nextrec 0.4.33__py3-none-any.whl → 0.5.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- nextrec/__version__.py +1 -1
- nextrec/basic/activation.py +10 -18
- nextrec/basic/asserts.py +1 -22
- nextrec/basic/callback.py +2 -2
- nextrec/basic/features.py +6 -37
- nextrec/basic/heads.py +13 -1
- nextrec/basic/layers.py +33 -123
- nextrec/basic/loggers.py +3 -2
- nextrec/basic/metrics.py +85 -4
- nextrec/basic/model.py +518 -7
- nextrec/basic/summary.py +88 -42
- nextrec/cli.py +117 -30
- nextrec/data/data_processing.py +8 -13
- nextrec/data/preprocessor.py +449 -844
- nextrec/loss/grad_norm.py +78 -76
- nextrec/models/multi_task/ple.py +1 -0
- nextrec/models/multi_task/share_bottom.py +1 -0
- nextrec/models/ranking/afm.py +4 -9
- nextrec/models/ranking/dien.py +7 -8
- nextrec/models/ranking/ffm.py +2 -2
- nextrec/models/retrieval/sdm.py +1 -2
- nextrec/models/sequential/hstu.py +0 -2
- nextrec/models/tree_base/base.py +1 -1
- nextrec/utils/__init__.py +2 -1
- nextrec/utils/config.py +1 -1
- nextrec/utils/console.py +1 -1
- nextrec/utils/onnx_utils.py +252 -0
- nextrec/utils/torch_utils.py +63 -56
- nextrec/utils/types.py +43 -0
- {nextrec-0.4.33.dist-info → nextrec-0.5.0.dist-info}/METADATA +10 -4
- {nextrec-0.4.33.dist-info → nextrec-0.5.0.dist-info}/RECORD +34 -42
- nextrec/models/multi_task/[pre]star.py +0 -192
- nextrec/models/representation/autorec.py +0 -0
- nextrec/models/representation/bpr.py +0 -0
- nextrec/models/representation/cl4srec.py +0 -0
- nextrec/models/representation/lightgcn.py +0 -0
- nextrec/models/representation/mf.py +0 -0
- nextrec/models/representation/s3rec.py +0 -0
- nextrec/models/sequential/sasrec.py +0 -0
- nextrec/utils/feature.py +0 -29
- {nextrec-0.4.33.dist-info → nextrec-0.5.0.dist-info}/WHEEL +0 -0
- {nextrec-0.4.33.dist-info → nextrec-0.5.0.dist-info}/entry_points.txt +0 -0
- {nextrec-0.4.33.dist-info → nextrec-0.5.0.dist-info}/licenses/LICENSE +0 -0
nextrec/__version__.py
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
__version__ = "0.
|
|
1
|
+
__version__ = "0.5.0"
|
nextrec/basic/activation.py
CHANGED
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
"""
|
|
2
|
-
Activation function definitions
|
|
2
|
+
Activation function definitions.
|
|
3
3
|
|
|
4
4
|
Date: create on 27/10/2025
|
|
5
|
-
Checkpoint: edit on
|
|
5
|
+
Checkpoint: edit on 20/01/2026
|
|
6
6
|
Author: Yang Zhou, zyaztec@gmail.com
|
|
7
7
|
"""
|
|
8
8
|
|
|
@@ -22,26 +22,18 @@ class Dice(nn.Module):
|
|
|
22
22
|
where p(x) = sigmoid((x - E[x]) / sqrt(Var[x] + epsilon))
|
|
23
23
|
"""
|
|
24
24
|
|
|
25
|
-
def __init__(self, emb_size: int, epsilon: float = 1e-
|
|
25
|
+
def __init__(self, emb_size: int, epsilon: float = 1e-3):
|
|
26
26
|
super(Dice, self).__init__()
|
|
27
|
-
self.epsilon = epsilon
|
|
28
27
|
self.alpha = nn.Parameter(torch.zeros(emb_size))
|
|
29
|
-
self.bn = nn.BatchNorm1d(emb_size)
|
|
28
|
+
self.bn = nn.BatchNorm1d(emb_size, eps=epsilon, affine=False)
|
|
30
29
|
|
|
31
30
|
def forward(self, x):
|
|
32
|
-
#
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
x = x.view(-1, emb_size)
|
|
39
|
-
x_norm = self.bn(x)
|
|
40
|
-
p = torch.sigmoid(x_norm)
|
|
41
|
-
output = p * x + (1 - p) * self.alpha * x
|
|
42
|
-
if len(original_shape) == 3:
|
|
43
|
-
output = output.view(original_shape)
|
|
44
|
-
return output
|
|
31
|
+
# keep original shape for reshaping back after batch norm
|
|
32
|
+
orig_shape = x.shape # x: [N, L, emb_size] or [N, emb_size]
|
|
33
|
+
x2 = x.reshape(-1, orig_shape[-1]) # x2:[N*L, emb_size] or [N, emb_size]
|
|
34
|
+
x_norm = self.bn(x2)
|
|
35
|
+
p = torch.sigmoid(x_norm).reshape(orig_shape)
|
|
36
|
+
return x * (self.alpha + (1 - self.alpha) * p)
|
|
45
37
|
|
|
46
38
|
|
|
47
39
|
def activation_layer(
|
nextrec/basic/asserts.py
CHANGED
|
@@ -8,7 +8,7 @@ Author: Yang Zhou, zyaztec@gmail.com
|
|
|
8
8
|
|
|
9
9
|
from __future__ import annotations
|
|
10
10
|
|
|
11
|
-
from nextrec.utils.types import TaskTypeName
|
|
11
|
+
from nextrec.utils.types import TaskTypeName
|
|
12
12
|
|
|
13
13
|
|
|
14
14
|
def assert_task(
|
|
@@ -49,24 +49,3 @@ def assert_task(
|
|
|
49
49
|
raise ValueError(
|
|
50
50
|
f"{model_name} requires task length {nums_task}, got {len(task)}."
|
|
51
51
|
)
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
def assert_training_mode(
|
|
55
|
-
training_mode: TrainingModeName | list[TrainingModeName],
|
|
56
|
-
nums_task: int,
|
|
57
|
-
*,
|
|
58
|
-
model_name: str,
|
|
59
|
-
) -> None:
|
|
60
|
-
valid_modes = {"pointwise", "pairwise", "listwise"}
|
|
61
|
-
if not isinstance(training_mode, list):
|
|
62
|
-
raise TypeError(
|
|
63
|
-
f"[{model_name}-init Error] training_mode must be a list with length {nums_task}."
|
|
64
|
-
)
|
|
65
|
-
if len(training_mode) != nums_task:
|
|
66
|
-
raise ValueError(
|
|
67
|
-
f"[{model_name}-init Error] training_mode list length must match number of tasks."
|
|
68
|
-
)
|
|
69
|
-
if any(mode not in valid_modes for mode in training_mode):
|
|
70
|
-
raise ValueError(
|
|
71
|
-
f"[{model_name}-init Error] training_mode must be one of {'pointwise', 'pairwise', 'listwise'}."
|
|
72
|
-
)
|
nextrec/basic/callback.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
Callback System for Training Process
|
|
3
3
|
|
|
4
4
|
Date: create on 27/10/2025
|
|
5
|
-
Checkpoint: edit on
|
|
5
|
+
Checkpoint: edit on 21/01/2026
|
|
6
6
|
Author: Yang Zhou, zyaztec@gmail.com
|
|
7
7
|
"""
|
|
8
8
|
|
|
@@ -69,7 +69,7 @@ class Callback:
|
|
|
69
69
|
|
|
70
70
|
class CallbackList:
|
|
71
71
|
"""
|
|
72
|
-
Generates a list of callbacks
|
|
72
|
+
Generates a list of callbacks, used to manage and invoke multiple callbacks during training.
|
|
73
73
|
"""
|
|
74
74
|
|
|
75
75
|
def __init__(self, callbacks: Optional[list[Callback]] = None):
|
nextrec/basic/features.py
CHANGED
|
@@ -8,10 +8,9 @@ Author: Yang Zhou, zyaztec@gmail.com
|
|
|
8
8
|
|
|
9
9
|
import torch
|
|
10
10
|
|
|
11
|
-
from typing import Literal
|
|
12
|
-
|
|
13
11
|
from nextrec.utils.embedding import get_auto_embedding_dim
|
|
14
|
-
from nextrec.utils.
|
|
12
|
+
from nextrec.utils.torch_utils import to_list
|
|
13
|
+
from nextrec.utils.types import EmbeddingInitType, SequenceCombinerType
|
|
15
14
|
|
|
16
15
|
|
|
17
16
|
class BaseFeature:
|
|
@@ -29,15 +28,7 @@ class EmbeddingFeature(BaseFeature):
|
|
|
29
28
|
embedding_name: str = "",
|
|
30
29
|
embedding_dim: int | None = None,
|
|
31
30
|
padding_idx: int = 0,
|
|
32
|
-
init_type:
|
|
33
|
-
"normal",
|
|
34
|
-
"uniform",
|
|
35
|
-
"xavier_uniform",
|
|
36
|
-
"xavier_normal",
|
|
37
|
-
"kaiming_uniform",
|
|
38
|
-
"kaiming_normal",
|
|
39
|
-
"orthogonal",
|
|
40
|
-
] = "normal",
|
|
31
|
+
init_type: EmbeddingInitType = "normal",
|
|
41
32
|
init_params: dict | None = None,
|
|
42
33
|
l1_reg: float = 0.0,
|
|
43
34
|
l2_reg: float = 0.0,
|
|
@@ -73,23 +64,9 @@ class SequenceFeature(EmbeddingFeature):
|
|
|
73
64
|
max_len: int = 50,
|
|
74
65
|
embedding_name: str = "",
|
|
75
66
|
embedding_dim: int | None = None,
|
|
76
|
-
combiner:
|
|
77
|
-
"mean",
|
|
78
|
-
"sum",
|
|
79
|
-
"concat",
|
|
80
|
-
"dot_attention",
|
|
81
|
-
"self_attention",
|
|
82
|
-
] = "mean",
|
|
67
|
+
combiner: SequenceCombinerType = "mean",
|
|
83
68
|
padding_idx: int = 0,
|
|
84
|
-
init_type:
|
|
85
|
-
"normal",
|
|
86
|
-
"uniform",
|
|
87
|
-
"xavier_uniform",
|
|
88
|
-
"xavier_normal",
|
|
89
|
-
"kaiming_uniform",
|
|
90
|
-
"kaiming_normal",
|
|
91
|
-
"orthogonal",
|
|
92
|
-
] = "normal",
|
|
69
|
+
init_type: EmbeddingInitType = "normal",
|
|
93
70
|
init_params: dict | None = None,
|
|
94
71
|
l1_reg: float = 0.0,
|
|
95
72
|
l2_reg: float = 0.0,
|
|
@@ -143,15 +120,7 @@ class SparseFeature(EmbeddingFeature):
|
|
|
143
120
|
embedding_name: str = "",
|
|
144
121
|
embedding_dim: int | None = None,
|
|
145
122
|
padding_idx: int = 0,
|
|
146
|
-
init_type:
|
|
147
|
-
"normal",
|
|
148
|
-
"uniform",
|
|
149
|
-
"xavier_uniform",
|
|
150
|
-
"xavier_normal",
|
|
151
|
-
"kaiming_uniform",
|
|
152
|
-
"kaiming_normal",
|
|
153
|
-
"orthogonal",
|
|
154
|
-
] = "normal",
|
|
123
|
+
init_type: EmbeddingInitType = "normal",
|
|
155
124
|
init_params: dict | None = None,
|
|
156
125
|
l1_reg: float = 0.0,
|
|
157
126
|
l2_reg: float = 0.0,
|
nextrec/basic/heads.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
Task head implementations for NextRec models.
|
|
3
3
|
|
|
4
4
|
Date: create on 23/12/2025
|
|
5
|
-
Checkpoint: edit on
|
|
5
|
+
Checkpoint: edit on 22/01/2026
|
|
6
6
|
Author: Yang Zhou, zyaztec@gmail.com
|
|
7
7
|
"""
|
|
8
8
|
|
|
@@ -24,6 +24,12 @@ class TaskHead(nn.Module):
|
|
|
24
24
|
|
|
25
25
|
This wraps PredictionLayer so models can depend on a "Head" abstraction
|
|
26
26
|
without changing their existing forward signatures.
|
|
27
|
+
|
|
28
|
+
Args:
|
|
29
|
+
task_type: The type of task(s) this head is responsible for.
|
|
30
|
+
task_dims: The dimensionality of each task's output.
|
|
31
|
+
use_bias: Whether to include a bias term in the prediction layer.
|
|
32
|
+
return_logits: Whether to return raw logits or apply activation.
|
|
27
33
|
"""
|
|
28
34
|
|
|
29
35
|
def __init__(
|
|
@@ -56,6 +62,12 @@ class RetrievalHead(nn.Module):
|
|
|
56
62
|
|
|
57
63
|
It computes similarity for pointwise training/inference, and returns
|
|
58
64
|
raw embeddings for in-batch negative sampling in pairwise/listwise modes.
|
|
65
|
+
|
|
66
|
+
Args:
|
|
67
|
+
similarity_metric: The metric used to compute similarity between embeddings.
|
|
68
|
+
temperature: Scaling factor for similarity scores.
|
|
69
|
+
training_mode: The training mode, which can be pointwise, pairwise, or listwise.
|
|
70
|
+
apply_sigmoid: Whether to apply sigmoid activation to the similarity scores in pointwise mode.
|
|
59
71
|
"""
|
|
60
72
|
|
|
61
73
|
def __init__(
|
nextrec/basic/layers.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
Layer implementations used across NextRec.
|
|
3
3
|
|
|
4
4
|
Date: create on 27/10/2025
|
|
5
|
-
Checkpoint: edit on
|
|
5
|
+
Checkpoint: edit on 25/01/2026
|
|
6
6
|
Author: Yang Zhou, zyaztec@gmail.com
|
|
7
7
|
"""
|
|
8
8
|
|
|
@@ -20,15 +20,13 @@ import torch.nn.functional as F
|
|
|
20
20
|
from nextrec.basic.activation import activation_layer
|
|
21
21
|
from nextrec.basic.features import DenseFeature, SequenceFeature, SparseFeature
|
|
22
22
|
from nextrec.utils.torch_utils import get_initializer
|
|
23
|
-
from nextrec.utils.types import ActivationName
|
|
23
|
+
from nextrec.utils.types import ActivationName, TaskTypeName
|
|
24
24
|
|
|
25
25
|
|
|
26
26
|
class PredictionLayer(nn.Module):
|
|
27
27
|
def __init__(
|
|
28
28
|
self,
|
|
29
|
-
task_type:
|
|
30
|
-
Literal["binary", "regression"] | list[Literal["binary", "regression"]]
|
|
31
|
-
) = "binary",
|
|
29
|
+
task_type: TaskTypeName | list[TaskTypeName] = "binary",
|
|
32
30
|
task_dims: int | list[int] | None = None,
|
|
33
31
|
use_bias: bool = True,
|
|
34
32
|
return_logits: bool = False,
|
|
@@ -81,10 +79,12 @@ class PredictionLayer(nn.Module):
|
|
|
81
79
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
82
80
|
if x.dim() == 1:
|
|
83
81
|
x = x.unsqueeze(0) # (1 * total_dim)
|
|
84
|
-
if
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
82
|
+
if not torch.onnx.is_in_onnx_export():
|
|
83
|
+
if x.shape[-1] != self.total_dim:
|
|
84
|
+
raise ValueError(
|
|
85
|
+
f"[PredictionLayer Error]: Input last dimension ({x.shape[-1]}) does not match expected total dimension ({self.total_dim})."
|
|
86
|
+
)
|
|
87
|
+
|
|
88
88
|
logits = x if self.bias is None else x + self.bias
|
|
89
89
|
outputs = []
|
|
90
90
|
for task_type, (start, end) in zip(self.task_types, self.task_slices):
|
|
@@ -92,10 +92,9 @@ class PredictionLayer(nn.Module):
|
|
|
92
92
|
if self.return_logits:
|
|
93
93
|
outputs.append(task_logits)
|
|
94
94
|
continue
|
|
95
|
-
|
|
96
|
-
if task == "binary":
|
|
95
|
+
if task_type == "binary":
|
|
97
96
|
outputs.append(torch.sigmoid(task_logits))
|
|
98
|
-
elif
|
|
97
|
+
elif task_type == "regression":
|
|
99
98
|
outputs.append(task_logits)
|
|
100
99
|
else:
|
|
101
100
|
raise ValueError(
|
|
@@ -219,7 +218,7 @@ class EmbeddingLayer(nn.Module):
|
|
|
219
218
|
|
|
220
219
|
elif isinstance(feature, SequenceFeature):
|
|
221
220
|
seq_input = x[feature.name].long()
|
|
222
|
-
if feature.max_len is not None
|
|
221
|
+
if feature.max_len is not None:
|
|
223
222
|
seq_input = seq_input[:, -feature.max_len :]
|
|
224
223
|
|
|
225
224
|
embed = self.embed_dict[feature.embedding_name]
|
|
@@ -282,10 +281,11 @@ class EmbeddingLayer(nn.Module):
|
|
|
282
281
|
value = value.view(value.size(0), -1) # [B, input_dim]
|
|
283
282
|
input_dim = feature.input_dim
|
|
284
283
|
assert_input_dim = self.dense_input_dims.get(feature.name, input_dim)
|
|
285
|
-
if
|
|
286
|
-
|
|
287
|
-
|
|
288
|
-
|
|
284
|
+
if not torch.onnx.is_in_onnx_export():
|
|
285
|
+
if value.shape[1] != assert_input_dim:
|
|
286
|
+
raise ValueError(
|
|
287
|
+
f"[EmbeddingLayer Error]:Dense feature '{feature.name}' expects {assert_input_dim} inputs but got {value.shape[1]}."
|
|
288
|
+
)
|
|
289
289
|
if not feature.use_projection:
|
|
290
290
|
return value
|
|
291
291
|
dense_layer = self.dense_transforms[feature.name]
|
|
@@ -331,29 +331,10 @@ class InputMask(nn.Module):
|
|
|
331
331
|
feature: SequenceFeature,
|
|
332
332
|
seq_tensor: torch.Tensor | None = None,
|
|
333
333
|
):
|
|
334
|
-
if seq_tensor is not None
|
|
335
|
-
|
|
336
|
-
else:
|
|
337
|
-
values = x[feature.name]
|
|
338
|
-
values = values.long()
|
|
334
|
+
values = seq_tensor if seq_tensor is not None else x[feature.name]
|
|
335
|
+
values = values.long().view(values.size(0), -1)
|
|
339
336
|
padding_idx = feature.padding_idx if feature.padding_idx is not None else 0
|
|
340
|
-
mask = values != padding_idx
|
|
341
|
-
|
|
342
|
-
if mask.dim() == 1:
|
|
343
|
-
# [B] -> [B, 1, 1]
|
|
344
|
-
mask = mask.unsqueeze(1).unsqueeze(2)
|
|
345
|
-
elif mask.dim() == 2:
|
|
346
|
-
# [B, L] -> [B, 1, L]
|
|
347
|
-
mask = mask.unsqueeze(1)
|
|
348
|
-
elif mask.dim() == 3:
|
|
349
|
-
# [B, 1, L]
|
|
350
|
-
# [B, L, 1] -> [B, L] -> [B, 1, L]
|
|
351
|
-
if mask.size(1) != 1 and mask.size(2) == 1:
|
|
352
|
-
mask = mask.squeeze(-1).unsqueeze(1)
|
|
353
|
-
else:
|
|
354
|
-
raise ValueError(
|
|
355
|
-
f"InputMask only supports 1D/2D/3D tensors, got shape {values.shape}"
|
|
356
|
-
)
|
|
337
|
+
mask = (values != padding_idx).unsqueeze(1)
|
|
357
338
|
return mask.float()
|
|
358
339
|
|
|
359
340
|
|
|
@@ -897,30 +878,7 @@ class AttentionPoolingLayer(nn.Module):
|
|
|
897
878
|
self,
|
|
898
879
|
embedding_dim: int,
|
|
899
880
|
hidden_units: list = [80, 40],
|
|
900
|
-
activation:
|
|
901
|
-
"dice",
|
|
902
|
-
"relu",
|
|
903
|
-
"relu6",
|
|
904
|
-
"elu",
|
|
905
|
-
"selu",
|
|
906
|
-
"leaky_relu",
|
|
907
|
-
"prelu",
|
|
908
|
-
"gelu",
|
|
909
|
-
"sigmoid",
|
|
910
|
-
"tanh",
|
|
911
|
-
"softplus",
|
|
912
|
-
"softsign",
|
|
913
|
-
"hardswish",
|
|
914
|
-
"mish",
|
|
915
|
-
"silu",
|
|
916
|
-
"swish",
|
|
917
|
-
"hardsigmoid",
|
|
918
|
-
"tanhshrink",
|
|
919
|
-
"softshrink",
|
|
920
|
-
"none",
|
|
921
|
-
"linear",
|
|
922
|
-
"identity",
|
|
923
|
-
] = "sigmoid",
|
|
881
|
+
activation: ActivationName = "sigmoid",
|
|
924
882
|
use_softmax: bool = False,
|
|
925
883
|
):
|
|
926
884
|
super().__init__()
|
|
@@ -954,39 +912,22 @@ class AttentionPoolingLayer(nn.Module):
|
|
|
954
912
|
output: [batch_size, embedding_dim] - attention pooled representation
|
|
955
913
|
"""
|
|
956
914
|
batch_size, sequence_length, embedding_dim = keys.shape
|
|
957
|
-
|
|
958
|
-
|
|
959
|
-
|
|
960
|
-
|
|
961
|
-
|
|
962
|
-
# keys_length: (batch_size,)
|
|
963
|
-
device = keys.device
|
|
964
|
-
seq_range = torch.arange(sequence_length, device=device).unsqueeze(
|
|
965
|
-
0
|
|
966
|
-
) # (1, sequence_length)
|
|
967
|
-
mask = (seq_range < keys_length.unsqueeze(1)).unsqueeze(-1).float()
|
|
968
|
-
if mask is not None:
|
|
969
|
-
if mask.dim() == 2:
|
|
970
|
-
# (B, L)
|
|
971
|
-
mask = mask.unsqueeze(-1)
|
|
972
|
-
elif (
|
|
973
|
-
mask.dim() == 3
|
|
974
|
-
and mask.shape[1] == 1
|
|
975
|
-
and mask.shape[2] == sequence_length
|
|
976
|
-
):
|
|
977
|
-
# (B, 1, L) -> (B, L, 1)
|
|
978
|
-
mask = mask.transpose(1, 2)
|
|
979
|
-
elif (
|
|
980
|
-
mask.dim() == 3
|
|
981
|
-
and mask.shape[1] == sequence_length
|
|
982
|
-
and mask.shape[2] == 1
|
|
983
|
-
):
|
|
984
|
-
pass
|
|
915
|
+
if mask is None:
|
|
916
|
+
if keys_length is None:
|
|
917
|
+
mask = torch.ones(
|
|
918
|
+
(batch_size, sequence_length), device=keys.device, dtype=keys.dtype
|
|
919
|
+
)
|
|
985
920
|
else:
|
|
921
|
+
device = keys.device
|
|
922
|
+
seq_range = torch.arange(sequence_length, device=device).unsqueeze(0)
|
|
923
|
+
mask = (seq_range < keys_length.unsqueeze(1)).to(keys.dtype)
|
|
924
|
+
else:
|
|
925
|
+
mask = mask.to(keys.dtype).reshape(batch_size, -1)
|
|
926
|
+
if mask.shape[1] != sequence_length:
|
|
986
927
|
raise ValueError(
|
|
987
928
|
f"[AttentionPoolingLayer Error]: Unsupported mask shape: {mask.shape}"
|
|
988
929
|
)
|
|
989
|
-
|
|
930
|
+
mask = mask.unsqueeze(-1)
|
|
990
931
|
# Expand query to (B, L, D)
|
|
991
932
|
query_expanded = query.unsqueeze(1).expand(-1, sequence_length, -1)
|
|
992
933
|
# [query, key, query-key, query*key] -> (B, L, 4D)
|
|
@@ -1026,34 +967,3 @@ class RMSNorm(torch.nn.Module):
|
|
|
1026
967
|
variance = torch.mean(x**2, dim=-1, keepdim=True)
|
|
1027
968
|
x_normalized = x * torch.rsqrt(variance + self.eps)
|
|
1028
969
|
return self.weight * x_normalized
|
|
1029
|
-
|
|
1030
|
-
|
|
1031
|
-
class DomainBatchNorm(nn.Module):
|
|
1032
|
-
"""Domain-specific BatchNorm (applied per-domain with a shared interface)."""
|
|
1033
|
-
|
|
1034
|
-
def __init__(self, num_features: int, num_domains: int):
|
|
1035
|
-
super().__init__()
|
|
1036
|
-
if num_domains < 1:
|
|
1037
|
-
raise ValueError("num_domains must be >= 1")
|
|
1038
|
-
self.bns = nn.ModuleList(
|
|
1039
|
-
[nn.BatchNorm1d(num_features) for _ in range(num_domains)]
|
|
1040
|
-
)
|
|
1041
|
-
|
|
1042
|
-
def forward(self, x: torch.Tensor, domain_mask: torch.Tensor) -> torch.Tensor:
|
|
1043
|
-
if x.dim() != 2:
|
|
1044
|
-
raise ValueError("DomainBatchNorm expects 2D inputs [B, D].")
|
|
1045
|
-
output = x.clone()
|
|
1046
|
-
if domain_mask.dim() == 1:
|
|
1047
|
-
domain_ids = domain_mask.long()
|
|
1048
|
-
for idx, bn in enumerate(self.bns):
|
|
1049
|
-
mask = domain_ids == idx
|
|
1050
|
-
if mask.any():
|
|
1051
|
-
output[mask] = bn(x[mask])
|
|
1052
|
-
return output
|
|
1053
|
-
if domain_mask.dim() != 2:
|
|
1054
|
-
raise ValueError("domain_mask must be 1D indices or 2D one-hot mask.")
|
|
1055
|
-
for idx, bn in enumerate(self.bns):
|
|
1056
|
-
mask = domain_mask[:, idx] > 0
|
|
1057
|
-
if mask.any():
|
|
1058
|
-
output[mask] = bn(x[mask])
|
|
1059
|
-
return output
|
nextrec/basic/loggers.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
NextRec Basic Loggers
|
|
3
3
|
|
|
4
4
|
Date: create on 27/10/2025
|
|
5
|
-
Checkpoint: edit on
|
|
5
|
+
Checkpoint: edit on 22/01/2026
|
|
6
6
|
Author: Yang Zhou, zyaztec@gmail.com
|
|
7
7
|
"""
|
|
8
8
|
|
|
@@ -99,7 +99,8 @@ def format_kv(label: str, value: Any, width: int = 34, indent: int = 0) -> str:
|
|
|
99
99
|
|
|
100
100
|
|
|
101
101
|
def setup_logger(session_id: str | os.PathLike | None = None):
|
|
102
|
-
"""
|
|
102
|
+
"""
|
|
103
|
+
Set up a logger that logs to both console and a file with ANSI formatting.
|
|
103
104
|
Only console output has colors; file output is stripped of ANSI codes.
|
|
104
105
|
|
|
105
106
|
Logs are stored under ``log/<experiment_id>/logs`` by default. A stable
|
nextrec/basic/metrics.py
CHANGED
|
@@ -23,7 +23,6 @@ from sklearn.metrics import (
|
|
|
23
23
|
)
|
|
24
24
|
from nextrec.utils.types import TaskTypeName, MetricsName
|
|
25
25
|
|
|
26
|
-
|
|
27
26
|
TASK_DEFAULT_METRICS = {
|
|
28
27
|
"binary": ["auc", "gauc", "ks", "logloss", "accuracy", "precision", "recall", "f1"],
|
|
29
28
|
"regression": ["mse", "mae", "rmse", "r2", "mape"],
|
|
@@ -334,6 +333,60 @@ def compute_map_at_k(
|
|
|
334
333
|
return float(np.mean(aps)) if aps else 0.0
|
|
335
334
|
|
|
336
335
|
|
|
336
|
+
def compute_topk_counts(
|
|
337
|
+
y_true: np.ndarray, y_pred: np.ndarray, k_percent: int
|
|
338
|
+
) -> tuple[int, int, int]:
|
|
339
|
+
"""Compute Top-K% sample size, hits, and positives for binary labels."""
|
|
340
|
+
y_true = (y_true > 0).astype(int)
|
|
341
|
+
n = y_true.size
|
|
342
|
+
if n == 0:
|
|
343
|
+
return 0, 0, 0
|
|
344
|
+
if k_percent <= 0:
|
|
345
|
+
return 0, 0, int(y_true.sum())
|
|
346
|
+
if k_percent >= 100:
|
|
347
|
+
k_count = n
|
|
348
|
+
else:
|
|
349
|
+
k_count = int(np.ceil(n * (k_percent / 100.0)))
|
|
350
|
+
k_count = max(k_count, 1)
|
|
351
|
+
order = np.argsort(y_pred)[::-1]
|
|
352
|
+
topk = order[:k_count]
|
|
353
|
+
hits = int(y_true[topk].sum())
|
|
354
|
+
total_pos = int(y_true.sum())
|
|
355
|
+
return k_count, hits, total_pos
|
|
356
|
+
|
|
357
|
+
|
|
358
|
+
def compute_topk_precision(
|
|
359
|
+
y_true: np.ndarray, y_pred: np.ndarray, k_percent: int
|
|
360
|
+
) -> float:
|
|
361
|
+
"""Compute Top-K% Precision."""
|
|
362
|
+
k_count, hits, _ = compute_topk_counts(y_true, y_pred, k_percent)
|
|
363
|
+
if k_count == 0:
|
|
364
|
+
return 0.0
|
|
365
|
+
return float(hits / k_count)
|
|
366
|
+
|
|
367
|
+
|
|
368
|
+
def compute_topk_recall(
|
|
369
|
+
y_true: np.ndarray, y_pred: np.ndarray, k_percent: int
|
|
370
|
+
) -> float:
|
|
371
|
+
"""Compute Top-K% Recall."""
|
|
372
|
+
_, hits, total_pos = compute_topk_counts(y_true, y_pred, k_percent)
|
|
373
|
+
if total_pos == 0:
|
|
374
|
+
return 0.0
|
|
375
|
+
return float(hits / total_pos)
|
|
376
|
+
|
|
377
|
+
|
|
378
|
+
def compute_lift_at_k(y_true: np.ndarray, y_pred: np.ndarray, k_percent: int) -> float:
|
|
379
|
+
"""Compute Lift@K from Top-K% precision and overall rate."""
|
|
380
|
+
k_count, hits, total_pos = compute_topk_counts(y_true, y_pred, k_percent)
|
|
381
|
+
if k_count == 0:
|
|
382
|
+
return 0.0
|
|
383
|
+
base_rate = total_pos / float(y_true.size)
|
|
384
|
+
if base_rate == 0.0:
|
|
385
|
+
return 0.0
|
|
386
|
+
precision = hits / float(k_count)
|
|
387
|
+
return float(precision / base_rate)
|
|
388
|
+
|
|
389
|
+
|
|
337
390
|
def compute_cosine_separation(y_true: np.ndarray, y_pred: np.ndarray) -> float:
|
|
338
391
|
"""Compute Cosine Separation."""
|
|
339
392
|
y_true = (y_true > 0).astype(int)
|
|
@@ -399,11 +452,11 @@ def configure_metrics(
|
|
|
399
452
|
if primary_task not in TASK_DEFAULT_METRICS:
|
|
400
453
|
raise ValueError(f"Unsupported task type: {primary_task}")
|
|
401
454
|
metrics_list = TASK_DEFAULT_METRICS[primary_task]
|
|
402
|
-
best_metrics_mode =
|
|
455
|
+
best_metrics_mode = get_best_metric_mode(metrics_list[0], primary_task)
|
|
403
456
|
return metrics_list, task_specific_metrics, best_metrics_mode
|
|
404
457
|
|
|
405
458
|
|
|
406
|
-
def
|
|
459
|
+
def get_best_metric_mode(first_metric: MetricsName, primary_task: TaskTypeName) -> str:
|
|
407
460
|
"""Determine if metric should be maximized or minimized."""
|
|
408
461
|
# Metrics that should be maximized
|
|
409
462
|
if first_metric in {
|
|
@@ -429,6 +482,9 @@ def getbest_metric_mode(first_metric: MetricsName, primary_task: TaskTypeName) -
|
|
|
429
482
|
or first_metric.startswith("mrr@")
|
|
430
483
|
or first_metric.startswith("ndcg@")
|
|
431
484
|
or first_metric.startswith("map@")
|
|
485
|
+
or first_metric.startswith("topk_recall@")
|
|
486
|
+
or first_metric.startswith("topk_precision@")
|
|
487
|
+
or first_metric.startswith("lift@")
|
|
432
488
|
):
|
|
433
489
|
return "max"
|
|
434
490
|
# Cosine separation should be maximized
|
|
@@ -457,6 +513,15 @@ def compute_single_metric(
|
|
|
457
513
|
|
|
458
514
|
y_p_binary = (y_pred > 0.5).astype(int)
|
|
459
515
|
try:
|
|
516
|
+
if metric.startswith("topk_recall@"):
|
|
517
|
+
k_percent = int(metric.split("@")[1])
|
|
518
|
+
return compute_topk_recall(y_true, y_pred, k_percent)
|
|
519
|
+
if metric.startswith("topk_precision@"):
|
|
520
|
+
k_percent = int(metric.split("@")[1])
|
|
521
|
+
return compute_topk_precision(y_true, y_pred, k_percent)
|
|
522
|
+
if metric.startswith("lift@"):
|
|
523
|
+
k_percent = int(metric.split("@")[1])
|
|
524
|
+
return compute_lift_at_k(y_true, y_pred, k_percent)
|
|
460
525
|
if metric.startswith("recall@"):
|
|
461
526
|
k = int(metric.split("@")[1])
|
|
462
527
|
return compute_recall_at_k(y_true, y_pred, user_ids, k) # type: ignore
|
|
@@ -650,7 +715,23 @@ def evaluate_metrics(
|
|
|
650
715
|
allowed_metrics = metric_allowlist.get(task_type)
|
|
651
716
|
for metric in metrics:
|
|
652
717
|
if allowed_metrics is not None and metric not in allowed_metrics:
|
|
653
|
-
|
|
718
|
+
if metric.startswith(
|
|
719
|
+
(
|
|
720
|
+
"recall@",
|
|
721
|
+
"precision@",
|
|
722
|
+
"hitrate@",
|
|
723
|
+
"hr@",
|
|
724
|
+
"mrr@",
|
|
725
|
+
"ndcg@",
|
|
726
|
+
"map@",
|
|
727
|
+
"topk_recall@",
|
|
728
|
+
"topk_precision@",
|
|
729
|
+
"lift@",
|
|
730
|
+
)
|
|
731
|
+
):
|
|
732
|
+
pass
|
|
733
|
+
else:
|
|
734
|
+
continue
|
|
654
735
|
y_true_task = y_true[:, task_idx]
|
|
655
736
|
y_pred_task = y_pred[:, task_idx]
|
|
656
737
|
task_user_ids = user_ids
|