# luxonis_model_summary

Python API: `luxonis_train.callbacks.luxonis_model_summary`

Prints the layer summary of the model, as a rich table or as plain text.

## Classes

### LuxonisModelSummary

Callback that prints the layer summary of the model.

The callback extends the Lightning `RichModelSummary` and prints the summary when a fit starts. With `rich` on, it prints `rich`
tables to the console and writes a copy without terminal styling to the log file. With `rich` off, it logs a plain `tabulate`
table and the totals. These records reach the console and the log file.
[LuxonisLightningModule.configure_callbacks](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/lightning/luxonis_lightning.md)
adds this callback with `max_depth=2`, and with `rich` equal to `rich_logging` of the config.

#### Methods

##### init

```python
def __init__(rich: bool = True, **kwargs):
```

Initialize the callback.

Parameters

 * `rich` (`bool`): Print `rich` tables. `False` logs a plain text table instead.
 * `**kwargs`: Keyword arguments for the Lightning `RichModelSummary`. `max_depth` sets the deepest level of nested modules in the
   table, and `0` turns the summary off. Lightning passes every other keyword to
   [LuxonisModelSummary.summarize](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/callbacks/luxonis_model_summary.md).

##### summarize

```python
def summarize(*args, **kwargs):
```

Print the layer summary as `rich` tables or as plain text.

The Lightning `ModelSummary.on_fit_start` hook calls this method at the start of a fit, on global rank 0 only. It skips the call
when `max_depth` is `0`.

With `rich` on, the method prints a `rich` table and a grid of totals to the global `rich` console. It also writes both without
terminal styling to the log file only. With `rich` off, the method logs a `tabulate` table in the `fancy_grid` format and one line
per total with `logger.info`. These lines reach the console and the log file.

Each column of `summary_data` becomes a table column, in order. The method sets the headers by position: an index column, `Name`,
`Type`, `Params`, and `Mode`. When `summary_data` has the columns `In sizes` and `Out sizes`, the `rich` table adds these two
headers after `Mode`. Lightning puts a `FLOPs` column after `Mode`. Thus these two headers go over the `FLOPs` and `In sizes`
data. In the `rich` table, a data column after the last header gets no header. `tabulate` puts the five headers over the last five
columns. With more than five columns, no header of the plain table is over its data.

The totals are:

 * the trainable, the non-trainable, and the total parameter counts;
 * the estimated size of the parameters in MB, without the decimal part;
 * the number of modules in train mode and in eval mode.

The counts and the size use a short form, for example `1.2 M`.

Parameters

 * `*args`: The positional arguments of the Lightning hook, in order: `summary_data`, the columns as `(header, values)` pairs;
   `total_parameters`; `trainable_parameters`; `model_size`, in MB; and `total_training_modes`, a dictionary with the keys
   `"train"` and `"eval"`.
 * `**kwargs`: The keyword arguments of the Lightning hook. The `rich` table reads `header_style`, which is `"bold magenta"` when
   it is not given. The method ignores all other keywords, such as `total_flops`.