, standard error, or other information.
Row limit (dialog)
Limits the number of rows of the input table loaded in the dialog.
Missing values (Int, Long)
By default, Int and Long columns containing missing values are converted to doubles in Python, because doubles are the only numeric type containing a built-in missing value representation (NaN). Converting longs to doubles, however, introduces imprecision. The conversion may be avoided by replacing missing values with so-called sentinel values. To convert missing values to sentinel values before script execution, check the first option, to convert sentinel values back to missing values after script execution, check the second option. To select a sentinel value, click on MIN_VAL (the smallest possible value for the column's data type), MAX_VAL (the largest possible value for the column's data type), or a manually enter an integer in the text field.
Rows per chunk
Large tables are broken into chunks when they are transferred between Python and Java. This option controls how many rows are included in each chunk. The default value is intended for tables containing few columns having numeric types. If larger data types like images or text are transferred, a single chunk may exceed the maximum allowed buffer size. In this case lowering the value of this parameter helps getting reasonably sized chunks.

Input Ports

Icon
The deep learning network. Referred to as input_network in the Python script. Can be any deep learning network which can be used in Python.
Icon
The input table. Referred to as input_table in the Python script, where it is a pandas.DataFrame.

Output Ports

Icon
The trained output deep learning network. This is the content of the variable output_network which has to be defined in the script and has to be a deep learning network known to KNIME. The network can then be used in common "DL Python" nodes or in nodes which are specific to the network type.

Popular Predecessors

Views

Standard output
The text sent by Python to the standard output during the execution of this node.
Error output
The text sent by Python to the error output during the execution of this node.

Workflows

Links

Developers

You want to see the source code for this node? Click the following button and we’ll use our super-powers to find it for you.