refactor: inference upds

src/deep_neurographs/machine_learning/inference.py

@@ -13,6 +13,176 @@
 from torch.nn.functional import sigmoid
 from torch.utils.data import DataLoader
 
+CHUNK_SHAPE = (512, 512, 512)
+
+
+from copy import deepcopy
+from random import sample
+
+import fastremap
+import networkx as nx
+
+from deep_neurographs import graph_utils as gutils
+from deep_neurographs import utils
+from deep_neurographs.neurograph import NeuroGraph
+
+
+
+
+def run(
+    neurograph,
+    model_type,
+    model_path,
+    labels_path,
+    proposals,
+    batch_size_proposals=2000,
+    confidence_threshold=0.7,
+    output_dir=None,
+    seeds=None,
+):
+    if seeds:
+        run_with_seeds(
+            neurograph,
+            model_type,
+            model_path,
+            labels_path,
+            proposals,
+            confidence_threshold=0.7,
+            output_dir=None,
+            seeds=None,
+        )
+    else:
+        run_without_seeds(
+            neurograph,
+            model_type,
+            model_path,
+            labels_path,
+            proposals,
+            confidence_threshold=0.7,
+            output_dir=None,
+        )
+
+
+
+def run_with_seeds(
+    neurograph,
+    model_type,
+    model_path,
+    labels_path,
+    proposals,
+    seeds,
+    confidence_threshold=0.7,
+    output_dir=None,
+):
+    # build seed graphs
+    # --> grow graph to some limit
+    # --> run inference on each seed graph in parellel
+    # --> sequentially combine results
+    # --> repeat
+    # check whether any proposals remain
+    # --> call run_without_seeds
+    pass
+
+
+def run_without_seeds(
+    neurograph,
+    model_type,
+    model_path,
+    labels_path,
+    proposals,
+    confidence_threshold=0.7,
+    output_dir=None,
+):
+    # form batches of proposals wrt distance
+    # --> generate features
+    # --> run prediction
+    # --> parse predictions and merge proposals
+
+
+def build_from_soma(
+    neurograph, labels_path, chunk_origin, chunk_shape=CHUNK_SHAPE, n_hops=1
+):
+    swc_ids = get_swc_ids(labels_path, chunk_origin, chunk_shape)
+    seed_neurograph = build_seed_neurograph(neurograph, swc_ids)
+
+    # loop
+    pass
+
+
+def get_swc_ids(path, xyz, chunk_shape, from_center=True):
+    img = utils.open_tensorstore(path, "neuroglancer_precomputed")
+    img = utils.read_tensorstore(
+        img, xyz, chunk_shape, from_center=from_center
+    )
+    return set(fastremap.unique(img).astype(int))
+
+
+def build_seed_neurograph(neurograph, swc_ids):
+    seed_neurograph = NeuroGraph()
+    for nodes in nx.connected_components(neurograph):
+        i = sample_singleton(nodes)
+        swc_id = int(neurograph.nodes[i]["swc_id"])
+        if swc_id in swc_ids:
+            seed_neurograph.update(neurograph.subgraph(nodes))
+            seed_neurograph.add_swc_id(swc_id)
+    return seed_neurograph
+
+
+def sample_singleton(my_container):
+    return sample(my_container, 1)[0]
+
+
+def build_from_boundary(neurograph, pred_neurograph, boundary_components):
+    new_bdd = list()
+    for c in boundary_components:
+        pred_neurograph, bdd = expand_boundary(neurograph, pred_neurograph, c)
+        new_bdd.extend(bdd)
+    return pred_neurograph, new_bdd
+
+
+def expand_boundary(neurograph, pred_neurograph, component):
+    bdd = list()
+    for i in component:
+        for j in neurograph.nodes[i]["proposals"]:
+            # Extract proposals info
+            edge = frozenset({i, j})
+            swc_id = int(neurograph.nodes[j]["swc_id"])
+
+            # Add component to graph
+            if swc_id not in pred_neurograph.swc_ids:
+                c = gutils.get_component(neurograph, j)
+                pred_neurograph.add_swc_id(swc_id)
+                pred_neurograph = ingest_subgraph(
+                    neurograph, pred_neurograph, c
+                )
+                bdd.append(c)
+
+            # Add proposal to graph
+            pred_neurograph.proposals[edge] = deepcopy(
+                neurograph.proposals[edge]
+            )
+            pred_neurograph.nodes[i]["proposals"].add(j)
+            pred_neurograph.nodes[j]["proposals"].add(i)
+
+    return pred_neurograph, bdd
+
+
+def ingest_subgraph(neurograph_1, neurograph_2, node_subset):
+    # Add structural components
+    subgraph = nx.subgraph(neurograph_1, node_subset)
+    neurograph_2.add_nodes_from(node_subset)
+    neurograph_2.add_edges_from(subgraph.edges)
+
+    # Add attributes
+    for node in node_subset:
+        neurograph_2.nodes[node].update(neurograph_1.nodes[node])
+        neurograph_2.nodes[node]["proposals"] = set()
+
+    for edge in subgraph.edges:
+        neurograph_2.edges[edge].update(neurograph_1.edges[edge])
+
+    return neurograph_2
+
 
 def predict(dataset, model, model_type):
     accuracy = []