Writing a configuration

One file, one run

The shortest thing that trains:

task: semantic_segmentation
data:
  train_path: data/train
  validation_path: data/validation
  colormap: [[0, 0, 0], [255, 255, 255]]
models:
  - name: first_try
    input_shape: [256, 256]

spec <- platypus_spec("run.yaml")
fit <- platypus_fit(spec)
evaluate(fit)

Four things are required and the rest have defaults: task, data.train_path, and a name and input_shape for each model. Everything else - the architecture, the loss, the optimiser, how many epochs - has a default that lives in the engine, so a short configuration is a complete one rather than an incomplete one. ?u_net has a Where the defaults live section saying why they are not written into the R signatures.

colormap says which colour in a mask means which class, darkest first. Masks stored as label maps rather than pictures use labels: [0, 1] instead, and giving both is an error: exactly one says how a mask is encoded.

Several models at once

models: is a list, and that is the reason this file format exists rather than only the code path. Each entry trains on the same data and lands in the same table, so comparing variants is one run rather than a loop somebody has to write.

task: semantic_segmentation
data:
  train_path: data/train
  validation_path: data/validation
  colormap: [[0, 0, 0], [255, 255, 255]]
models:
  - name: shallow
    input_shape: [256, 256]
    blocks: 3
    filters: 16
  - name: deep
    input_shape: [256, 256]
    blocks: 5
    filters: 32
    architecture: u_net_plus_plus

fit <- platypus_fit(platypus_spec("run.yaml"))
evaluate(fit)            # one row per model
training_history(fit)    # one row per model per epoch

name has to be unique, because it is how a model is identified in every score table, in every record and in predict(fit, model = ...). Two models called the same thing is refused rather than resolved, by name: model names must be unique; repeated: shallow.

Two entries differing only in blocks is a legitimate configuration and a common one: this is for comparing variants, not only architectures.

Dividing one folder

With no validation_path, say how to divide train_path:

task: semantic_segmentation
data:
  train_path: data/all
  split:
    fractions: [0.7, 0.15, 0.15]
    group_by: "^(patient\\d+)_"
  colormap: [[0, 0, 0], [255, 255, 255]]
models:
  - name: by_patient
    input_shape: [256, 256]

Two fractions give train and validation, three add a test set. Nothing is written to disk - the samples are divided after they are discovered, so there are no new folders and no CSV. split_dataset() is still the tool when the CSVs themselves are the point.

group_by is required and may be NULL. It is a regular expression whose first group is the thing that must not appear on both sides - a patient, a scan, a slide. The mistake it prevents leaves no trace: slices of one patient in both sets make validation measure memory rather than generalisation, and every number in the output still looks reasonable. An omitted key would be that choice made silently; null is the same choice made on purpose, for data where every sample is independent.

Exactly one of split and validation_path. If a run genuinely has no validation set - a final fit on everything - say validation: false, which is the third answer rather than an absence, because “I have no validation set” and “I forgot” look identical otherwise.

The same from R, where the argument answers the same question three ways:

segmentation_data("data/all", split = list(fractions = c(0.8, 0.2), group_by = NULL))
segmentation_data("data/train", validation = "data/validation")
segmentation_data("data/all", validation = FALSE)

Augmentation, callbacks, and the rest of a model

task: semantic_segmentation
data:
  train_path: data/train
  validation_path: data/validation
  colormap: [[0, 0, 0], [255, 255, 255]]
models:
  - name: augmented
    input_shape: [256, 256]
    epochs: 40
    loss: {name: dice}
    metrics: [{name: iou}, {name: dice}]
    optimizer: {name: adam, learning_rate: 0.001}
    callbacks:
      - {name: early_stopping, monitor: val_loss, patience: 10}
      - {name: model_checkpoint, monitor: val_iou, path: runs/best.safetensors}
    augmentation:
      - {name: HorizontalFlip, params: {p: 0.5}}
      - {name: RandomBrightnessContrast, params: {p: 0.3}}

Every one of those blocks has an R constructor that builds the same thing, and the help pages for them are where the measurements are: ?losses, ?metrics, ?optimizers, ?callbacks.

u_net(
  "augmented",
  input_shape = c(256, 256),
  epochs = 40,
  loss = loss_dice(),
  metrics = list(metric_iou(), metric_dice()),
  optimizer = optimizer_adam(learning_rate = 0.001),
  callbacks = list(
    callback_early_stopping(monitor = "val_loss", patience = 10),
    callback_model_checkpoint(monitor = "val_iou", path = "runs/best.safetensors")
  ),
  augmentation = list(
    augmentation_step("HorizontalFlip", p = 0.5),
    augmentation_step("RandomBrightnessContrast", p = 0.3)
  )
)

There is no mode on a callback that watches a number. The direction is taken from the name: anything ending in loss is better lower, anything else better higher. So monitor = "val_iou" maximises and monitor = "val_loss" minimises, and there is nothing to set inconsistently.

Augmentation applies to the training data only - never to validation, where the point is to measure the same thing every epoch. The names are albumentations’ own, and available_transforms() lists what the installed version offers; at rank = 3 the list is shorter, because support for volumes is uneven. Each step is probed at the model’s own input size while the pipeline is built, so a transform that cannot do the job says so before the first epoch rather than during it.

Overlapping classes instead of one per voxel

colormap and labels both say the same thing about a voxel: it is in exactly one class. regions is the third way of naming classes and the only one that drops that assumption - classes named as unions of a label map’s values, which is how BraTS and the lesion-inside-an-organ family actually store their ground truth.

task: semantic_segmentation
data:
  train_path: data/train
  validation_path: data/validation
  regions:
    - {name: whole tumour, of: [1, 2, 4]}
    - {name: tumour core, of: [1, 4]}
    - {name: enhancing, of: [4]}
models:
  - name: brats
    input_shape: [240, 240]
    loss: {name: focal}

The same from R. A region is a list(name = , of = ), and of takes the label-map values that belong to it:

spec <- platypus_spec(
  data = segmentation_data(
    "data/train", "data/validation",
    regions = list(
      list(name = "whole tumour", of = c(1, 2, 4)),
      list(name = "tumour core", of = c(1, 4)),
      list(name = "enhancing", of = 4)
    )
  ),
  models = list(u_net("brats", input_shape = c(240, 240), loss = loss_focal()))
)

The last region shows that a single value is ordinary rather than an edge case: BraTS’s enhancing tumour is one label.

The task is still semantic_segmentation. What changes is what a channel means.

There is no background entry, so the class count is not the one you would write as labels. Those same masks as labels = c(0, 1, 2, 4) are four classes; as the three regions above they are three channels. And a voxel whose value is 4 is in all three of them at once, which is the point rather than a tolerated side effect.

Each channel is an independent binary question - a sigmoid per channel rather than one softmax across them - so the channels do not compete and do not sum to one. That is what makes the overlap expressible, and it is also why a prediction cannot be written back as a single picture with save_masks(): one voxel gets one colour, and a voxel here may belong to three classes.

So save_masks() refuses such a prediction rather than taking the strongest channel, and predict() refuses type = "class" for the same reason - there is no one class per voxel to return. Ask for type = "probability" and write it with save_class_masks(), which gives one binary file per region, named after the region. The channel order is the order the regions were written in, which is also the order inspect_masks() names them in and the order every per-class score is reported in.

Two regions are the minimum, because one region is ordinary binary segmentation and labels = c(0, 1) already says that - with a background channel, which regions has none of. Disjoint regions are allowed and are not a mistake: they group several values into one class and drop the background channel, which is a useful thing to ask for and is refused nowhere.

The failure this invites is quiet, so it is refused loudly. A region whose of matches nothing in the data - of = 3 against a label map holding 1, 2 and 4 - is a channel of zeros. It trains to predict nothing and scores a Dice of 0 for that region while every other number in the run looks ordinary, which is the kind of result that gets read as “that class is hard”. platypus_fit() refuses it before training, and inspect_masks() asks the same question over as much of the data as you want - naming the region rather than giving its position, because missing_classes: 1 would leave you counting entries in the file above:

inspect_masks(spec)
#>   split unmatched samples_checked total_samples present_classes      missing_classes
#> 1 train    0.0000             500          1251 whole tumour, ...    enhancing

spec <- platypus_spec(
  data = segmentation_data(
    "data/train", "data/validation",
    classes_from = c("_vessel\\.png$", "_organ\\.png$")
  ),
  models = list(u_net("crossing", input_shape = c(256, 256)))
)

Instances instead of one mask

task: instance_segmentation keeps the objects apart rather than uniting them. The data block gets shorter rather than longer: no colormap and no labels, because there are always exactly two classes - background and object - and what names the objects is the directory of per-object mask files itself.

task: instance_segmentation
data:
  train_path: data/train
  validation_path: data/validation
  mode: nested_dirs
models:
  - name: rim
    input_shape: [256, 256]
    separation: {name: rim, width: 2}
    loss: {name: cce_dice}
    decode: [{name: nearest_seed}, {name: connected_components}]
  - name: flows
    input_shape: [256, 256]
    separation:
      name: flows
      loss: {name: mse, where: foreground}

The same two models from R. instance_data() is what says which task this is - an instance data block declares nothing a segmentation one lacks, so the constructor is the word:

spec <- platypus_spec(
  data = instance_data("data/train", "data/validation"),
  models = list(
    u_net(
      "rim",
      input_shape = c(256, 256),
      separation = separation_rim(width = 2),
      loss = loss_cce_dice(),
      decode = list(decode_nearest_seed(), decode_connected_components())
    ),
    u_net(
      "flows",
      input_shape = c(256, 256),
      separation = separation_flows(loss = loss_mse(where = "foreground"))
    )
  )
)
fit <- platypus_fit(spec)
evaluate(fit)

Two axes, and they are deliberately different shapes. separation is one block per model and has no default: it is what the network is trained to predict, so it sets the width of the head - three channels for rim, four for flows at two classes - and that width is derived rather than written anywhere. decode is a list, because several decodings share one forward pass: a second decoding of 134 images costs 1.41 s against an hour and a half for a second training run. evaluate() then returns a row per model and decoding.

Which mask column to read is a decision the table makes for you and it used to make it wrongly. Every row carries foreground_dice - the binary overlap of anything the model called an object against anything that is one, the same arithmetic whatever the separation is - and that is the one to read across models. The dice beside it is the training metric, a mean over the class channels the head has, so three for separation_rim() against two for separation_flows(): read down one model’s rows, never across two. On the Data Science Bowl’s validation cases the class column showed two masks 0.09 apart that are 0.0062 apart, and in the other direction.

No value appears in both lists, which is what settles which field a new idea belongs in: separation is a noun and decode is a verb. separation_rim() can be decoded by decode_nearest_seed() or decode_watershed(), separation_flows() by decode_follow_flows(), and decode_connected_components() reads no signal at all and works with every target. Leave decode out and it derives to the signal’s own decoding plus connected_components; give a list and you get the list you gave.

A model’s own loss and metrics score the mask at either separation, so they are an overlap or a cross-entropy loss here as everywhere else. separation_flows() carries its own loss for the two vector channels, which is where loss_mse(), loss_l1() and loss_smooth_l1() belong: where = "foreground" averages the error over the voxels the target is defined on, and on these images that is 13.97% of them.

What this task is worth is measured, and it is not flattering. On the Data Science Bowl the separation signal is worth +0.068 of instance average precision at best while a perfect mask is worth +0.2398, so the mask is the limit by a factor of 3.6 - and a perfect semantic mask decoded by connected_components still scores only 0.7655. That is what the yardstick row is for: a run that improved the mask and did nothing for the separation improves every row at once, and without the row that ignores the signal there is nothing to notice it against.

One thing is missing from R rather than from the engine, and it is missing on purpose rather than by oversight: predict() on an instance fit refuses. The run and its scores work - evaluate() gives the row per model and decoding above - but there is no call yet that brings the objects themselves back into R as a label map, so predict() says so instead of returning a mask that is not what you asked for. platypus#103.

Detection instead of segmentation

task: object_detection changes which fields the file accepts. Boxes rather than masks, so classes and an annotation format rather than a colormap:

task: object_detection
data:
  train_path: data/train
  validation_path: data/validation
  annotation_format: pascal_voc
  classes: [rbc, wbc, platelets]
models:
  - name: cells
    input_shape: [416, 416]

Leave anchors out and they are fitted to your own boxes by k-means on an IoU distance, anchors_per_grid of them per grid. That is worth about 0.23 of mean overlap against COCO’s nine on blood cells - measured, 0.65 against 0.88 - so the default is the useful one and naming anchors by hand is for reproducing a published model. detection_anchors() reports them and plot_anchors() draws them against the boxes they were fitted to.

One label instead of a mask

task: classification answers with one label for a whole image rather than one per voxel. The data block loses everything that describes a mask and gains the one thing a label needs - classes, in class order:

task: classification
data:
  train_path: data/train
  validation_path: data/validation
  mode: nested_dirs
  classes: [benign, malignant]
models:
  - name: ours
    input_shape: [128, 128]
    blocks: 4
    filters: 16
    loss: {name: cce}
    metrics: [{name: auc}, {name: balanced_accuracy}]
  - name: resnet
    input_shape: [224, 224]
    encoder: resnet34
    pretrained: true
    freeze_encoder: 2
    pooling: max
    head_dropout: 0.2
    loss: {name: focal, gamma: 2}

The same thing from code is classification_data() and classifier():

spec <- platypus_spec(
  data = classification_data("data/train", "data/validation",
                             classes = c("benign", "malignant")),
  models = list(
    classifier("ours", input_shape = c(128, 128), blocks = 4, filters = 16,
               loss = loss_cce(),
               metrics = list(metric_auc(), metric_balanced_accuracy())),
    classifier("resnet", input_shape = c(224, 224), encoder = "resnet34",
               pretrained = TRUE, freeze_encoder = 2, pooling = "max",
               head_dropout = 0.2, loss = loss_focal(gamma = 2))
  )
)

Where the labelling is written follows from mode rather than being said twice. In nested_dirs the directories under train_path are the classes, with the images inside them - the layout nearly every image dataset ships in - and in config_file the CSV’s label column holds the name. subdirs is refused here, which is what keeps the two readings of nested_dirs apart: a sample is one image rather than a directory with the image in one subdirectory and what labels it in the other.

classifier() is the one model constructor in this package not named after an architecture, and the missing field is the point rather than an omission. The network is a contracting path, pooled, with one dense layer on top, so encoder is the whole answer: unset it is ours, sized by blocks and filters, and named it is any of available_encoders(). The two models above are the comparison this task makes easiest - ours against a pretrained backbone on the same split. blocks, filters and block_width are refused beside encoder rather than ignored, because a backbone’s own depth is then the only one there is, and available_models("classification") says all of this instead of returning rows.

metric_auc() is the default and that is a measurement rather than a preference. On 40000 images at a prevalence of 3%, with the model’s own probabilities as the truth so that only its discrimination varies, accuracy spans 0.0100 where the AUC spans 0.4631 - and accuracy is not monotone in quality: the model with no signal at all scores above two that have one. metric_balanced_accuracy() and metric_f1() do not rescue it, moving 0.0000 and -0.0005 between a real signal and a collapse, because a calibrated model at that prevalence never crosses the threshold and everything read off the hard prediction sees the same all-negative answer. class_balance() is what makes that prevalence visible before a run rather than after it.

A split cut out of the training folder is stratified, which is the one part of the walk this task overrides. Measured over 400 seeds, a random 70/15/15 of 100 images at a 3% prevalence leaves some part with none of the rare class 91% of the time, 64% at 200 images and 20% at 500 - and a per-class metric cannot average over a class its pass never held, so the run dies at the end of the first epoch with the epoch already paid for. group_by beside it is refused when a group’s images do not share a label, because keeping a patient on one side and keeping a class on both are then two instructions that cannot both be followed.

What predict() gives back is a table rather than an array, which is the shape the rest of R wants. type = "class" is one row per image with key, label, predicted and probability, and type = "probability" is a matrix of images by classes with the keys as row names - so either one joins to whatever else you hold per image. space and the cases/each pair are refused by name: they describe a grid, and a split of masks too large to hold at once, and neither is a thing one label per image has. evaluate_cases() is refused for a different reason - a case here is one image whose score is 0 or 1, so a split’s worst case is 0 for any model that gets a single image wrong.

task: classification
data:
  train_path: data/train.csv
  validation_path: data/validation.csv
  mode: config_file
  classes: [atelectasis, effusion, nodule, mass]
  labels_from: [Atelectasis, Effusion, Nodule, Mass]
models:
  - name: ours
    input_shape: [224, 224]
    blocks: 4
    filters: 16
    loss: {name: cce}
    metrics: [{name: auc}, {name: f1}]

spec <- platypus_spec(
  data = classification_data(
    "data/train.csv", "data/validation.csv",
    classes = c("atelectasis", "effusion", "nodule", "mass"),
    labels_from = c("Atelectasis", "Effusion", "Nodule", "Mass"),
    mode = "config_file"
  ),
  models = list(
    classifier("ours", input_shape = c(224, 224), blocks = 4, filters = 16,
               loss = loss_cce(), metrics = list(metric_auc(), metric_f1()))
  )
)

task: classification
data:
  train_path: data/train.csv
  validation_path: data/validation.csv
  test_path: data/to-predict.csv
  mode: config_file
  classes: [benign, malignant]
models:
  - name: ours
    input_shape: [224, 224]
    blocks: 4
    filters: 16

Not training at all

task: semantic_segmentation
data:
  train_path: data/images
  validation_path: data/images
  colormap: [[0, 0, 0], [255, 255, 255]]
models:
  - name: published
    input_shape: [256, 256]
    weights: dsbowl-unet
    fit: false

weights takes a registry name, a local path, or hf://owner/repo/file@commit - and the hf:// form requires the commit, because weights that change under a stable name are the worst kind of irreproducibility. A registry name carries one already; available_weights() lists them.

With fit: false nothing trains and predict(fit) works immediately. The weights are checked against the model they are being loaded into before torch is touched: a shape, a channel count or a class count that does not match is a refusal naming the field, not a tensor-size error from inside a loss.

What is checked, and when

Validation happens in this order, and the first two cost no epochs:

1. The file, against the schema. Unknown fields, missing required ones, a name: that no block has, a callback watching a number nothing will produce. Before any path is read - and before Python starts, for the checks R can make itself.
2. The data on disk, when the run begins: that the folders exist, that images and masks correspond, and that every class in colormap or labels actually appears in the masks, because a declared class appearing in none has a target channel of zeros, so the model is never shown what it is being asked to find. inspect_masks() runs that check on its own, and platypus_fit(check_masks = FALSE) skips it.
3. Everything else, during the run.

A refusal names the field and says what to do. The engine’s own messages are translated into R conditions rather than passed through as Python tracebacks, so tryCatch(platypus_spec("run.yaml"), error = function(e) conditionMessage(e)) gives a sentence rather than a stack.

The same run from code

The file is not a second implementation. These describe the same experiment:

spec <- platypus_spec("run.yaml")

spec <- platypus_spec(
  data = segmentation_data(
    "data/train",
    validation = "data/validation",
    colormap = list(c(0, 0, 0), c(255, 255, 255))
  ),
  models = list(u_net("first_try", input_shape = c(256, 256)))
)

And as.list(spec) is the file, so a run built from arguments can be written out and kept:

yaml::write_yaml(as.list(spec), "run.yaml")

Which to use is a question about how many models and how often, not about what is possible. A file is better when there are several models, when the run is going to be repeated, or when somebody else has to see what was done; the arguments are better inside a loop over something the file cannot express, like a vector of datasets.

Note that as.list() is also how the fields are reached: the object returned by platypus_spec() carries the specification on the Python side, so spec$task is NULL while as.list(spec)$task is what you want.