[init] docs

Author: ctr26

docs/cli.md

@@ -0,0 +1,100 @@
+The cli is mostly handled by hydra (https://hydra.cc/docs/intro/). The main commands are:
+
+bie_train: Train a model
+bie_predict: Predict with a model
+
+# Training
+
+To train a model, you can use the following command:
+
+```bash
+bie_train
+```
+
+To see all the available options, you can use the `--help` flag:
+
+```bash
+bie_train --help
+```
+
+## Data
+
+Out of the box bie_train is configured to try to use torchvision.datasets.ImageFolder to load data.
+This can be endlessly overwritte using Hydra's configuration system (e.g. _target_ ).
+However, for most applications using the stock ImageFolder class will work.
+To then point the model to useful data you need to set the 'receipe.data' key like so:
+
+```bash
+bie_train recipe.data=/path/to/data
+```
+
+ImageFolder will use PIL to load images, so you can use any image format that PIL supports, this includes jpg, png, bmp, etc, tif.
+
+More exotic formats will require a custom dataset class, which is not covered here; realisitically you should convert your data to a more common format.
+PNG for instance is a lossless format that loads quickly from disk due to it's efficient compression.
+The bie_train defaults tend to be sane, for instance the data is shuffled, and the data is split into train and validation sets.
+
+It is worth noting that ImageFolder expects the data to be organised into "classes" even though default bie_train does not use the class labels during training.
+To denote these classes, you should organise your data into folders, where each folder is a class, and the images in that folder are instances of that class.
+See here for more information: https://pytorch.org/vision/stable/datasets.html#imagefolder
+
+## Models
+
+The default model backbone a "resnet18" with a "vae" architecture for autoencoding, but you can specify a different model using the `receipe.model` flag:
+
+```bash
+bie_train recipe.model=resnet50_vqvae receipe.data=/path/to/data
+```
+
+N.B. the resnet series of models expect the tensor input to (3,224,224) in shape,
+
+
+### Supervised vs Unsupervised models
+
+By default the model is unsupervised, meaning the class labels are ignored during training.
+However, a (experimental) supervised model can be selected by setting:
+
+```bash
+bie_train lit_model.model=_target_="bioimage_embed.lightning.torch.AutoEncoderSupervised" receipe.data=/path/to/data
+```
+
+This uses contrastive learning using the labelled data, specifically SimCLR: https://arxiv.org/abs/2002.05709
+
+## Reciepes
+
+The major components of the training process are controlled by the "reciepe" schema.
+These values are also what is used for generating the uuid of the training run.
+This means that the model can infact resume from a crash or be retrained with the same configuration aswell as multiple models being trained in parallel using the same directory.
+This is useful for hyperparameter search, or for training multiple models on the same data.
+
+### lr_scheduler and optimizer
+
+The lr_scheduler and optimizer are mimics of the timm library and built using create_optimizer and create_scheduler.
+https://timm.fast.ai/Optimizers
+and
+https://timm.fast.ai/schedulerss
+
+The default optimizer is "adamw" and the default scheduler is "cosine", aswell as some other hyperparameters borrowed from: https://arxiv.org/abs/2110.00476
+
+The way the timm create_* functions work is they receive a generic SimpleNamespace, and only take the keys they need.
+The consequence is that timm creates a controlled vocabulary for the hyperparameters in receipe; this makes it possible to choose from the wide variety of optimizers and schedulers in timm.
+https://timm.fast.ai
+
+## Augmentation
+
+The package includes a default augmentation, which is stored in the configruation file.
+The default augmentation is written using albumentations, which is a powerful library for image augmentation.
+https://albumentations.ai/docs/
+
+
+The default augmentation is a simple set of augmentations that are useful for biological_images, crucially it mostly neglects any RGB and non-physical augmentation effects.
+It is recommended to edit the default augmentations in the configuration file and not in the CLI as the commands can get quite long.
+
+
+## Config file
+
+This will train a model using the default configuration. You can also specify a configuration file using the `--config` flag:
+
+```bash
+bie_train --config path/to/config.yaml
+```