max:<n> balancing strategy

Manifest.toml

@@ -312,11 +312,11 @@ version = "1.9.0"
 
 [[deps.PlanarLarvae]]
 deps = ["DelimitedFiles", "HDF5", "JSON3", "LinearAlgebra", "MAT", "Meshes", "OrderedCollections", "Random", "SHA", "StaticArrays", "Statistics", "StatsBase", "StructTypes"]
-git-tree-sha1 = "ef6169e9f8705569925bef897704c7514b4d5f18"
+git-tree-sha1 = "025970ba7e5b0b9455b2487c1ad2150a17edba0c"
 repo-rev = "main"
 repo-url = "https://gitlab.pasteur.fr/nyx/planarlarvae.jl"
 uuid = "c2615984-ef14-4d40-b148-916c85b43307"
-version = "0.12.0"
+version = "0.13.0"
 
 [[deps.PrecompileTools]]
 deps = ["Preferences"]

src/LarvaDatasets.jl

@@ -556,6 +556,7 @@ function write_larva_dataset_hdf5(output_dir::String,
         seed=nothing,
         distributed_sampling=true,
         past_future_extensions=true,
+        balancing_strategy=nothing,
         )
     if distributed_sampling
         new_write_larva_dataset_hdf5(output_dir, input_data;
@@ -571,10 +572,12 @@ function write_larva_dataset_hdf5(output_dir::String,
             fixmwt=fixmwt,
             frameinterval=frameinterval,
             includeall=includeall,
+            balancing_strategy=balancing_strategy,
             past_future_extensions=past_future_extensions,
             seed=seed)
     else
         past_future_extensions || throw("not implemented")
+        isnothing(balancing_strategy) || throw("not implemented")
         legacy_write_larva_dataset_hdf5(output_dir, input_data, window_length;
             labels=labels,
             labelpointers=labelpointers,
@@ -595,7 +598,8 @@ function new_write_larva_dataset_hdf5(output_dir, input_data;
         window_length=20,
         labels=nothing,
         file_filter=nothing,
-        balance=true,
+        balance=nothing,
+        balancing_strategy=nothing,
         frameinterval=0.1,
         includeall="edited",
         past_future_extensions=true,
@@ -603,6 +607,8 @@ function new_write_larva_dataset_hdf5(output_dir, input_data;
         kwargs...)
     repo = if input_data isa String
         isnothing(file_filter) ? Repository(input_data) : Repository(input_data, file_filter)
+    elseif input_data isa Repository
+        input_data
     else
         Repository(repo = pwd(), files = eltype(input_data) === String ? preload.(input_data) : input_data)
     end
@@ -613,23 +619,31 @@ function new_write_larva_dataset_hdf5(output_dir, input_data;
     @assert !isnothing(frameinterval)
     window = TimeWindow(window_length * frameinterval, round(Int, 1 / frameinterval);
                         maggotuba_compatibility=past_future_extensions)
-    selectors = isnothing(labels) ? getprimarylabels(first(Dataloaders.files(repo))) : labels
+    index = if startswith(balancing_strategy, "max:")
+        # `includeall` not supported
+        maxcount = parse(Int, balancing_strategy[5:end])
+        capacitysampling(labels, maxcount; seed=seed)
+    else
+        if isnothing(balance)
+            balance = lowercase(balancing_strategy) == "maggotuba"
+        end
         min_max_ratio = balance ? 2 : 20
-    index = if isnothing(includeall)
-        ratiobasedsampling(selectors, min_max_ratio; seed=seed)
+        if isnothing(includeall)
+            ratiobasedsampling(labels, min_max_ratio; seed=seed)
         else
-        ratiobasedsampling(selectors, min_max_ratio, prioritylabel(includeall); seed=seed)
+            ratiobasedsampling(labels, min_max_ratio, prioritylabel(includeall); seed=seed)
+        end
     end
     loader = DataLoader(repo, window, index)
     try
         buildindex(loader; unload=true)
-    catch ArgumentError
+    catch
         # most likely error message: "collection must be non-empty"
         @error "Most likely cause: no time segments could be isolated"
         rethrow()
     end
-    total_sample_size = length(loader.index)
-    classcounts, _ = Dataloaders.groupby(selectors, loader.index.targetcounts)
+    total_sample_size = length(index)
+    classcounts, _ = Dataloaders.groupby(index.sampler.selectors, index.targetcounts)
     #
     extended_window_length = past_future_extensions ? 3 * window_length : window_length
     date = Dates.format(Dates.now(), "yyyy_mm_dd")

src/taggingbackends/explorer.py

@@ -484,9 +484,9 @@ run `poetry add {pkg}` from directory: \n
         return input_files, labels
 
     def generate_dataset(self, input_files,
-            labels=None, window_length=20, sample_size=None, balance=True,
+            labels=None, window_length=20, sample_size=None, balance=None,
             include_all=None, frame_interval=None, seed=None,
-            past_future_extensions=None):
+            past_future_extensions=None, balancing_strategy=None):
         """
         Generate a *larva_dataset hdf5* file in data/interim/{instance}/
         """
@@ -502,6 +502,7 @@ run `poetry add {pkg}` from directory: \n
                 balance=balance,
                 includeall=include_all,
                 frameinterval=frame_interval,
+                balancing_strategy=balancing_strategy,
                 past_future_extensions=past_future_extensions,
                 seed=seed)
 

src/taggingbackends/main.py

@@ -45,8 +45,8 @@ algorithm aims to minimize. Weights are also specified as a comma-separated
 list of floating-point values. As many weights as labels are expected.
 
 In addition to class penalties, the majority classes can be subsampled in
-different ways. Argument `--balancing-strategy` can take either "maggotuba" or
-"auto", that correspond to specific subsampling strategies.
+different ways. Argument `--balancing-strategy` can take values "maggotuba",
+"auto" or "max:<n>", that correspond to specific subsampling strategies.
 
 --balancing-strategy maggotuba
     Denoting n the size of the minority class, classes of size less than 10n are
@@ -60,6 +60,10 @@ different ways. Argument `--balancing-strategy` can take either "maggotuba" or
     In addition, if class weights are not defined, they are set as the inverse
     of the corresponding class size.
 
+--balancing-strategy max:<n>
+    <n> is the maximum count for any given class. Classes with fewer occurences
+    in the dataset are fully sampled.
+
 Subsampling is done at random. However, data observations bearing a specific
 secondary label can be included with priority, up to the target size if too
 many, and then complemented with different randomly picked observations. To