Av softmax

[VincentAuriau]

fixes limit case of unavailable large utility

[gemini-code-assist]

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.

[VincentAuriau]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces the tied_embeddings option to the AleaCarta basket model, allowing items to have distinct embeddings depending on whether they are in the basket. It also updates the softmax utility calculation and adds corresponding unit and integration tests. The review identified a critical bug in softmax_with_availabilities where multiplying logits by availability can corrupt the maximum utility calculation when all available items have negative utilities. Additionally, setting a global TensorFlow seed inside the model's instantiate method was flagged as a bad practice due to global side-effects.

[gemini-code-assist]

Bug Description

Multiplying items_logit_by_choice by available_items_by_choice (which is 0.0 for unavailable items) sets the logits of unavailable items to 0.0.

If all available items have negative utilities (e.g., [-10.0, -20.0]), the maximum utility (normalizer) will be incorrectly computed as 0.0 (from the unavailable items) instead of -10.0. This defeats the overflow/underflow prevention and can cause the probabilities of available items to underflow to 0.0 or result in NaN values.

Solution

Instead of multiplying by the availability mask before taking the maximum, use tf.where to mask the unavailable items with a large negative value (e.g., -1e9). This ensures that:

1. Unavailable items do not affect the maximum utility calculation.
2. Exponentiating the masked logits safely underflows to 0.0 for unavailable items, avoiding any NaN or overflow issues.

Suggested change

	numerator = tf.multiply(items_logit_by_choice, available_items_by_choice)
	normalizer = tf.reduce_max(numerator, axis=axis, keepdims=True)
	numerator = tf.exp(numerator - normalizer)
	large_negative = tf.constant(-1e9, dtype=items_logit_by_choice.dtype)
	masked_logits = tf.where(available_items_by_choice > 0.5, items_logit_by_choice, large_negative)
	normalizer = tf.reduce_max(masked_logits, axis=axis, keepdims=True)
	numerator = tf.exp(masked_logits - normalizer)

[gemini-code-assist]

Issue

Calling tf.random.set_seed(42) inside the instantiate method of a model class introduces a global side-effect. It resets the global random seed of the entire TensorFlow runtime, which can silently break the reproducibility of other models, training loops, or random processes in the user's application.

Recommendation

Avoid setting global seeds inside library code. Instead, allow the user to control the global seed externally, or accept an optional seed parameter in the model's constructor and use it locally (e.g., passing it to the initializers or using a tf.random.Generator). If reproducibility is desired by default, you can pass a seed directly to the initializers or document that the user should set the global seed.

[github-actions]

Coverage Report for Python 3.10

File	Stmts	Miss	Cover	Missing
choice_learn
__init__.py	2	0	100%
tf_ops.py	64	1	98%	286
choice_learn/basket_models
__init__.py	4	0	100%
alea_carta.py	155	3	98%	111, 136, 326
base_basket_model.py	248	29	88%	115–116, 127, 145, 189, 259, 381, 489, 589–591, 680, 785, 793, 803, 851, 854–864, 925–928, 967–968
basic_attention_model.py	89	4	96%	424, 427, 433, 440
self_attention_model.py	133	9	93%	71, 73, 75, 450–454, 651
shopper.py	184	9	95%	130, 159, 325, 345, 360, 363, 377, 489, 618
choice_learn/basket_models/data
__init__.py	2	0	100%
basket_dataset.py	195	62	68%	74–77, 223–231, 299–301, 411, 544–580, 608–648, 671, 678–683, 753–764, 812
preprocessing.py	94	78	17%	43–45, 128–364
choice_learn/basket_models/datasets
__init__.py	3	0	100%
badminton.py	81	6	93%	62, 194–199, 247
bakery.py	38	3	92%	47, 51, 61
choice_learn/basket_models/utils
__init__.py	0	0	100%
permutation.py	22	1	95%	37
choice_learn/data
__init__.py	3	0	100%
choice_dataset.py	649	33	95%	198, 250, 283, 421, 463–464, 589, 724, 738, 840, 842, 937, 957–961, 1140, 1159–1161, 1179–1181, 1209, 1214, 1223, 1240, 1281, 1293, 1307, 1346, 1361, 1366, 1395, 1408, 1443–1444
indexer.py	241	23	90%	20, 31, 45, 60–67, 202–204, 219–230, 265, 291, 582
storage.py	161	6	96%	22, 33, 51, 56, 61, 71
store.py	72	72	0%	3–275
choice_learn/datasets
__init__.py	4	0	100%
base.py	400	5	99%	42–43, 153–154, 714
expedia.py	102	83	19%	37–301
tafeng.py	49	0	100%
choice_learn/datasets/data
__init__.py	0	0	100%
choice_learn/models
__init__.py	14	2	86%	15–16
base_model.py	335	36	89%	145, 187, 289, 297, 303, 312, 352, 356–357, 362, 391, 395–396, 413, 426, 434, 475–476, 485–486, 587, 589, 605, 609, 611, 734–735, 908, 935, 939–953
baseline_models.py	49	0	100%
conditional_logit.py	269	26	90%	49, 52, 54, 85, 88, 91–95, 98–102, 136, 206, 212–216, 351, 388, 445, 520–526, 651, 685, 822, 826
halo_mnl.py	124	2	98%	186, 374
latent_class_base_model.py	286	39	86%	55–61, 273–279, 288, 325–330, 497–500, 605, 624, 665–701, 715, 720, 751–752, 774–775, 869–870, 974
latent_class_mnl.py	62	6	90%	257–261, 296
learning_mnl.py	67	3	96%	157, 182, 188
nested_logit.py	291	12	96%	55, 77, 160, 269, 351, 484, 530, 600, 679, 848, 900, 904
reslogit.py	132	6	95%	285, 360, 369, 374, 382, 432
rumnet.py	236	3	99%	748–751, 982
simple_mnl.py	139	6	96%	167, 275, 347, 355, 357, 359
tastenet.py	94	3	97%	142, 180, 188
choice_learn/toolbox
__init__.py	0	0	100%
assortment_optimizer.py	27	6	78%	28–30, 93–95, 160–162
gurobi_opt.py	238	238	0%	3–675
or_tools_opt.py	230	11	95%	103, 107, 296–305, 315, 319, 607, 611
choice_learn/utils
metrics.py	116	7	94%	73, 151–153, 219–221, 287–289
TOTAL	5704	833	85%

Tests	Skipped	Failures	Errors	Time
228	0 💤	0 ❌	0 🔥	5m 22s ⏱️

[github-actions]

Coverage Report for Python 3.9

File	Stmts	Miss	Cover	Missing
choice_learn
__init__.py	2	0	100%
tf_ops.py	64	1	98%	286
choice_learn/basket_models
__init__.py	4	0	100%
alea_carta.py	155	3	98%	111, 136, 326
base_basket_model.py	248	29	88%	115–116, 127, 145, 189, 259, 381, 489, 589–591, 680, 785, 793, 803, 851, 854–864, 925–928, 967–968
basic_attention_model.py	89	4	96%	424, 427, 433, 440
self_attention_model.py	133	9	93%	71, 73, 75, 450–454, 651
shopper.py	184	9	95%	130, 159, 325, 345, 360, 363, 377, 489, 618
choice_learn/basket_models/data
__init__.py	2	0	100%
basket_dataset.py	195	62	68%	74–77, 223–231, 299–301, 411, 544–580, 608–648, 671, 678–683, 753–764, 812
preprocessing.py	94	78	17%	43–45, 128–364
choice_learn/basket_models/datasets
__init__.py	3	0	100%
badminton.py	81	6	93%	62, 194–199, 247
bakery.py	38	3	92%	47, 51, 61
choice_learn/basket_models/utils
__init__.py	0	0	100%
permutation.py	22	1	95%	37
choice_learn/data
__init__.py	3	0	100%
choice_dataset.py	649	33	95%	198, 250, 283, 421, 463–464, 589, 724, 738, 840, 842, 937, 957–961, 1140, 1159–1161, 1179–1181, 1209, 1214, 1223, 1240, 1281, 1293, 1307, 1346, 1361, 1366, 1395, 1408, 1443–1444
indexer.py	241	23	90%	20, 31, 45, 60–67, 202–204, 219–230, 265, 291, 582
storage.py	161	6	96%	22, 33, 51, 56, 61, 71
store.py	72	72	0%	3–275
choice_learn/datasets
__init__.py	4	0	100%
base.py	400	5	99%	42–43, 153–154, 714
expedia.py	102	83	19%	37–301
tafeng.py	49	0	100%
choice_learn/datasets/data
__init__.py	0	0	100%
choice_learn/models
__init__.py	14	2	86%	15–16
base_model.py	335	35	90%	145, 187, 289, 297, 303, 312, 352, 356–357, 362, 391, 395–396, 413, 426, 434, 475–476, 485–486, 587, 589, 605, 609, 734–735, 908, 935, 939–953
baseline_models.py	49	0	100%
conditional_logit.py	269	26	90%	49, 52, 54, 85, 88, 91–95, 98–102, 136, 206, 212–216, 351, 388, 445, 520–526, 651, 685, 822, 826
halo_mnl.py	124	2	98%	186, 374
latent_class_base_model.py	286	39	86%	55–61, 273–279, 288, 325–330, 497–500, 605, 624, 665–701, 715, 720, 751–752, 774–775, 869–870, 974
latent_class_mnl.py	62	6	90%	257–261, 296
learning_mnl.py	67	3	96%	157, 182, 188
nested_logit.py	291	12	96%	55, 77, 160, 269, 351, 484, 530, 600, 679, 848, 900, 904
reslogit.py	132	7	95%	122, 285, 360, 369, 374, 382, 432
rumnet.py	236	3	99%	748–751, 982
simple_mnl.py	139	6	96%	167, 275, 347, 355, 357, 359
tastenet.py	94	3	97%	142, 180, 188
choice_learn/toolbox
__init__.py	0	0	100%
assortment_optimizer.py	27	6	78%	28–30, 93–95, 160–162
gurobi_opt.py	236	236	0%	3–675
or_tools_opt.py	230	11	95%	103, 107, 296–305, 315, 319, 607, 611
choice_learn/utils
metrics.py	116	73	37%	71–105, 115, 149–173, 183, 217–240, 250, 285–309, 319
TOTAL	5702	897	84%

Tests	Skipped	Failures	Errors	Time
228	0 💤	0 ❌	0 🔥	5m 32s ⏱️

[github-actions]

Coverage Report for Python 3.12

File	Stmts	Miss	Cover	Missing
choice_learn
__init__.py	2	0	100%
tf_ops.py	64	1	98%	286
choice_learn/basket_models
__init__.py	4	0	100%
alea_carta.py	155	3	98%	111, 136, 326
base_basket_model.py	248	29	88%	115–116, 127, 145, 189, 259, 381, 489, 589–591, 680, 785, 793, 803, 851, 854–864, 925–928, 967–968
basic_attention_model.py	89	4	96%	424, 427, 433, 440
self_attention_model.py	133	9	93%	71, 73, 75, 450–454, 651
shopper.py	184	9	95%	130, 159, 325, 345, 360, 363, 377, 489, 618
choice_learn/basket_models/data
__init__.py	2	0	100%
basket_dataset.py	195	62	68%	74–77, 223–231, 299–301, 411, 544–580, 608–648, 671, 678–683, 753–764, 812
preprocessing.py	94	78	17%	43–45, 128–364
choice_learn/basket_models/datasets
__init__.py	3	0	100%
badminton.py	81	6	93%	62, 194–199, 247
bakery.py	38	3	92%	47, 53, 61
choice_learn/basket_models/utils
__init__.py	0	0	100%
permutation.py	22	1	95%	37
choice_learn/data
__init__.py	3	0	100%
choice_dataset.py	649	33	95%	198, 250, 283, 421, 463–464, 589, 724, 738, 840, 842, 937, 957–961, 1140, 1159–1161, 1179–1181, 1209, 1214, 1223, 1240, 1281, 1293, 1307, 1346, 1361, 1366, 1395, 1408, 1443–1444
indexer.py	241	23	90%	20, 31, 45, 60–67, 202–204, 219–230, 265, 291, 582
storage.py	161	6	96%	22, 33, 51, 56, 61, 71
store.py	72	72	0%	3–275
choice_learn/datasets
__init__.py	4	0	100%
base.py	400	5	99%	42–43, 153–154, 714
expedia.py	102	83	19%	37–301
tafeng.py	49	0	100%
choice_learn/datasets/data
__init__.py	0	0	100%
choice_learn/models
__init__.py	14	2	86%	15–16
base_model.py	335	35	90%	145, 187, 289, 297, 303, 312, 352, 356–357, 362, 391, 395–396, 413, 426, 434, 475–476, 485–486, 587, 589, 605, 609, 611, 734–735, 935, 939–953
baseline_models.py	49	0	100%
conditional_logit.py	269	26	90%	49, 52, 54, 85, 88, 91–95, 98–102, 136, 206, 212–216, 351, 388, 445, 520–526, 651, 685, 822, 826
halo_mnl.py	124	18	85%	186, 341, 360, 364–380
latent_class_base_model.py	286	39	86%	55–61, 273–279, 288, 325–330, 497–500, 605, 624, 665–701, 715, 720, 751–752, 774–775, 869–870, 974
latent_class_mnl.py	62	6	90%	257–261, 296
learning_mnl.py	67	3	96%	157, 182, 188
nested_logit.py	291	12	96%	55, 77, 160, 269, 351, 484, 530, 600, 679, 848, 900, 904
reslogit.py	132	6	95%	285, 360, 369, 374, 382, 432
rumnet.py	236	3	99%	748–751, 982
simple_mnl.py	139	6	96%	167, 275, 347, 355, 357, 359
tastenet.py	94	3	97%	142, 180, 188
choice_learn/toolbox
__init__.py	0	0	100%
assortment_optimizer.py	27	6	78%	28–30, 93–95, 160–162
gurobi_opt.py	238	238	0%	3–675
or_tools_opt.py	230	11	95%	103, 107, 296–305, 315, 319, 607, 611
choice_learn/utils
metrics.py	116	7	94%	73, 151–153, 219–221, 287–289
TOTAL	5704	848	85%

Tests	Skipped	Failures	Errors	Time
228	0 💤	1 ❌	0 🔥	7m 34s ⏱️

[github-actions]

Coverage Report for Python 3.11

File	Stmts	Miss	Cover	Missing
choice_learn
__init__.py	2	0	100%
tf_ops.py	64	1	98%	286
choice_learn/basket_models
__init__.py	4	0	100%
alea_carta.py	155	3	98%	111, 136, 326
base_basket_model.py	248	29	88%	115–116, 127, 145, 189, 259, 381, 489, 589–591, 680, 785, 793, 803, 851, 854–864, 925–928, 967–968
basic_attention_model.py	89	4	96%	424, 427, 433, 440
self_attention_model.py	133	9	93%	71, 73, 75, 450–454, 651
shopper.py	184	9	95%	130, 159, 325, 345, 360, 363, 377, 489, 618
choice_learn/basket_models/data
__init__.py	2	0	100%
basket_dataset.py	195	62	68%	74–77, 223–231, 299–301, 411, 544–580, 608–648, 671, 678–683, 753–764, 812
preprocessing.py	94	78	17%	43–45, 128–364
choice_learn/basket_models/datasets
__init__.py	3	0	100%
badminton.py	81	6	93%	62, 194–199, 247
bakery.py	38	3	92%	47, 51, 61
choice_learn/basket_models/utils
__init__.py	0	0	100%
permutation.py	22	1	95%	37
choice_learn/data
__init__.py	3	0	100%
choice_dataset.py	649	33	95%	198, 250, 283, 421, 463–464, 589, 724, 738, 840, 842, 937, 957–961, 1140, 1159–1161, 1179–1181, 1209, 1214, 1223, 1240, 1281, 1293, 1307, 1346, 1361, 1366, 1395, 1408, 1443–1444
indexer.py	241	23	90%	20, 31, 45, 60–67, 202–204, 219–230, 265, 291, 582
storage.py	161	6	96%	22, 33, 51, 56, 61, 71
store.py	72	72	0%	3–275
choice_learn/datasets
__init__.py	4	0	100%
base.py	400	5	99%	42–43, 153–154, 714
expedia.py	102	83	19%	37–301
tafeng.py	49	0	100%
choice_learn/datasets/data
__init__.py	0	0	100%
choice_learn/models
__init__.py	14	2	86%	15–16
base_model.py	335	35	90%	145, 187, 289, 297, 303, 312, 352, 356–357, 362, 391, 395–396, 413, 426, 434, 475–476, 485–486, 587, 589, 605, 609, 611, 734–735, 935, 939–953
baseline_models.py	49	0	100%
conditional_logit.py	269	26	90%	49, 52, 54, 85, 88, 91–95, 98–102, 136, 206, 212–216, 351, 388, 445, 520–526, 651, 685, 822, 826
halo_mnl.py	124	18	85%	186, 341, 360, 364–380
latent_class_base_model.py	286	39	86%	55–61, 273–279, 288, 325–330, 497–500, 605, 624, 665–701, 715, 720, 751–752, 774–775, 869–870, 974
latent_class_mnl.py	62	6	90%	257–261, 296
learning_mnl.py	67	3	96%	157, 182, 188
nested_logit.py	291	12	96%	55, 77, 160, 269, 351, 484, 530, 600, 679, 848, 900, 904
reslogit.py	132	6	95%	285, 360, 369, 374, 382, 432
rumnet.py	236	3	99%	748–751, 982
simple_mnl.py	139	6	96%	167, 275, 347, 355, 357, 359
tastenet.py	94	3	97%	142, 180, 188
choice_learn/toolbox
__init__.py	0	0	100%
assortment_optimizer.py	27	6	78%	28–30, 93–95, 160–162
gurobi_opt.py	238	238	0%	3–675
or_tools_opt.py	230	11	95%	103, 107, 296–305, 315, 319, 607, 611
choice_learn/utils
metrics.py	116	7	94%	73, 151–153, 219–221, 287–289
TOTAL	5704	848	85%

Tests	Skipped	Failures	Errors	Time
228	0 💤	1 ❌	0 🔥	5m 38s ⏱️