File size: 8,056 Bytes
1be89f3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 |
import gc
import json
from pathlib import Path
from typing import Any
import click
import polars as pl
import torch
from yambda.constants import Constants
from yambda.evaluation.metrics import calc_metrics
from yambda.evaluation.ranking import Embeddings, Targets, rank_items
from yambda.processing import timesplit
from yambda.utils import argmax
DEFAULT_GRIDS = {
"50m": {
"likes": [0, 0.001, 0.25, 0.5, 1, 2, 3, 4, 6],
"listens": [0, 0.001, 0.002, 0.004, 0.008, 0.016, 0.032, 0.064, 0.128, 0.256, 0.5, 1.0, 2],
},
"500m": {
"likes": [0, 0.002, 0.004, 0.008, 0.016, 0.032, 0.064, 0.128, 0.256, 0.5, 1.0],
"listens": [0, 0.001, 0.002, 0.004, 0.008],
},
}
@click.command()
@click.option(
'--data_dir', required=True, type=str, default="../../data/flat", show_default=True, help="Expects flat data"
)
@click.option(
'--size',
required=True,
type=click.Choice(['50m', '500m']), # 5b is not supported
default="50m",
multiple=False,
show_default=True,
help="5b is not supported due to (num_user, num_user) matrix",
)
@click.option(
'--interaction',
required=True,
type=click.Choice(['likes', 'listens']),
default="likes",
multiple=False,
show_default=True,
)
@click.option(
'--hours',
required=True,
type=float,
default=[-1],
multiple=True,
show_default=True,
help="Hyperparameter. If -1 default grid will be used",
)
@click.option('--validation_metric', required=True, type=str, default="ndcg@100", show_default=True)
@click.option('--report_metrics', required=True, type=str, default=Constants.METRICS, multiple=True, show_default=True)
@click.option('--device', required=True, type=str, default="cuda:0", show_default=True)
def main(
data_dir: str,
size: str,
interaction: str,
hours: list[float],
validation_metric: str,
report_metrics: list[str],
device: str,
):
print(f"REPORT METRICS: {report_metrics}")
print(f"SIZE {size}, INTERACTION {interaction}")
result = item_knn(
data_dir,
size,
interaction,
device,
hours=hours if hours[0] != -1 else DEFAULT_GRIDS[size][interaction],
validation_metric=validation_metric,
report_metrics=report_metrics,
)
print(json.dumps(result, indent=2))
def scan(path: str, dataset_size: str, dataset_name: str) -> pl.LazyFrame:
path = Path(path) / dataset_size / dataset_name
df = pl.scan_parquet(path.with_suffix(".parquet"))
return df
def preprocess(
df: pl.LazyFrame, interaction: str, val_size: int
) -> tuple[pl.LazyFrame, pl.LazyFrame | None, pl.LazyFrame]:
if interaction == "listens":
df = df.filter(pl.col("played_ratio_pct") >= Constants.TRACK_LISTEN_THRESHOLD)
train, val, test = timesplit.flat_split_train_val_test(
df, val_size=val_size, test_timestamp=Constants.TEST_TIMESTAMP
)
return (
train,
val.collect(engine="streaming").lazy() if val is not None else None,
test.collect(engine="streaming").lazy(),
)
def eliminate_zeros(x: torch.Tensor, threshold: float = 1e-9) -> torch.Tensor:
mask = (x._values() > threshold).nonzero()
nv = x._values().index_select(0, mask.view(-1))
ni = x._indices().index_select(1, mask.view(-1))
return torch.sparse_coo_tensor(ni, nv, x.shape)
def create_weighted_sparse_tensor(train: pl.LazyFrame, tau: float) -> torch.Tensor:
uid_mapping = (
train.select("uid").unique().with_columns(pl.col("uid").rank(method="dense").alias("uid_idx") - 1).collect()
)
item_mapping = (
train.select("item_id")
.unique()
.with_columns(pl.col("item_id").rank(method="dense").alias("item_idx") - 1)
.collect()
)
processed = (
train.with_columns(pl.max("timestamp").over("uid").alias("max_timestamp"))
.with_columns((pl.col("max_timestamp") - pl.col("timestamp")).alias("delta"))
.with_columns((tau ** pl.col("delta")).alias("weight"))
.join(uid_mapping.lazy(), on="uid", how="inner")
.join(item_mapping.lazy(), on="item_id", how="inner")
)
coo_data = processed.group_by(["uid_idx", "item_idx"]).agg(pl.sum("weight").alias("total_weight")).collect()
indices = torch.concat([coo_data["uid_idx"].to_torch()[None, :], coo_data["item_idx"].to_torch()[None, :]], dim=0)
values = torch.tensor(coo_data["total_weight"].to_numpy(), dtype=torch.float)
return eliminate_zeros(
torch.sparse_coo_tensor(
indices=indices, values=values, size=(uid_mapping["uid_idx"].max() + 1, item_mapping["item_idx"].max() + 1)
)
)
def sparse_normalize(sparse_tensor: torch.Tensor, dim=0, eps=1e-12):
indices = sparse_tensor.coalesce().indices()
values = sparse_tensor.coalesce().values()
unique_dim_indices, inverse = torch.unique(indices[dim], return_inverse=True)
squared_values = values**2
sum_squared = torch.zeros_like(unique_dim_indices, dtype=torch.float32)
sum_squared.scatter_add_(0, inverse, squared_values)
norms = torch.sqrt(sum_squared + eps)
normalized_values = values / norms[inverse]
return torch.sparse_coo_tensor(indices, normalized_values, sparse_tensor.size())
def training(
train: pl.LazyFrame, hour: float, user_item: torch.Tensor, user_ids: torch.Tensor, device: str, decay: float = 0.9
) -> Embeddings:
tau = 0.0 if hour == 0 else decay ** (1 / 24 / 60 / 60 / (hour / 24))
user_item_with_tau = create_weighted_sparse_tensor(train, tau)
user_embeddings = (user_item_with_tau @ user_item.T).to_dense()
user_embeddings = torch.nn.functional.normalize(user_embeddings, dim=-1)
return Embeddings(user_ids, user_embeddings.to(device))
def evaluation(
train: pl.LazyFrame, val: pl.LazyFrame, device: str, hours: list[float], metrics: list[str]
) -> list[dict[str, Any]]:
num_ranked_items = max([int(x.split("@")[1]) for x in metrics])
unique_user_ids = train.select("uid").unique().sort("uid").collect(engine="streaming")["uid"].to_torch().to(device)
unique_item_ids = (
train.select("item_id").unique().sort("item_id").collect(engine="streaming")["item_id"].to_torch().to(device)
)
user_item = create_weighted_sparse_tensor(train, 1.0)
item_embeddings = sparse_normalize(user_item.T.to(device), dim=-1)
item_embeddings = Embeddings(unique_item_ids, item_embeddings)
targets = Targets.from_sequential(
val.group_by('uid', maintain_order=True).agg(pl.all().exclude('uid')).select(['uid', 'item_id']),
device,
)
hour2metrics = []
for hour in hours:
user_embeddings = training(
train=train,
hour=hour,
user_item=user_item,
user_ids=unique_user_ids,
device=device,
)
ranked = rank_items(
users=user_embeddings,
items=item_embeddings,
num_items=num_ranked_items,
batch_size=128,
)
del user_embeddings
gc.collect()
hour2metrics.append(calc_metrics(ranked, targets, metrics))
del unique_user_ids
del unique_item_ids
del item_embeddings
del targets
gc.collect()
return hour2metrics
def item_knn(
data_dir: str,
size: str,
interaction: str,
device: str,
hours: list[float],
validation_metric: str,
report_metrics: list[str],
) -> dict[str, Any]:
df = scan(data_dir, size, interaction)
# hyperopt by validation
train, val, _ = preprocess(df, interaction, val_size=Constants.VAL_SIZE)
results = evaluation(train, val, device, hours, [validation_metric])
metric_name, k = validation_metric.split('@')
best_hour = hours[argmax(results, lambda x: x[metric_name][int(k)])]
print(f"FINAL HYPERPARAMS {best_hour=}")
# train final model
train, _, test = preprocess(df, interaction, val_size=0)
return evaluation(train, test, device, [best_hour], report_metrics)[0]
if __name__ == "__main__":
main()
|