Skip to content

SigLIP (Sigmoid Loss for Language Image Pre-Training)

SigLIP (Sigmoid Loss for Language Image Pre-Training) is a vision-language model that builds upon the principles of CLIP but introduces a key architectural change: it uses a sigmoid loss function instead of the softmax-based contrastive loss. Additionally, there are some slight implementation differences (no attention_mask for the text encoder, padding the text inputs, multihead attention pooling for the vision encoder rather than a linear projection layer).

This modification simplifies the training objective by treating the problem as a binary classification for each image-text pair (i.e., are they a positive or negative match?). This approach avoids the need for a global normalization over all pairs in a batch, which makes it more scalable and robust to noisy, web-scale data.

Key features of SigLIP: 1. Vision Encoder: A Vision Transformer (ViT) with a Multi-Head Attention Pooling (MAP) head. 2. Text Encoder: A standard Transformer model. 3. Sigmoid Loss: Enables training on larger batches and noisier datasets without requiring careful data curation or complex negative sampling strategies.

SigLIP was introduced in the paper "Sigmoid Loss for Language Image Pre-Training" and has demonstrated improved performance and training efficiency.

jimm.models.siglip.SigLIPVisionModel

Bases: Module

Source code in src/jimm/models/siglip.py 
 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
class SigLIPVisionModel(nnx.Module):
    def __init__(
        self,
        image_resolution: int,
        vision_layers: int,
        vision_width: int,
        vision_patch_size: int,
        use_gradient_checkpointing: bool = False,
        rngs: rnglib.Rngs = nnx.Rngs(0),
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        mesh: Mesh | None = None,
    ):
        self.vision_layers = vision_layers
        self.vision_width = vision_width
        self.vision_patch_size = vision_patch_size
        self.dtype = dtype

        vision_heads = vision_width // 64

        self.encoder = VisionTransformerBase(
            img_size=image_resolution,
            patch_size=vision_patch_size,
            in_channels=3,
            hidden_size=vision_width,
            num_layers=vision_layers,
            num_heads=vision_heads,
            mlp_dim=vision_width * 4,
            use_pre_norm=False,
            use_patch_bias=True,
            use_quick_gelu=False,
            use_gradient_checkpointing=use_gradient_checkpointing,
            pooling_type="MAP",
            layernorm_epsilon=1e-6,
            dtype=dtype,
            param_dtype=param_dtype,
            mesh=mesh,
            rngs=rngs,
        )

    def __call__(self, image: Float[Array, "batch height width channels"]) -> Float[Array, "batch vision_width"]:
        """
        Encode images into embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.

        Returns:
            Float[Array, "batch transformer_width"]: Image embeddings.
        """
        return self.encoder(image)

    @classmethod
    def from_pretrained(
        cls,
        model_name_or_path: str,
        use_pytorch: bool = False,
        mesh: Mesh | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        use_gradient_checkpointing: bool = False,
        rngs: rnglib.Rngs = nnx.Rngs(0),
    ) -> "SigLIPVisionModel":
        params_fstate, config_dict = load_params_and_config(model_name_or_path, use_pytorch)

        vision_patch_size = params_fstate["vision_model.embeddings.patch_embedding.weight"].shape[3]
        vision_width = params_fstate["vision_model.embeddings.patch_embedding.bias"].shape[0]
        vision_num_layers = 0
        for k in params_fstate:
            if k.startswith("vision_model.encoder.layers.") and k.endswith(".mlp.fc2.bias"):
                vision_num_layers = max(vision_num_layers, int(k.split(".")[3]) + 1)

        vision_model = cls(
            image_resolution=config_dict["vision_config"]["image_size"],
            vision_layers=vision_num_layers,
            vision_width=vision_width,
            vision_patch_size=vision_patch_size,
            use_gradient_checkpointing=use_gradient_checkpointing,
            mesh=mesh,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
        )

        flax_model_params_fstate = dict(nnx.to_flat_state(nnx.state(vision_model, nnx.Param)))

        vision_mapping_list = [
            (("encoder", "patch_embeddings", "kernel"), ("vision_model", "embeddings", "patch_embedding", "weight")),
            (("encoder", "patch_embeddings", "bias"), ("vision_model", "embeddings", "patch_embedding", "bias")),
            (("encoder", "position_embeddings"), ("vision_model", "embeddings", "position_embedding", "weight")),
            (("encoder", "ln_post", "scale"), ("vision_model", "post_layernorm", "weight")),
            (("encoder", "ln_post", "bias"), ("vision_model", "post_layernorm", "bias")),
            (("encoder", "MAPHead", "probe"), ("vision_model", "head", "probe")),
            (("encoder", "MAPHead", "layernorm", "scale"), ("vision_model", "head", "layernorm", "weight")),
            (("encoder", "MAPHead", "layernorm", "bias"), ("vision_model", "head", "layernorm", "bias")),
            (("encoder", "MAPHead", "mlp", "layers", 0, "kernel"), ("vision_model", "head", "mlp", "fc1", "weight")),
            (("encoder", "MAPHead", "mlp", "layers", 0, "bias"), ("vision_model", "head", "mlp", "fc1", "bias")),
            (("encoder", "MAPHead", "mlp", "layers", 2, "kernel"), ("vision_model", "head", "mlp", "fc2", "weight")),
            (("encoder", "MAPHead", "mlp", "layers", 2, "bias"), ("vision_model", "head", "mlp", "fc2", "bias")),
            (("encoder", "MAPHead", "attn", "query", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("encoder", "MAPHead", "attn", "query", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("encoder", "MAPHead", "attn", "key", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("encoder", "MAPHead", "attn", "key", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("encoder", "MAPHead", "attn", "value", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("encoder", "MAPHead", "attn", "value", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("encoder", "MAPHead", "attn", "out", "kernel"), ("vision_model", "head", "attention", "out_proj", "weight")),
            (("encoder", "MAPHead", "attn", "out", "bias"), ("vision_model", "head", "attention", "out_proj", "bias")),
        ]

        vision_heads = vision_width // 64

        for i in range(vision_num_layers):
            flax_base = ("encoder", "encoder", "layers", i)
            hf_base = ("vision_model", "encoder", "layers", str(i))
            vision_mapping_list.extend(
                [
                    (flax_base + ("attn", "query", "kernel"), hf_base + ("self_attn", "q_proj", "weight")),
                    (flax_base + ("attn", "query", "bias"), hf_base + ("self_attn", "q_proj", "bias")),
                    (flax_base + ("attn", "key", "kernel"), hf_base + ("self_attn", "k_proj", "weight")),
                    (flax_base + ("attn", "key", "bias"), hf_base + ("self_attn", "k_proj", "bias")),
                    (flax_base + ("attn", "value", "kernel"), hf_base + ("self_attn", "v_proj", "weight")),
                    (flax_base + ("attn", "value", "bias"), hf_base + ("self_attn", "v_proj", "bias")),
                    (flax_base + ("attn", "out", "kernel"), hf_base + ("self_attn", "out_proj", "weight")),
                    (flax_base + ("attn", "out", "bias"), hf_base + ("self_attn", "out_proj", "bias")),
                    (flax_base + ("norm1", "scale"), hf_base + ("layer_norm1", "weight")),
                    (flax_base + ("norm1", "bias"), hf_base + ("layer_norm1", "bias")),
                    (flax_base + ("norm2", "scale"), hf_base + ("layer_norm2", "weight")),
                    (flax_base + ("norm2", "bias"), hf_base + ("layer_norm2", "bias")),
                    (flax_base + ("mlp", "layers", 0, "kernel"), hf_base + ("mlp", "fc1", "weight")),
                    (flax_base + ("mlp", "layers", 0, "bias"), hf_base + ("mlp", "fc1", "bias")),
                    (flax_base + ("mlp", "layers", 3, "kernel"), hf_base + ("mlp", "fc2", "weight")),
                    (flax_base + ("mlp", "layers", 3, "bias"), hf_base + ("mlp", "fc2", "bias")),
                ]
            )

        params_name_mapping = dict(vision_mapping_list)
        nonvisited = set(flax_model_params_fstate.keys())
        used_hf_keys: Set[str] = set()

        for flax_dst_key_tuple, hf_src_key_tuple in params_name_mapping.items():
            hf_src_key_as_string = ".".join(hf_src_key_tuple)

            nonvisited.discard(flax_dst_key_tuple)
            used_hf_keys.add(hf_src_key_as_string)
            src_value = params_fstate[hf_src_key_as_string]
            dst_value_obj = flax_model_params_fstate[flax_dst_key_tuple]

            if flax_dst_key_tuple == ("encoder", "patch_embeddings", "kernel"):
                src_value = jnp.transpose(src_value, (2, 3, 1, 0))
            elif flax_dst_key_tuple == ("encoder", "position_embeddings"):
                src_value = src_value.reshape(1, src_value.shape[0], src_value.shape[1])
            elif hf_src_key_tuple[-1] == "weight" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
                src_value = jnp.transpose(src_value, (1, 0))
                num_heads = vision_heads
                head_dim = vision_width // num_heads
                src_value = src_value.reshape((vision_width, num_heads, head_dim))
            elif hf_src_key_tuple[-1] == "bias" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
                num_heads = vision_heads
                head_dim = vision_width // num_heads
                src_value = src_value.reshape((num_heads, head_dim))
            elif hf_src_key_tuple[-2:] == ("out_proj", "weight"):
                src_value = jnp.transpose(src_value, (1, 0))
                num_heads = vision_heads
                head_dim = vision_width // num_heads
                src_value = src_value.reshape((num_heads, head_dim, vision_width))
            elif hf_src_key_tuple[-1] == "in_proj_weight":
                num_heads = vision_heads
                head_dim = vision_width // num_heads
                q_w, k_w, v_w = jnp.split(src_value, 3, axis=0)
                w_map = {"query": q_w, "key": k_w, "value": v_w}
                src_value = jnp.transpose(w_map[flax_dst_key_tuple[-2]], (1, 0)).reshape(vision_width, num_heads, head_dim)
            elif hf_src_key_tuple[-1] == "in_proj_bias":
                num_heads = vision_heads
                head_dim = vision_width // num_heads
                q_b, k_b, v_b = jnp.split(src_value, 3, axis=0)
                b_map = {"query": q_b, "key": k_b, "value": v_b}
                src_value = b_map[flax_dst_key_tuple[-2]].reshape(num_heads, head_dim)
            elif hf_src_key_tuple[-1] == "weight" and src_value.ndim == 2:
                src_value = jnp.transpose(src_value, (1, 0))

            if src_value.shape != dst_value_obj.value.shape:
                raise ValueError(f"Shape mismatch for {flax_dst_key_tuple} vs {hf_src_key_as_string}: {dst_value_obj.value.shape} (expected) != {src_value.shape} (actual)")

            src_value = src_value.astype(param_dtype)
            dst_value_obj.value = src_value

        nnx.update(vision_model, nnx.from_flat_state(flax_model_params_fstate))
        known_buffer_keys = {("encoder", "vision_position_ids")}
        unexpected_nonvisited = nonvisited - known_buffer_keys
        if unexpected_nonvisited:
            print(f"Warning: Some {cls.__name__} parameters were not loaded: {sorted(list(unexpected_nonvisited))}")

        return vision_model

__call__(image)

Encode images into embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required

Returns:

Type Description
Float[Array, 'batch vision_width']

Float[Array, "batch transformer_width"]: Image embeddings.
Source code in src/jimm/models/siglip.py 
58
59
60
61
62
63
64
65
66
67
68
def __call__(self, image: Float[Array, "batch height width channels"]) -> Float[Array, "batch vision_width"]:
    """
    Encode images into embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.

    Returns:
        Float[Array, "batch transformer_width"]: Image embeddings.
    """
    return self.encoder(image)

jimm.models.siglip.SigLIP

Bases: Module

Source code in src/jimm/models/siglip.py 
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
class SigLIP(nnx.Module):
    def __init__(
        self,
        image_resolution: int,
        vision_layers: int,
        vision_width: int,
        vision_patch_size: int,
        context_length: int,
        vocab_size: int,
        transformer_width: int,
        transformer_heads: int,
        transformer_layers: int,
        use_gradient_checkpointing: bool = False,
        rngs: rnglib.Rngs = nnx.Rngs(0),
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        mesh: Mesh | None = None,
    ):
        """
        Initialize the SigLIP model.

        Args:
            image_resolution (int): The resolution of the input images.
            vision_layers (int): The number of layers in the vision transformer.
            vision_width (int): The width of the vision transformer.
            vision_patch_size (int): The patch size of the vision transformer.
            context_length (int): The length of the context.
            vocab_size (int): The size of the vocabulary.
            transformer_width (int): The width of the transformer.
            transformer_heads (int): The number of attention heads in the transformer.
            transformer_layers (int): The number of layers in the transformer.
            use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
            rngs (rnglib.Rngs): The random number generator state. Defaults to nnx.Rngs(0).
            dtype (DTypeLike): The data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike): The data type for parameters. Defaults to jnp.float32.
            mesh (Mesh | None): Optional device mesh for parameter sharding. Defaults to None.
        """
        self.vision_layers = vision_layers
        self.vision_width = vision_width
        self.vision_patch_size = vision_patch_size
        self.context_length = context_length
        self.transformer_width = transformer_width
        self.transformer_heads = transformer_heads
        self.transformer_layers = transformer_layers
        self.dtype = dtype
        self._original_config = None

        self.vision_heads = vision_width // 64
        self.vision_model = SigLIPVisionModel(
            image_resolution=image_resolution,
            vision_layers=vision_layers,
            vision_width=vision_width,
            vision_patch_size=vision_patch_size,
            use_gradient_checkpointing=use_gradient_checkpointing,
            rngs=rngs,
            dtype=dtype,
            param_dtype=param_dtype,
            mesh=mesh,
        )

        self.text_model = Transformer(
            width=transformer_width,
            mlp_dim=transformer_width * 4,
            layers=transformer_layers,
            num_heads=transformer_heads,
            dropout_rate=0.0,
            layernorm_epsilon=1e-6,
            use_quick_gelu=False,
            use_gradient_checkpointing=use_gradient_checkpointing,
            dtype=dtype,
            param_dtype=param_dtype,
            mesh=mesh,
            rngs=rngs,
        )
        self.vocab_size = vocab_size
        self.token_embedding = nnx.Embed(
            num_embeddings=vocab_size,
            features=transformer_width,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            embedding_init=sharded_init(nnx.initializers.xavier_uniform(), P("model", None), mesh),
        )
        self.positional_embedding = nnx.Param(sharded_init(nnx.initializers.truncated_normal(stddev=0.02), P("model", None), mesh)(rngs.params(), (context_length, transformer_width)))
        self.ln_final = nnx.LayerNorm(
            transformer_width,
            epsilon=1e-6,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            scale_init=sharded_init(nnx.initializers.ones_init(), P("model"), mesh),
            bias_init=sharded_init(nnx.initializers.zeros_init(), P("model"), mesh),
        )
        self.text_projection = nnx.Linear(
            transformer_width,
            transformer_width,
            use_bias=True,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
            kernel_init=sharded_init(nnx.initializers.xavier_uniform(), P("model", None), mesh),
        )
        self.logit_scale = nnx.Param(sharded_init(nnx.initializers.ones_init(), P(), mesh)(rngs.params(), ()))
        self.logit_bias = nnx.Param(sharded_init(nnx.initializers.ones_init(), P(), mesh)(rngs.params(), ()))

    def _create_config(self) -> dict[str, Any]:
        return {
            "model_type": "siglip",
            "text_config": {
                "hidden_size": self.transformer_width,
                "num_attention_heads": self.transformer_heads,
                "num_hidden_layers": self.transformer_layers,
                "max_position_embeddings": self.context_length,
                "vocab_size": self.vocab_size,
            },
            "vision_config": {
                "hidden_size": self.vision_width,
                "num_attention_heads": self.vision_width // 64,
                "num_hidden_layers": self.vision_layers,
                "image_size": self.vision_model.img_size,
                "patch_size": self.vision_patch_size,
            },
        }

    def encode_image(self, image: Float[Array, "batch height width channels"]) -> Float[Array, "batch transformer_width"]:
        """
        Encode images into embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.

        Returns:
            Float[Array, "batch transformer_width"]: Image embeddings.
        """
        return self.vision_model(image)

    def encode_text(self, text: Int[Array, "batch context_length"]) -> Float[Array, "batch transformer_width"]:
        """
        Encode text tokens into embeddings.

        Args:
            text (Int[Array, "batch context_length"]): Batch of token sequences.

        Returns:
            Float[Array, "batch transformer_width"]: Text embeddings.
        """
        seq_len = text.shape[1]
        x: Float[Array, "batch context_length transformer_width"] = self.token_embedding(text)
        x: Float[Array, "batch context_length transformer_width"] = x + self.positional_embedding.value[:seq_len]
        x: Float[Array, "batch context_length transformer_width"] = self.text_model(x)
        x: Float[Array, "batch context_length transformer_width"] = self.ln_final(x)

        pooled_output = x[:, -1, :]
        x: Float[Array, "batch transformer_width"] = self.text_projection(pooled_output)
        return x

    def __call__(self, image: Float[Array, "batch height width channels"], text: Int[Array, "batch context_length"]) -> Float[Array, "batch batch"]:
        """
        Calculate similarity between image and text embeddings.

        Args:
            image (Float[Array, "batch height width channels"]): Batch of input images.
            text (Int[Array, "batch context_length"]): Batch of token sequences.

        Returns:
            Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
        """
        image_features: Float[Array, "batch transformer_width"] = self.encode_image(image)
        text_features: Float[Array, "batch transformer_width"] = self.encode_text(text)

        image_features: Float[Array, "batch transformer_width"] = image_features / jnp.linalg.norm(image_features, axis=-1, keepdims=True)
        text_features: Float[Array, "batch transformer_width"] = text_features / jnp.linalg.norm(text_features, axis=-1, keepdims=True)

        logit_scale: Float[Array, ""] = jnp.exp(self.logit_scale.value)
        logits: Float[Array, "batch batch"] = logit_scale * image_features @ text_features.T + self.logit_bias.value
        return logits

    def save_pretrained(self, save_directory: str):
        _SPECIAL_MAPPINGS = {
            "ln_final.weight": "text_model.final_layer_norm.weight",
            "ln_final.bias": "text_model.final_layer_norm.bias",
            "vision_model.encoder.ln_pre.weight": "vision_model.pre_layrnorm.weight",
            "vision_model.encoder.ln_pre.bias": "vision_model.pre_layrnorm.bias",
            "vision_model.encoder.ln_post.weight": "vision_model.post_layernorm.weight",
            "vision_model.encoder.ln_post.bias": "vision_model.post_layernorm.bias",
            "vision_model.encoder.position_embeddings": "vision_model.embeddings.position_embedding.weight",
            "vision_model.encoder.patch_embeddings.weight": "vision_model.embeddings.patch_embedding.weight",
            "vision_model.encoder.patch_embeddings.bias": "vision_model.embeddings.patch_embedding.bias",
            "positional_embedding": "text_model.embeddings.position_embedding.weight",
            "text_position_ids": "text_model.embeddings.position_ids",
            "token_embedding.embedding": "text_model.embeddings.token_embedding.weight",
            "text_projection.weight": "text_model.head.weight",
            "text_projection.bias": "text_model.head.bias",
            "visual_projection.weight": "visual_projection.weight",
            # MAPHead mappings
            "vision_model.encoder.MAPHead.probe": "vision_model.head.probe",
            "vision_model.encoder.MAPHead.layernorm.weight": "vision_model.head.layernorm.weight",
            "vision_model.encoder.MAPHead.layernorm.bias": "vision_model.head.layernorm.bias",
            "vision_model.encoder.MAPHead.mlp.fc1.weight": "vision_model.head.mlp.fc1.weight",
            "vision_model.encoder.MAPHead.mlp.fc1.bias": "vision_model.head.mlp.fc1.bias",
            "vision_model.encoder.MAPHead.mlp.layers.2.weight": "vision_model.head.mlp.fc2.weight",
            "vision_model.encoder.MAPHead.mlp.layers.2.bias": "vision_model.head.mlp.fc2.bias",
            "vision_model.encoder.MAPHead.attn.in_proj_weight": "vision_model.head.attention.in_proj_weight",
            "vision_model.encoder.MAPHead.attn.in_proj_bias": "vision_model.head.attention.in_proj_bias",
            "vision_model.encoder.MAPHead.self_attn.out_proj.weight": "vision_model.head.attention.out_proj.weight",
            "vision_model.encoder.MAPHead.self_attn.out_proj.bias": "vision_model.head.attention.out_proj.bias",
        }
        _SPECIAL_RENAMINGS = {
            "text_model.layers": "text_model.encoder.layers",
            "vision_model.encoder.encoder.layers": "vision_model.encoder.layers",
            ".attn.query.": ".self_attn.q_proj.",
            ".attn.key.": ".self_attn.k_proj.",
            ".attn.value.": ".self_attn.v_proj.",
            ".attn.out.": ".self_attn.out_proj.",
            ".mlp.layers.0.": ".mlp.fc1.",
            ".mlp.layers.3.": ".mlp.fc2.",
            ".norm1.": ".layer_norm1.",
            ".norm2.": ".layer_norm2.",
        }
        os.makedirs(save_directory, exist_ok=True)

        config = self._original_config.copy() if self._original_config else self._create_config()
        with open(os.path.join(save_directory, "config.json"), "w") as f:
            json.dump(config, f, indent=2)

        _, state = nnx.split(self)
        state_dict = nnx.to_pure_dict(state)

        # Combine Q, K, V projections for MAPHead attention before conversion
        if "vision_model" in state_dict and "MAPHead" in state_dict["vision_model"]["encoder"]:
            maphead = state_dict["vision_model"]["encoder"]["MAPHead"]
            if "attn" in maphead:
                attn = maphead["attn"]

                # Combine weights: [q_weight, k_weight, v_weight] -> in_proj_weight
                if all(k in attn for k in ["query", "key", "value"]):
                    q_weight = attn["query"]["kernel"]  # shape: (vision_width, num_heads, head_dim)
                    k_weight = attn["key"]["kernel"]
                    v_weight = attn["value"]["kernel"]

                    # Reshape to (vision_width, vision_width) and concatenate
                    q_flat = q_weight.reshape(q_weight.shape[0], -1).T  # (vision_width, vision_width)
                    k_flat = k_weight.reshape(k_weight.shape[0], -1).T
                    v_flat = v_weight.reshape(v_weight.shape[0], -1).T

                    in_proj_weight = jnp.concatenate([q_flat, k_flat, v_flat], axis=0)  # (3*vision_width, vision_width)

                    # Combine biases: [q_bias, k_bias, v_bias] -> in_proj_bias
                    q_bias = attn["query"]["bias"]  # shape: (num_heads, head_dim)
                    k_bias = attn["key"]["bias"]
                    v_bias = attn["value"]["bias"]

                    q_bias_flat = q_bias.flatten()
                    k_bias_flat = k_bias.flatten()
                    v_bias_flat = v_bias.flatten()

                    in_proj_bias = jnp.concatenate([q_bias_flat, k_bias_flat, v_bias_flat], axis=0)  # (3*vision_width,)

                    del attn["query"]
                    del attn["key"]
                    del attn["value"]

                    attn["in_proj_weight"] = in_proj_weight
                    attn["in_proj_bias"] = in_proj_bias

                # Handle out_proj: flatten from (num_heads, head_dim, vision_width) to (vision_width, vision_width)
                if "out" in attn:
                    out_weight = attn["out"]["kernel"]  # shape: (num_heads, head_dim, vision_width)
                    out_bias = attn["out"]["bias"]  # shape: (vision_width,)

                    # Flatten the out_proj weight to (vision_width, vision_width)
                    out_weight_flat = out_weight.reshape(-1, out_weight.shape[-1])  # (num_heads*head_dim, vision_width)

                    del attn["out"]
                    if "self_attn" not in maphead:
                        maphead["self_attn"] = {}
                    if "out_proj" not in maphead["self_attn"]:
                        maphead["self_attn"]["out_proj"] = {}
                    maphead["self_attn"]["out_proj"]["weight"] = out_weight_flat
                    maphead["self_attn"]["out_proj"]["bias"] = out_bias

        hf_state = convert_state_to_hf_format(state_dict, _SPECIAL_MAPPINGS, _SPECIAL_RENAMINGS)

        for key in ["logit_scale", "logit_bias"]:
            if key in hf_state and hf_state[key].ndim == 0:
                hf_state[key] = jnp.expand_dims(hf_state[key], 0)

        hf_state.pop("vision_model.encoder.vision_position_ids", None)

        save_safetensors(hf_state, os.path.join(save_directory, "model.safetensors"))

    @classmethod
    def from_pretrained(
        cls,
        model_name_or_path: str,
        use_pytorch: bool = False,
        mesh: Mesh | None = None,
        dtype: DTypeLike = jnp.float32,
        param_dtype: DTypeLike = jnp.float32,
        use_gradient_checkpointing: bool = False,
        rngs: rnglib.Rngs = nnx.Rngs(0),
    ) -> "SigLIP":
        """Load a pretrained SigLIP model from a local path or HuggingFace Hub.

        Args:
            model_name_or_path (str): Path to local weights or HuggingFace model ID.
            use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
            mesh (Mesh | None): Optional device mesh for parameter sharding. Defaults to None.
            dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
            param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
            use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
            rngs (rnglib.Rngs): Random number generator keys. Defaults to nnx.Rngs(0).

        Returns:
            SigLIP: Pretrained SigLIP model
        """
        params_fstate, config_dict = load_params_and_config(model_name_or_path, use_pytorch)
        config: dict[str, Any] = config_dict

        vision_patch_size = params_fstate["vision_model.embeddings.patch_embedding.weight"].shape[3]
        vision_width = params_fstate["vision_model.embeddings.patch_embedding.bias"].shape[0]
        vision_num_layers = 0
        for k in params_fstate:
            if k.startswith("vision_model.encoder.layers.") and k.endswith(".mlp.fc2.bias"):
                vision_num_layers = max(vision_num_layers, int(k.split(".")[3]) + 1)

        context_length = params_fstate["text_model.embeddings.position_embedding.weight"].shape[0]
        vocab_size = params_fstate["text_model.embeddings.token_embedding.weight"].shape[0]
        text_hidden_size = params_fstate["text_model.embeddings.token_embedding.weight"].shape[1]
        text_num_layers = 0
        for k_param in params_fstate:
            if k_param.startswith("text_model.encoder.layers.") and k_param.endswith(".self_attn.q_proj.weight"):
                layer_idx = int(k_param.split(".")[3])
                text_num_layers = max(text_num_layers, layer_idx + 1)

        model = cls(
            image_resolution=config["vision_config"]["image_size"],
            vision_layers=vision_num_layers,
            vision_width=vision_width,
            vision_patch_size=vision_patch_size,
            context_length=context_length,
            vocab_size=vocab_size,
            transformer_width=text_hidden_size,
            transformer_heads=text_hidden_size // 64,
            transformer_layers=text_num_layers,
            use_gradient_checkpointing=use_gradient_checkpointing,
            mesh=mesh,
            dtype=dtype,
            param_dtype=param_dtype,
            rngs=rngs,
        )

        flax_model_params_fstate = dict(nnx.to_flat_state(nnx.state(model, nnx.Param)))
        nonvisited = set(flax_model_params_fstate.keys())
        used_hf_keys: Set[str] = set()

        mapping_list = [
            (("logit_scale",), ("logit_scale",)),
            (("logit_bias",), ("logit_bias",)),
            (("positional_embedding",), ("text_model", "embeddings", "position_embedding", "weight")),
            (("token_embedding", "embedding"), ("text_model", "embeddings", "token_embedding", "weight")),
            (("ln_final", "scale"), ("text_model", "final_layer_norm", "weight")),
            (("ln_final", "bias"), ("text_model", "final_layer_norm", "bias")),
            (("text_projection", "kernel"), ("text_model", "head", "weight")),
            (("text_projection", "bias"), ("text_model", "head", "bias")),
            (("vision_model", "encoder", "patch_embeddings", "kernel"), ("vision_model", "embeddings", "patch_embedding", "weight")),
            (("vision_model", "encoder", "patch_embeddings", "bias"), ("vision_model", "embeddings", "patch_embedding", "bias")),
            (("vision_model", "encoder", "position_embeddings"), ("vision_model", "embeddings", "position_embedding", "weight")),
            (("vision_model", "encoder", "ln_post", "scale"), ("vision_model", "post_layernorm", "weight")),
            (("vision_model", "encoder", "ln_post", "bias"), ("vision_model", "post_layernorm", "bias")),
            (("vision_model", "encoder", "MAPHead", "probe"), ("vision_model", "head", "probe")),
            (("vision_model", "encoder", "MAPHead", "layernorm", "scale"), ("vision_model", "head", "layernorm", "weight")),
            (("vision_model", "encoder", "MAPHead", "layernorm", "bias"), ("vision_model", "head", "layernorm", "bias")),
            (("vision_model", "encoder", "MAPHead", "mlp", "layers", 0, "kernel"), ("vision_model", "head", "mlp", "fc1", "weight")),
            (("vision_model", "encoder", "MAPHead", "mlp", "layers", 0, "bias"), ("vision_model", "head", "mlp", "fc1", "bias")),
            (("vision_model", "encoder", "MAPHead", "mlp", "layers", 2, "kernel"), ("vision_model", "head", "mlp", "fc2", "weight")),
            (("vision_model", "encoder", "MAPHead", "mlp", "layers", 2, "bias"), ("vision_model", "head", "mlp", "fc2", "bias")),
            (("vision_model", "encoder", "MAPHead", "attn", "query", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("vision_model", "encoder", "MAPHead", "attn", "query", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("vision_model", "encoder", "MAPHead", "attn", "key", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("vision_model", "encoder", "MAPHead", "attn", "key", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("vision_model", "encoder", "MAPHead", "attn", "value", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
            (("vision_model", "encoder", "MAPHead", "attn", "value", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
            (("vision_model", "encoder", "MAPHead", "attn", "out", "kernel"), ("vision_model", "head", "attention", "out_proj", "weight")),
            (("vision_model", "encoder", "MAPHead", "attn", "out", "bias"), ("vision_model", "head", "attention", "out_proj", "bias")),
        ]

        for i in range(text_num_layers):
            flax_base = ("text_model", "layers", i)
            hf_base = ("text_model", "encoder", "layers", str(i))

            mapping_list.extend(
                [
                    (flax_base + ("attn", "query", "kernel"), hf_base + ("self_attn", "q_proj", "weight")),
                    (flax_base + ("attn", "query", "bias"), hf_base + ("self_attn", "q_proj", "bias")),
                    (flax_base + ("attn", "key", "kernel"), hf_base + ("self_attn", "k_proj", "weight")),
                    (flax_base + ("attn", "key", "bias"), hf_base + ("self_attn", "k_proj", "bias")),
                    (flax_base + ("attn", "value", "kernel"), hf_base + ("self_attn", "v_proj", "weight")),
                    (flax_base + ("attn", "value", "bias"), hf_base + ("self_attn", "v_proj", "bias")),
                    (flax_base + ("attn", "out", "kernel"), hf_base + ("self_attn", "out_proj", "weight")),
                    (flax_base + ("attn", "out", "bias"), hf_base + ("self_attn", "out_proj", "bias")),
                    (flax_base + ("norm1", "scale"), hf_base + ("layer_norm1", "weight")),
                    (flax_base + ("norm1", "bias"), hf_base + ("layer_norm1", "bias")),
                    (flax_base + ("norm2", "scale"), hf_base + ("layer_norm2", "weight")),
                    (flax_base + ("norm2", "bias"), hf_base + ("layer_norm2", "bias")),
                    (flax_base + ("mlp", "layers", 0, "kernel"), hf_base + ("mlp", "fc1", "weight")),
                    (flax_base + ("mlp", "layers", 0, "bias"), hf_base + ("mlp", "fc1", "bias")),
                    (flax_base + ("mlp", "layers", 3, "kernel"), hf_base + ("mlp", "fc2", "weight")),
                    (flax_base + ("mlp", "layers", 3, "bias"), hf_base + ("mlp", "fc2", "bias")),
                ]
            )

        for i in range(vision_num_layers):
            flax_base = ("vision_model", "encoder", "encoder", "layers", i)
            hf_base = ("vision_model", "encoder", "layers", str(i))
            mapping_list.extend(
                [
                    (flax_base + ("attn", "query", "kernel"), hf_base + ("self_attn", "q_proj", "weight")),
                    (flax_base + ("attn", "query", "bias"), hf_base + ("self_attn", "q_proj", "bias")),
                    (flax_base + ("attn", "key", "kernel"), hf_base + ("self_attn", "k_proj", "weight")),
                    (flax_base + ("attn", "key", "bias"), hf_base + ("self_attn", "k_proj", "bias")),
                    (flax_base + ("attn", "value", "kernel"), hf_base + ("self_attn", "v_proj", "weight")),
                    (flax_base + ("attn", "value", "bias"), hf_base + ("self_attn", "v_proj", "bias")),
                    (flax_base + ("attn", "out", "kernel"), hf_base + ("self_attn", "out_proj", "weight")),
                    (flax_base + ("attn", "out", "bias"), hf_base + ("self_attn", "out_proj", "bias")),
                    (flax_base + ("norm1", "scale"), hf_base + ("layer_norm1", "weight")),
                    (flax_base + ("norm1", "bias"), hf_base + ("layer_norm1", "bias")),
                    (flax_base + ("norm2", "scale"), hf_base + ("layer_norm2", "weight")),
                    (flax_base + ("norm2", "bias"), hf_base + ("layer_norm2", "bias")),
                    (flax_base + ("mlp", "layers", 0, "kernel"), hf_base + ("mlp", "fc1", "weight")),
                    (flax_base + ("mlp", "layers", 0, "bias"), hf_base + ("mlp", "fc1", "bias")),
                    (flax_base + ("mlp", "layers", 3, "kernel"), hf_base + ("mlp", "fc2", "weight")),
                    (flax_base + ("mlp", "layers", 3, "bias"), hf_base + ("mlp", "fc2", "bias")),
                ]
            )

        params_name_mapping = dict(mapping_list)

        for flax_dst_key_tuple, hf_src_key_tuple in params_name_mapping.items():
            hf_src_key_as_string = ".".join(hf_src_key_tuple)
            nonvisited.discard(flax_dst_key_tuple)
            used_hf_keys.add(hf_src_key_as_string)
            src_value = params_fstate[hf_src_key_as_string]
            dst_value_obj = flax_model_params_fstate[flax_dst_key_tuple]

            if flax_dst_key_tuple == ("vision_model", "encoder", "patch_embeddings", "kernel"):
                src_value = jnp.transpose(src_value, (2, 3, 1, 0))
            elif flax_dst_key_tuple == ("vision_model", "encoder", "position_embeddings"):
                src_value = src_value.reshape(1, src_value.shape[0], src_value.shape[1])
            elif flax_dst_key_tuple in [("logit_scale",), ("logit_bias",)]:
                src_value = jnp.squeeze(src_value)
            elif hf_src_key_tuple[-1] == "weight" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
                src_value = jnp.transpose(src_value, (1, 0))
                if "text_model" in hf_src_key_as_string:
                    num_heads = model.transformer_heads
                    head_dim = model.transformer_width // num_heads
                    src_value = src_value.reshape((model.transformer_width, num_heads, head_dim))
                else:
                    num_heads = model.vision_heads
                    head_dim = vision_width // num_heads
                    src_value = src_value.reshape((vision_width, num_heads, head_dim))
            elif hf_src_key_tuple[-1] == "bias" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
                if "text_model" in hf_src_key_as_string:
                    num_heads = model.transformer_heads
                    head_dim = model.transformer_width // num_heads
                else:
                    num_heads = model.vision_heads
                    head_dim = vision_width // num_heads
                src_value = src_value.reshape((num_heads, head_dim))
            elif hf_src_key_tuple[-2:] == ("out_proj", "weight"):
                src_value = jnp.transpose(src_value, (1, 0))
                if "text_model" in hf_src_key_as_string:
                    num_heads = model.transformer_heads
                    head_dim = model.transformer_width // num_heads
                    src_value = src_value.reshape((num_heads, head_dim, model.transformer_width))
                else:
                    num_heads = model.vision_heads
                    head_dim = vision_width // num_heads
                    src_value = src_value.reshape((num_heads, head_dim, vision_width))
            elif hf_src_key_tuple[-1] == "in_proj_weight":
                num_heads = model.vision_heads
                head_dim = vision_width // num_heads
                q_w, k_w, v_w = jnp.split(src_value, 3, axis=0)
                w_map = {"query": q_w, "key": k_w, "value": v_w}
                src_value = jnp.transpose(w_map[flax_dst_key_tuple[-2]], (1, 0)).reshape(vision_width, num_heads, head_dim)
            elif hf_src_key_tuple[-1] == "in_proj_bias":
                num_heads = model.vision_heads
                head_dim = vision_width // num_heads
                q_b, k_b, v_b = jnp.split(src_value, 3, axis=0)
                b_map = {"query": q_b, "key": k_b, "value": v_b}
                src_value = b_map[flax_dst_key_tuple[-2]].reshape(num_heads, head_dim)
            elif hf_src_key_tuple[-1] == "weight" and src_value.ndim == 2:
                if "position_embedding" not in hf_src_key_as_string and "token_embedding" not in hf_src_key_as_string:
                    src_value = jnp.transpose(src_value, (1, 0))
            if src_value.shape != dst_value_obj.value.shape:
                raise ValueError(f"Shape mismatch for {flax_dst_key_tuple} (Flax) vs {hf_src_key_as_string} (HF): {dst_value_obj.value.shape} (expected) != {src_value.shape} (actual)")

            src_value = src_value.astype(param_dtype)
            dst_value_obj.value = src_value

        nnx.update(model, nnx.from_flat_state(flax_model_params_fstate))

        hf_checkpoint_keys: Set[str] = set(params_fstate.keys())
        leftover_hf_keys = hf_checkpoint_keys - used_hf_keys
        known_unused_hf_buffer_keys = {
            "text_model.embeddings.position_ids",
            "vision_model.embeddings.position_ids",
        }
        unexpected_leftover_hf_keys = leftover_hf_keys - known_unused_hf_buffer_keys

        assert len(unexpected_leftover_hf_keys) == 0, f"Some unexpected HuggingFace checkpoint parameters were not used: {sorted(list(unexpected_leftover_hf_keys))}"
        model._original_config = config
        return model

__call__(image, text)

Calculate similarity between image and text embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required
text Int[Array, 'batch context_length']

Batch of token sequences.

 required

Returns:

Type Description
Float[Array, 'batch batch']

Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
Source code in src/jimm/models/siglip.py 
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
def __call__(self, image: Float[Array, "batch height width channels"], text: Int[Array, "batch context_length"]) -> Float[Array, "batch batch"]:
    """
    Calculate similarity between image and text embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.
        text (Int[Array, "batch context_length"]): Batch of token sequences.

    Returns:
        Float[Array, "batch batch"]: Similarity scores between all pairs of images and texts.
    """
    image_features: Float[Array, "batch transformer_width"] = self.encode_image(image)
    text_features: Float[Array, "batch transformer_width"] = self.encode_text(text)

    image_features: Float[Array, "batch transformer_width"] = image_features / jnp.linalg.norm(image_features, axis=-1, keepdims=True)
    text_features: Float[Array, "batch transformer_width"] = text_features / jnp.linalg.norm(text_features, axis=-1, keepdims=True)

    logit_scale: Float[Array, ""] = jnp.exp(self.logit_scale.value)
    logits: Float[Array, "batch batch"] = logit_scale * image_features @ text_features.T + self.logit_bias.value
    return logits

__init__(image_resolution, vision_layers, vision_width, vision_patch_size, context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, use_gradient_checkpointing=False, rngs=nnx.Rngs(0), dtype=jnp.float32, param_dtype=jnp.float32, mesh=None)

Initialize the SigLIP model.

Parameters:

Name Type Description Default
image_resolution int

The resolution of the input images.

 required
vision_layers int

The number of layers in the vision transformer.

 required
vision_width int

The width of the vision transformer.

 required
vision_patch_size int

The patch size of the vision transformer.

 required
context_length int

The length of the context.

 required
vocab_size int

The size of the vocabulary.

 required
transformer_width int

The width of the transformer.

 required
transformer_heads int

The number of attention heads in the transformer.

 required
transformer_layers int

The number of layers in the transformer.

 required
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
rngs Rngs

The random number generator state. Defaults to nnx.Rngs(0).

 Rngs(0)
dtype DTypeLike

The data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

The data type for parameters. Defaults to jnp.float32.

 float32
mesh Mesh | None

Optional device mesh for parameter sharding. Defaults to None.

 None
Source code in src/jimm/models/siglip.py 
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
def __init__(
    self,
    image_resolution: int,
    vision_layers: int,
    vision_width: int,
    vision_patch_size: int,
    context_length: int,
    vocab_size: int,
    transformer_width: int,
    transformer_heads: int,
    transformer_layers: int,
    use_gradient_checkpointing: bool = False,
    rngs: rnglib.Rngs = nnx.Rngs(0),
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    mesh: Mesh | None = None,
):
    """
    Initialize the SigLIP model.

    Args:
        image_resolution (int): The resolution of the input images.
        vision_layers (int): The number of layers in the vision transformer.
        vision_width (int): The width of the vision transformer.
        vision_patch_size (int): The patch size of the vision transformer.
        context_length (int): The length of the context.
        vocab_size (int): The size of the vocabulary.
        transformer_width (int): The width of the transformer.
        transformer_heads (int): The number of attention heads in the transformer.
        transformer_layers (int): The number of layers in the transformer.
        use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
        rngs (rnglib.Rngs): The random number generator state. Defaults to nnx.Rngs(0).
        dtype (DTypeLike): The data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike): The data type for parameters. Defaults to jnp.float32.
        mesh (Mesh | None): Optional device mesh for parameter sharding. Defaults to None.
    """
    self.vision_layers = vision_layers
    self.vision_width = vision_width
    self.vision_patch_size = vision_patch_size
    self.context_length = context_length
    self.transformer_width = transformer_width
    self.transformer_heads = transformer_heads
    self.transformer_layers = transformer_layers
    self.dtype = dtype
    self._original_config = None

    self.vision_heads = vision_width // 64
    self.vision_model = SigLIPVisionModel(
        image_resolution=image_resolution,
        vision_layers=vision_layers,
        vision_width=vision_width,
        vision_patch_size=vision_patch_size,
        use_gradient_checkpointing=use_gradient_checkpointing,
        rngs=rngs,
        dtype=dtype,
        param_dtype=param_dtype,
        mesh=mesh,
    )

    self.text_model = Transformer(
        width=transformer_width,
        mlp_dim=transformer_width * 4,
        layers=transformer_layers,
        num_heads=transformer_heads,
        dropout_rate=0.0,
        layernorm_epsilon=1e-6,
        use_quick_gelu=False,
        use_gradient_checkpointing=use_gradient_checkpointing,
        dtype=dtype,
        param_dtype=param_dtype,
        mesh=mesh,
        rngs=rngs,
    )
    self.vocab_size = vocab_size
    self.token_embedding = nnx.Embed(
        num_embeddings=vocab_size,
        features=transformer_width,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        embedding_init=sharded_init(nnx.initializers.xavier_uniform(), P("model", None), mesh),
    )
    self.positional_embedding = nnx.Param(sharded_init(nnx.initializers.truncated_normal(stddev=0.02), P("model", None), mesh)(rngs.params(), (context_length, transformer_width)))
    self.ln_final = nnx.LayerNorm(
        transformer_width,
        epsilon=1e-6,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        scale_init=sharded_init(nnx.initializers.ones_init(), P("model"), mesh),
        bias_init=sharded_init(nnx.initializers.zeros_init(), P("model"), mesh),
    )
    self.text_projection = nnx.Linear(
        transformer_width,
        transformer_width,
        use_bias=True,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
        kernel_init=sharded_init(nnx.initializers.xavier_uniform(), P("model", None), mesh),
    )
    self.logit_scale = nnx.Param(sharded_init(nnx.initializers.ones_init(), P(), mesh)(rngs.params(), ()))
    self.logit_bias = nnx.Param(sharded_init(nnx.initializers.ones_init(), P(), mesh)(rngs.params(), ()))

encode_image(image)

Encode images into embeddings.

Parameters:

Name Type Description Default
image Float[Array, 'batch height width channels']

Batch of input images.

 required

Returns:

Type Description
Float[Array, 'batch transformer_width']

Float[Array, "batch transformer_width"]: Image embeddings.
Source code in src/jimm/models/siglip.py 
337
338
339
340
341
342
343
344
345
346
347
def encode_image(self, image: Float[Array, "batch height width channels"]) -> Float[Array, "batch transformer_width"]:
    """
    Encode images into embeddings.

    Args:
        image (Float[Array, "batch height width channels"]): Batch of input images.

    Returns:
        Float[Array, "batch transformer_width"]: Image embeddings.
    """
    return self.vision_model(image)

encode_text(text)

Encode text tokens into embeddings.

Parameters:

Name Type Description Default
text Int[Array, 'batch context_length']

Batch of token sequences.

 required

Returns:

Type Description
Float[Array, 'batch transformer_width']

Float[Array, "batch transformer_width"]: Text embeddings.
Source code in src/jimm/models/siglip.py 
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
def encode_text(self, text: Int[Array, "batch context_length"]) -> Float[Array, "batch transformer_width"]:
    """
    Encode text tokens into embeddings.

    Args:
        text (Int[Array, "batch context_length"]): Batch of token sequences.

    Returns:
        Float[Array, "batch transformer_width"]: Text embeddings.
    """
    seq_len = text.shape[1]
    x: Float[Array, "batch context_length transformer_width"] = self.token_embedding(text)
    x: Float[Array, "batch context_length transformer_width"] = x + self.positional_embedding.value[:seq_len]
    x: Float[Array, "batch context_length transformer_width"] = self.text_model(x)
    x: Float[Array, "batch context_length transformer_width"] = self.ln_final(x)

    pooled_output = x[:, -1, :]
    x: Float[Array, "batch transformer_width"] = self.text_projection(pooled_output)
    return x

from_pretrained(model_name_or_path, use_pytorch=False, mesh=None, dtype=jnp.float32, param_dtype=jnp.float32, use_gradient_checkpointing=False, rngs=nnx.Rngs(0)) classmethod

Load a pretrained SigLIP model from a local path or HuggingFace Hub.

Parameters:

Name Type Description Default
model_name_or_path str

Path to local weights or HuggingFace model ID.

 required
use_pytorch bool

Whether to load from PyTorch weights. Defaults to False.

 False
mesh Mesh | None

Optional device mesh for parameter sharding. Defaults to None.

 None
dtype DTypeLike

Data type for computations. Defaults to jnp.float32.

 float32
param_dtype DTypeLike

Data type for parameters. Defaults to jnp.float32.

 float32
use_gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
rngs Rngs

Random number generator keys. Defaults to nnx.Rngs(0).

 Rngs(0)

Returns:

Name Type Description
SigLIP SigLIP

Pretrained SigLIP model
Source code in src/jimm/models/siglip.py 
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
@classmethod
def from_pretrained(
    cls,
    model_name_or_path: str,
    use_pytorch: bool = False,
    mesh: Mesh | None = None,
    dtype: DTypeLike = jnp.float32,
    param_dtype: DTypeLike = jnp.float32,
    use_gradient_checkpointing: bool = False,
    rngs: rnglib.Rngs = nnx.Rngs(0),
) -> "SigLIP":
    """Load a pretrained SigLIP model from a local path or HuggingFace Hub.

    Args:
        model_name_or_path (str): Path to local weights or HuggingFace model ID.
        use_pytorch (bool): Whether to load from PyTorch weights. Defaults to False.
        mesh (Mesh | None): Optional device mesh for parameter sharding. Defaults to None.
        dtype (DTypeLike): Data type for computations. Defaults to jnp.float32.
        param_dtype (DTypeLike): Data type for parameters. Defaults to jnp.float32.
        use_gradient_checkpointing (bool): Whether to use gradient checkpointing. Defaults to False.
        rngs (rnglib.Rngs): Random number generator keys. Defaults to nnx.Rngs(0).

    Returns:
        SigLIP: Pretrained SigLIP model
    """
    params_fstate, config_dict = load_params_and_config(model_name_or_path, use_pytorch)
    config: dict[str, Any] = config_dict

    vision_patch_size = params_fstate["vision_model.embeddings.patch_embedding.weight"].shape[3]
    vision_width = params_fstate["vision_model.embeddings.patch_embedding.bias"].shape[0]
    vision_num_layers = 0
    for k in params_fstate:
        if k.startswith("vision_model.encoder.layers.") and k.endswith(".mlp.fc2.bias"):
            vision_num_layers = max(vision_num_layers, int(k.split(".")[3]) + 1)

    context_length = params_fstate["text_model.embeddings.position_embedding.weight"].shape[0]
    vocab_size = params_fstate["text_model.embeddings.token_embedding.weight"].shape[0]
    text_hidden_size = params_fstate["text_model.embeddings.token_embedding.weight"].shape[1]
    text_num_layers = 0
    for k_param in params_fstate:
        if k_param.startswith("text_model.encoder.layers.") and k_param.endswith(".self_attn.q_proj.weight"):
            layer_idx = int(k_param.split(".")[3])
            text_num_layers = max(text_num_layers, layer_idx + 1)

    model = cls(
        image_resolution=config["vision_config"]["image_size"],
        vision_layers=vision_num_layers,
        vision_width=vision_width,
        vision_patch_size=vision_patch_size,
        context_length=context_length,
        vocab_size=vocab_size,
        transformer_width=text_hidden_size,
        transformer_heads=text_hidden_size // 64,
        transformer_layers=text_num_layers,
        use_gradient_checkpointing=use_gradient_checkpointing,
        mesh=mesh,
        dtype=dtype,
        param_dtype=param_dtype,
        rngs=rngs,
    )

    flax_model_params_fstate = dict(nnx.to_flat_state(nnx.state(model, nnx.Param)))
    nonvisited = set(flax_model_params_fstate.keys())
    used_hf_keys: Set[str] = set()

    mapping_list = [
        (("logit_scale",), ("logit_scale",)),
        (("logit_bias",), ("logit_bias",)),
        (("positional_embedding",), ("text_model", "embeddings", "position_embedding", "weight")),
        (("token_embedding", "embedding"), ("text_model", "embeddings", "token_embedding", "weight")),
        (("ln_final", "scale"), ("text_model", "final_layer_norm", "weight")),
        (("ln_final", "bias"), ("text_model", "final_layer_norm", "bias")),
        (("text_projection", "kernel"), ("text_model", "head", "weight")),
        (("text_projection", "bias"), ("text_model", "head", "bias")),
        (("vision_model", "encoder", "patch_embeddings", "kernel"), ("vision_model", "embeddings", "patch_embedding", "weight")),
        (("vision_model", "encoder", "patch_embeddings", "bias"), ("vision_model", "embeddings", "patch_embedding", "bias")),
        (("vision_model", "encoder", "position_embeddings"), ("vision_model", "embeddings", "position_embedding", "weight")),
        (("vision_model", "encoder", "ln_post", "scale"), ("vision_model", "post_layernorm", "weight")),
        (("vision_model", "encoder", "ln_post", "bias"), ("vision_model", "post_layernorm", "bias")),
        (("vision_model", "encoder", "MAPHead", "probe"), ("vision_model", "head", "probe")),
        (("vision_model", "encoder", "MAPHead", "layernorm", "scale"), ("vision_model", "head", "layernorm", "weight")),
        (("vision_model", "encoder", "MAPHead", "layernorm", "bias"), ("vision_model", "head", "layernorm", "bias")),
        (("vision_model", "encoder", "MAPHead", "mlp", "layers", 0, "kernel"), ("vision_model", "head", "mlp", "fc1", "weight")),
        (("vision_model", "encoder", "MAPHead", "mlp", "layers", 0, "bias"), ("vision_model", "head", "mlp", "fc1", "bias")),
        (("vision_model", "encoder", "MAPHead", "mlp", "layers", 2, "kernel"), ("vision_model", "head", "mlp", "fc2", "weight")),
        (("vision_model", "encoder", "MAPHead", "mlp", "layers", 2, "bias"), ("vision_model", "head", "mlp", "fc2", "bias")),
        (("vision_model", "encoder", "MAPHead", "attn", "query", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
        (("vision_model", "encoder", "MAPHead", "attn", "query", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
        (("vision_model", "encoder", "MAPHead", "attn", "key", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
        (("vision_model", "encoder", "MAPHead", "attn", "key", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
        (("vision_model", "encoder", "MAPHead", "attn", "value", "kernel"), ("vision_model", "head", "attention", "in_proj_weight")),
        (("vision_model", "encoder", "MAPHead", "attn", "value", "bias"), ("vision_model", "head", "attention", "in_proj_bias")),
        (("vision_model", "encoder", "MAPHead", "attn", "out", "kernel"), ("vision_model", "head", "attention", "out_proj", "weight")),
        (("vision_model", "encoder", "MAPHead", "attn", "out", "bias"), ("vision_model", "head", "attention", "out_proj", "bias")),
    ]

    for i in range(text_num_layers):
        flax_base = ("text_model", "layers", i)
        hf_base = ("text_model", "encoder", "layers", str(i))

        mapping_list.extend(
            [
                (flax_base + ("attn", "query", "kernel"), hf_base + ("self_attn", "q_proj", "weight")),
                (flax_base + ("attn", "query", "bias"), hf_base + ("self_attn", "q_proj", "bias")),
                (flax_base + ("attn", "key", "kernel"), hf_base + ("self_attn", "k_proj", "weight")),
                (flax_base + ("attn", "key", "bias"), hf_base + ("self_attn", "k_proj", "bias")),
                (flax_base + ("attn", "value", "kernel"), hf_base + ("self_attn", "v_proj", "weight")),
                (flax_base + ("attn", "value", "bias"), hf_base + ("self_attn", "v_proj", "bias")),
                (flax_base + ("attn", "out", "kernel"), hf_base + ("self_attn", "out_proj", "weight")),
                (flax_base + ("attn", "out", "bias"), hf_base + ("self_attn", "out_proj", "bias")),
                (flax_base + ("norm1", "scale"), hf_base + ("layer_norm1", "weight")),
                (flax_base + ("norm1", "bias"), hf_base + ("layer_norm1", "bias")),
                (flax_base + ("norm2", "scale"), hf_base + ("layer_norm2", "weight")),
                (flax_base + ("norm2", "bias"), hf_base + ("layer_norm2", "bias")),
                (flax_base + ("mlp", "layers", 0, "kernel"), hf_base + ("mlp", "fc1", "weight")),
                (flax_base + ("mlp", "layers", 0, "bias"), hf_base + ("mlp", "fc1", "bias")),
                (flax_base + ("mlp", "layers", 3, "kernel"), hf_base + ("mlp", "fc2", "weight")),
                (flax_base + ("mlp", "layers", 3, "bias"), hf_base + ("mlp", "fc2", "bias")),
            ]
        )

    for i in range(vision_num_layers):
        flax_base = ("vision_model", "encoder", "encoder", "layers", i)
        hf_base = ("vision_model", "encoder", "layers", str(i))
        mapping_list.extend(
            [
                (flax_base + ("attn", "query", "kernel"), hf_base + ("self_attn", "q_proj", "weight")),
                (flax_base + ("attn", "query", "bias"), hf_base + ("self_attn", "q_proj", "bias")),
                (flax_base + ("attn", "key", "kernel"), hf_base + ("self_attn", "k_proj", "weight")),
                (flax_base + ("attn", "key", "bias"), hf_base + ("self_attn", "k_proj", "bias")),
                (flax_base + ("attn", "value", "kernel"), hf_base + ("self_attn", "v_proj", "weight")),
                (flax_base + ("attn", "value", "bias"), hf_base + ("self_attn", "v_proj", "bias")),
                (flax_base + ("attn", "out", "kernel"), hf_base + ("self_attn", "out_proj", "weight")),
                (flax_base + ("attn", "out", "bias"), hf_base + ("self_attn", "out_proj", "bias")),
                (flax_base + ("norm1", "scale"), hf_base + ("layer_norm1", "weight")),
                (flax_base + ("norm1", "bias"), hf_base + ("layer_norm1", "bias")),
                (flax_base + ("norm2", "scale"), hf_base + ("layer_norm2", "weight")),
                (flax_base + ("norm2", "bias"), hf_base + ("layer_norm2", "bias")),
                (flax_base + ("mlp", "layers", 0, "kernel"), hf_base + ("mlp", "fc1", "weight")),
                (flax_base + ("mlp", "layers", 0, "bias"), hf_base + ("mlp", "fc1", "bias")),
                (flax_base + ("mlp", "layers", 3, "kernel"), hf_base + ("mlp", "fc2", "weight")),
                (flax_base + ("mlp", "layers", 3, "bias"), hf_base + ("mlp", "fc2", "bias")),
            ]
        )

    params_name_mapping = dict(mapping_list)

    for flax_dst_key_tuple, hf_src_key_tuple in params_name_mapping.items():
        hf_src_key_as_string = ".".join(hf_src_key_tuple)
        nonvisited.discard(flax_dst_key_tuple)
        used_hf_keys.add(hf_src_key_as_string)
        src_value = params_fstate[hf_src_key_as_string]
        dst_value_obj = flax_model_params_fstate[flax_dst_key_tuple]

        if flax_dst_key_tuple == ("vision_model", "encoder", "patch_embeddings", "kernel"):
            src_value = jnp.transpose(src_value, (2, 3, 1, 0))
        elif flax_dst_key_tuple == ("vision_model", "encoder", "position_embeddings"):
            src_value = src_value.reshape(1, src_value.shape[0], src_value.shape[1])
        elif flax_dst_key_tuple in [("logit_scale",), ("logit_bias",)]:
            src_value = jnp.squeeze(src_value)
        elif hf_src_key_tuple[-1] == "weight" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
            src_value = jnp.transpose(src_value, (1, 0))
            if "text_model" in hf_src_key_as_string:
                num_heads = model.transformer_heads
                head_dim = model.transformer_width // num_heads
                src_value = src_value.reshape((model.transformer_width, num_heads, head_dim))
            else:
                num_heads = model.vision_heads
                head_dim = vision_width // num_heads
                src_value = src_value.reshape((vision_width, num_heads, head_dim))
        elif hf_src_key_tuple[-1] == "bias" and hf_src_key_tuple[-2] in ("q_proj", "k_proj", "v_proj"):
            if "text_model" in hf_src_key_as_string:
                num_heads = model.transformer_heads
                head_dim = model.transformer_width // num_heads
            else:
                num_heads = model.vision_heads
                head_dim = vision_width // num_heads
            src_value = src_value.reshape((num_heads, head_dim))
        elif hf_src_key_tuple[-2:] == ("out_proj", "weight"):
            src_value = jnp.transpose(src_value, (1, 0))
            if "text_model" in hf_src_key_as_string:
                num_heads = model.transformer_heads
                head_dim = model.transformer_width // num_heads
                src_value = src_value.reshape((num_heads, head_dim, model.transformer_width))
            else:
                num_heads = model.vision_heads
                head_dim = vision_width // num_heads
                src_value = src_value.reshape((num_heads, head_dim, vision_width))
        elif hf_src_key_tuple[-1] == "in_proj_weight":
            num_heads = model.vision_heads
            head_dim = vision_width // num_heads
            q_w, k_w, v_w = jnp.split(src_value, 3, axis=0)
            w_map = {"query": q_w, "key": k_w, "value": v_w}
            src_value = jnp.transpose(w_map[flax_dst_key_tuple[-2]], (1, 0)).reshape(vision_width, num_heads, head_dim)
        elif hf_src_key_tuple[-1] == "in_proj_bias":
            num_heads = model.vision_heads
            head_dim = vision_width // num_heads
            q_b, k_b, v_b = jnp.split(src_value, 3, axis=0)
            b_map = {"query": q_b, "key": k_b, "value": v_b}
            src_value = b_map[flax_dst_key_tuple[-2]].reshape(num_heads, head_dim)
        elif hf_src_key_tuple[-1] == "weight" and src_value.ndim == 2:
            if "position_embedding" not in hf_src_key_as_string and "token_embedding" not in hf_src_key_as_string:
                src_value = jnp.transpose(src_value, (1, 0))
        if src_value.shape != dst_value_obj.value.shape:
            raise ValueError(f"Shape mismatch for {flax_dst_key_tuple} (Flax) vs {hf_src_key_as_string} (HF): {dst_value_obj.value.shape} (expected) != {src_value.shape} (actual)")

        src_value = src_value.astype(param_dtype)
        dst_value_obj.value = src_value

    nnx.update(model, nnx.from_flat_state(flax_model_params_fstate))

    hf_checkpoint_keys: Set[str] = set(params_fstate.keys())
    leftover_hf_keys = hf_checkpoint_keys - used_hf_keys
    known_unused_hf_buffer_keys = {
        "text_model.embeddings.position_ids",
        "vision_model.embeddings.position_ids",
    }
    unexpected_leftover_hf_keys = leftover_hf_keys - known_unused_hf_buffer_keys

    assert len(unexpected_leftover_hf_keys) == 0, f"Some unexpected HuggingFace checkpoint parameters were not used: {sorted(list(unexpected_leftover_hf_keys))}"
    model._original_config = config
    return model