e6localmap: new version using ALS factors

.gitattributes

@@ -1,3 +1,4 @@
 *.txt filter=lfs diff=lfs merge=lfs -text
 *.json filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text

data/{model-verybighalfbrain2e60.safetensors → implicit_tag_factors.safetensors}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:37fc80f31f158d5de026446b277b6da112677b72196994740b6d26d43e7c88f5
-size 70397688
+oid sha256:dfea51d9db1bd413aedb1781cc4b936c41c117a78221bc18dfbc6dface513ef7
+size 33554520

data/{tags.txt → tag2idx.json.gz}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10cca1eba52aa3b36145f6c913911067221c660b51000e51dbaf598edb77b289
-size 424308
+oid sha256:a4c98ec5f2b8488abd68ef617425c4d1fc1afd5962898d1dd78963444075485a
+size 4253419

data/tags.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1d0910d099345c89fedb6eb8acdac1312d18758e8c3678910602f7a42435803
+size 2875605

e6db/utils.py

@@ -0,0 +1,65 @@
+from pathlib import Path
+import gzip
+import json
+
+# FIXME: split this module by dependencies
+import numpy as np
+import scipy.sparse
+import torch
+import polars as pl
+from polars import col
+
+id2tagcat = {
+    0: "general",
+    1: "artist",
+    3: "copyright",
+    4: "character",
+    5: "species",
+    6: "invalid",
+    7: "meta",
+    8: "lore",
+}
+tagcat2id = {v: k for k, v in id2tagcat.items()}
+
+
+def load_tags(data_dir):
+    data_dir = Path(data_dir)
+    with gzip.open(data_dir / "tags.txt.gz", 'rt') as fd:
+        idx2tag = fd.read().split("\n")
+        if not idx2tag[-1]:
+            idx2tag = idx2tag[:-1]
+    with gzip.open(data_dir / "tag2idx.json.gz", "rb") as fp:
+        tag2idx = json.load(fp)
+    return tag2idx, idx2tag
+
+
+def filter_tag_list(column, vocab_size):
+
+    return col(column).list.eval(pl.element().filter(pl.element() < vocab_size))
+
+
+def tags_to_csr(df_posts, column="stripped_tags"):
+    from polars import col
+
+    indices = df_posts[column].explode().drop_nulls().to_numpy()
+    indptr = df_posts.select(
+        offset=col(column).list.len().cum_sum().shift(1, fill_value=0)
+    )["offset"].to_numpy()
+    return indices, indptr
+
+
+def tags_to_scipy_csr(df_posts, column="stripped_tags", vocab_size=None):
+    indices, indptr = tags_to_csr(df_posts, column=column)
+    indptr = np.r_[indptr, len(indices)]
+    R = scipy.sparse.csr_array(
+        (np.ones(indices.shape[0], dtype=np.uint8), indices, indptr),
+        shape=(len(df_posts), vocab_size or indices.max() + 1),
+    )
+    return R
+
+
+def tags_to_tensors(df_posts, column="stripped_tags", device=None, dtype=torch.int32):
+    indices, indptr = tags_to_csr(df_posts, column=column)
+    tags = torch.tensor(indices, dtype=dtype, device=device)
+    offsets = torch.tensor(indptr, dtype=dtype, device=device)
+    return tags, offsets