diff --git a/.doctrees/environment.pickle b/.doctrees/environment.pickle index fdcfad4e..afb64f4f 100644 Binary files a/.doctrees/environment.pickle and b/.doctrees/environment.pickle differ diff --git a/.doctrees/tutorial.doctree b/.doctrees/tutorial.doctree index 8e24b43b..7bba85c2 100644 Binary files a/.doctrees/tutorial.doctree and b/.doctrees/tutorial.doctree differ diff --git a/.doctrees/tutorial_multi.doctree b/.doctrees/tutorial_multi.doctree index d60ed41b..55f31ffe 100644 Binary files a/.doctrees/tutorial_multi.doctree and b/.doctrees/tutorial_multi.doctree differ diff --git a/_sources/tutorial.rst.txt b/_sources/tutorial.rst.txt index 369d71ec..c1817b93 100644 --- a/_sources/tutorial.rst.txt +++ b/_sources/tutorial.rst.txt @@ -49,7 +49,7 @@ and multispectral data at the same time is not currently supported. Stick to one data type per model (or stack the RGB bands with the multispectral bands and treat as in the case of multispectral data). -Preparing data (RGB and multispectral) +Preparing data (RGB/multispectral) -------------------------------------- An example of the recommended file structure when training a new model is as follows: @@ -271,7 +271,7 @@ multispectral (``.tif``) tiles. display(Image.fromarray(image)) -Training a model (RGB) +Training (RGB) ---------------------- Before training can commence, it is necessary to register the training data. It is possible to set a validation fold for @@ -357,7 +357,7 @@ Training outputs, including model weights and training metrics, will be stored i Early stopping is implemented and will be triggered by a sustained failure to improve on the performance of predictions on the validation fold. This is measured as the AP50 score of the validation predictions. -Training a model (multispectral) +Training (multispectral) -------------------------------- The process for training a multispectral model is similar to that for RGB data but there are some key steps that are @@ -547,7 +547,7 @@ model is to be used on a range of different resolutions, random resizing can hel different scales. -Post-training (check training convergence) +Post-training (check convergence) ------------------------------------------ It is important to check that the model has converged and is not overfitting. This can be done by plotting the training diff --git a/_sources/tutorial_multi.rst.txt b/_sources/tutorial_multi.rst.txt index 125e06ef..91548842 100644 --- a/_sources/tutorial_multi.rst.txt +++ b/_sources/tutorial_multi.rst.txt @@ -17,8 +17,8 @@ The key steps are: 4. Making landscape level predictions -Preparing data (RGB and multispectral) --------------------------------------- +Preparing data +-------------- Data can be prepared in a similar way to the single class case but the classes and their order (mapping) need to be saved so that they can be accessed diff --git a/index.html b/index.html index 5b262996..e0e17e86 100644 --- a/index.html +++ b/index.html @@ -106,18 +106,18 @@
An example of the recommended file structure when training a new model is as follows:
├── Danum (site directory)
│ ├── rgb
@@ -325,8 +325,8 @@ 2.1. Preparing data (RGB and multispectr
Before training can commence, it is necessary to register the training data. It is possible to set a validation fold for model evaluation (which can be helpful for tuning models). The validation fold can be changed over different training steps to expose the model to the full range of available training data. Register as many different folders as necessary
@@ -398,8 +398,8 @@The process for training a multispectral model is similar to that for RGB data but there are some key steps that are
different. Data will be read from .tif
files of 4 or more bands instead of the 3-band .png
files.
Data should be registered as before:
@@ -549,8 +549,8 @@It is important to check that the model has converged and is not overfitting. This can be done by plotting the training
and validation loss over time. The detectron2
training routine will output a metrics.json
file that can be used
to plot the training and validation loss. The following code can be used to plot the loss:
Data can be prepared in a similar way to the single class case but the classes and their order (mapping) need to be saved so that they can be accessed consistently across training and prediction. The classes are saved in a json