Spaces:
Running
Running
remove bias and minor fixes
Browse files- dalle_mini/modeling_bart_flax.py +30 -55
dalle_mini/modeling_bart_flax.py
CHANGED
@@ -44,7 +44,7 @@ from transformers.modeling_flax_utils import (
|
|
44 |
from transformers.utils import logging
|
45 |
|
46 |
|
47 |
-
from configuration_bart import BartConfig
|
48 |
|
49 |
|
50 |
logger = logging.get_logger(__name__)
|
@@ -80,7 +80,7 @@ class FlaxBartAttention(nn.Module):
|
|
80 |
dense = partial(
|
81 |
nn.Dense,
|
82 |
self.embed_dim,
|
83 |
-
use_bias=
|
84 |
dtype=self.dtype,
|
85 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
86 |
)
|
@@ -242,10 +242,14 @@ class FlaxBartEncoderLayer(nn.Module):
|
|
242 |
self.fc1 = nn.Dense(
|
243 |
self.config.encoder_ffn_dim,
|
244 |
dtype=self.dtype,
|
|
|
245 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
246 |
)
|
247 |
self.fc2 = nn.Dense(
|
248 |
-
self.embed_dim,
|
|
|
|
|
|
|
249 |
)
|
250 |
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
251 |
|
@@ -325,14 +329,18 @@ class FlaxBartDecoderLayer(nn.Module):
|
|
325 |
dropout=self.config.attention_dropout,
|
326 |
dtype=self.dtype,
|
327 |
)
|
328 |
-
self.encoder_attn_layer_norm = nn
|
329 |
self.fc1 = nn.Dense(
|
330 |
self.config.encoder_ffn_dim,
|
331 |
dtype=self.dtype,
|
|
|
332 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
333 |
)
|
334 |
self.fc2 = nn.Dense(
|
335 |
-
self.embed_dim,
|
|
|
|
|
|
|
336 |
)
|
337 |
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
338 |
|
@@ -414,7 +422,6 @@ class FlaxBartDecoderLayerCollection(nn.Module):
|
|
414 |
class FlaxBartEncoder(nn.Module):
|
415 |
config: BartConfig
|
416 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
417 |
-
embed_tokens: Optional[nn.Embed] = None
|
418 |
|
419 |
def setup(self):
|
420 |
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
@@ -424,16 +431,15 @@ class FlaxBartEncoder(nn.Module):
|
|
424 |
self.max_source_positions = self.config.max_position_embeddings
|
425 |
self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
|
426 |
|
427 |
-
|
428 |
-
self.
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
)
|
433 |
|
434 |
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
435 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
436 |
-
self.offset =
|
437 |
self.embed_positions = nn.Embed(
|
438 |
self.config.max_position_embeddings + self.offset,
|
439 |
embed_dim,
|
@@ -472,7 +478,6 @@ class FlaxBartEncoder(nn.Module):
|
|
472 |
class FlaxBartDecoder(nn.Module):
|
473 |
config: BartConfig
|
474 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
475 |
-
embed_tokens: Optional[nn.Embed] = None
|
476 |
|
477 |
def setup(self):
|
478 |
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
@@ -482,18 +487,17 @@ class FlaxBartDecoder(nn.Module):
|
|
482 |
self.max_target_positions = self.config.max_position_embeddings
|
483 |
self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
|
484 |
|
485 |
-
|
486 |
-
self.
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
)
|
491 |
|
492 |
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
493 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
494 |
-
self.offset =
|
495 |
self.embed_positions = nn.Embed(
|
496 |
-
self.config.
|
497 |
embed_dim,
|
498 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
499 |
)
|
@@ -546,20 +550,8 @@ class FlaxBartModule(nn.Module):
|
|
546 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
547 |
|
548 |
def setup(self):
|
549 |
-
self.
|
550 |
-
|
551 |
-
self.config.d_model,
|
552 |
-
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
553 |
-
)
|
554 |
-
# a separate embedding is used for the decoder
|
555 |
-
self.decoder_embed = nn.Embed(
|
556 |
-
self.config.decoder_vocab_size,
|
557 |
-
self.config.d_model,
|
558 |
-
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
559 |
-
)
|
560 |
-
|
561 |
-
self.encoder = FlaxBartEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)
|
562 |
-
self.decoder = FlaxBartDecoder(self.config, dtype=self.dtype, embed_tokens=self.decoder_embed)
|
563 |
|
564 |
def _get_encoder_module(self):
|
565 |
return self.encoder
|
@@ -575,8 +567,6 @@ class FlaxBartModule(nn.Module):
|
|
575 |
decoder_attention_mask,
|
576 |
position_ids,
|
577 |
decoder_position_ids,
|
578 |
-
output_attentions: bool = False,
|
579 |
-
output_hidden_states: bool = False,
|
580 |
return_dict: bool = True,
|
581 |
deterministic: bool = True,
|
582 |
):
|
@@ -584,9 +574,6 @@ class FlaxBartModule(nn.Module):
|
|
584 |
input_ids=input_ids,
|
585 |
attention_mask=attention_mask,
|
586 |
position_ids=position_ids,
|
587 |
-
output_attentions=output_attentions,
|
588 |
-
output_hidden_states=output_hidden_states,
|
589 |
-
return_dict=return_dict,
|
590 |
deterministic=deterministic,
|
591 |
)
|
592 |
|
@@ -596,9 +583,6 @@ class FlaxBartModule(nn.Module):
|
|
596 |
position_ids=decoder_position_ids,
|
597 |
encoder_hidden_states=encoder_outputs[0],
|
598 |
encoder_attention_mask=attention_mask,
|
599 |
-
output_attentions=output_attentions,
|
600 |
-
output_hidden_states=output_hidden_states,
|
601 |
-
return_dict=return_dict,
|
602 |
deterministic=deterministic,
|
603 |
)
|
604 |
|
@@ -629,8 +613,8 @@ class FlaxBartPreTrainedModel(FlaxPreTrainedModel):
|
|
629 |
dtype: jnp.dtype = jnp.float32,
|
630 |
**kwargs,
|
631 |
):
|
632 |
-
module = self.module_class(config=config, dtype=dtype
|
633 |
-
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
|
634 |
|
635 |
def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
|
636 |
# init input tensors
|
@@ -755,17 +739,11 @@ class FlaxBartPreTrainedModel(FlaxPreTrainedModel):
|
|
755 |
decoder_attention_mask: Optional[jnp.ndarray] = None,
|
756 |
position_ids: Optional[jnp.ndarray] = None,
|
757 |
decoder_position_ids: Optional[jnp.ndarray] = None,
|
758 |
-
output_attentions: Optional[bool] = None,
|
759 |
-
output_hidden_states: Optional[bool] = None,
|
760 |
return_dict: Optional[bool] = None,
|
761 |
train: bool = False,
|
762 |
params: dict = None,
|
763 |
dropout_rng: PRNGKey = None,
|
764 |
):
|
765 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
766 |
-
output_hidden_states = (
|
767 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
768 |
-
)
|
769 |
return_dict = return_dict if return_dict is not None else self.config.return_dict
|
770 |
|
771 |
# prepare encoder inputs
|
@@ -817,7 +795,6 @@ class FlaxBartForConditionalGenerationModule(nn.Module):
|
|
817 |
dtype=self.dtype,
|
818 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
819 |
)
|
820 |
-
self.final_logits_bias = self.param("final_logits_bias", self.bias_init, (1, self.config.decoder_vocab_size))
|
821 |
|
822 |
def _get_encoder_module(self):
|
823 |
return self.model.encoder
|
@@ -853,8 +830,6 @@ class FlaxBartForConditionalGenerationModule(nn.Module):
|
|
853 |
else:
|
854 |
lm_logits = self.lm_head(hidden_states)
|
855 |
|
856 |
-
lm_logits += self.final_logits_bias
|
857 |
-
|
858 |
return FlaxSeq2SeqLMOutput(
|
859 |
logits=lm_logits,
|
860 |
decoder_hidden_states=outputs.decoder_hidden_states,
|
|
|
44 |
from transformers.utils import logging
|
45 |
|
46 |
|
47 |
+
from .configuration_bart import BartConfig
|
48 |
|
49 |
|
50 |
logger = logging.get_logger(__name__)
|
|
|
80 |
dense = partial(
|
81 |
nn.Dense,
|
82 |
self.embed_dim,
|
83 |
+
use_bias=False,
|
84 |
dtype=self.dtype,
|
85 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
86 |
)
|
|
|
242 |
self.fc1 = nn.Dense(
|
243 |
self.config.encoder_ffn_dim,
|
244 |
dtype=self.dtype,
|
245 |
+
use_bias=False,
|
246 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
247 |
)
|
248 |
self.fc2 = nn.Dense(
|
249 |
+
self.embed_dim,
|
250 |
+
dtype=self.dtype,
|
251 |
+
use_bias=False,
|
252 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
253 |
)
|
254 |
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
255 |
|
|
|
329 |
dropout=self.config.attention_dropout,
|
330 |
dtype=self.dtype,
|
331 |
)
|
332 |
+
self.encoder_attn_layer_norm = nn
|
333 |
self.fc1 = nn.Dense(
|
334 |
self.config.encoder_ffn_dim,
|
335 |
dtype=self.dtype,
|
336 |
+
use_bias=False,
|
337 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
338 |
)
|
339 |
self.fc2 = nn.Dense(
|
340 |
+
self.embed_dim,
|
341 |
+
dtype=self.dtype,
|
342 |
+
use_bias=False,
|
343 |
+
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
344 |
)
|
345 |
self.final_layer_norm = nn.LayerNorm(dtype=self.dtype)
|
346 |
|
|
|
422 |
class FlaxBartEncoder(nn.Module):
|
423 |
config: BartConfig
|
424 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
|
|
425 |
|
426 |
def setup(self):
|
427 |
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
|
|
431 |
self.max_source_positions = self.config.max_position_embeddings
|
432 |
self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
|
433 |
|
434 |
+
self.embed_tokens = nn.Embed(
|
435 |
+
self.config.vocab_size,
|
436 |
+
embed_dim,
|
437 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
438 |
+
)
|
|
|
439 |
|
440 |
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
441 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
442 |
+
self.offset = 0
|
443 |
self.embed_positions = nn.Embed(
|
444 |
self.config.max_position_embeddings + self.offset,
|
445 |
embed_dim,
|
|
|
478 |
class FlaxBartDecoder(nn.Module):
|
479 |
config: BartConfig
|
480 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
|
|
481 |
|
482 |
def setup(self):
|
483 |
self.dropout_layer = nn.Dropout(rate=self.config.dropout)
|
|
|
487 |
self.max_target_positions = self.config.max_position_embeddings
|
488 |
self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
|
489 |
|
490 |
+
self.embed_tokens = nn.Embed(
|
491 |
+
self.config.decoder_vocab_size,
|
492 |
+
embed_dim,
|
493 |
+
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
494 |
+
)
|
|
|
495 |
|
496 |
# Bart is set up so that if padding_idx is specified then offset the embedding ids by 2
|
497 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
498 |
+
self.offset = 0
|
499 |
self.embed_positions = nn.Embed(
|
500 |
+
self.config.decoder_max_position_embeddings + self.offset,
|
501 |
embed_dim,
|
502 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
503 |
)
|
|
|
550 |
dtype: jnp.dtype = jnp.float32 # the dtype of the computation
|
551 |
|
552 |
def setup(self):
|
553 |
+
self.encoder = FlaxBartEncoder(self.config, dtype=self.dtype)
|
554 |
+
self.decoder = FlaxBartDecoder(self.config, dtype=self.dtype)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
555 |
|
556 |
def _get_encoder_module(self):
|
557 |
return self.encoder
|
|
|
567 |
decoder_attention_mask,
|
568 |
position_ids,
|
569 |
decoder_position_ids,
|
|
|
|
|
570 |
return_dict: bool = True,
|
571 |
deterministic: bool = True,
|
572 |
):
|
|
|
574 |
input_ids=input_ids,
|
575 |
attention_mask=attention_mask,
|
576 |
position_ids=position_ids,
|
|
|
|
|
|
|
577 |
deterministic=deterministic,
|
578 |
)
|
579 |
|
|
|
583 |
position_ids=decoder_position_ids,
|
584 |
encoder_hidden_states=encoder_outputs[0],
|
585 |
encoder_attention_mask=attention_mask,
|
|
|
|
|
|
|
586 |
deterministic=deterministic,
|
587 |
)
|
588 |
|
|
|
613 |
dtype: jnp.dtype = jnp.float32,
|
614 |
**kwargs,
|
615 |
):
|
616 |
+
module = self.module_class(config=config, dtype=dtype)
|
617 |
+
super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, **kwargs)
|
618 |
|
619 |
def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
|
620 |
# init input tensors
|
|
|
739 |
decoder_attention_mask: Optional[jnp.ndarray] = None,
|
740 |
position_ids: Optional[jnp.ndarray] = None,
|
741 |
decoder_position_ids: Optional[jnp.ndarray] = None,
|
|
|
|
|
742 |
return_dict: Optional[bool] = None,
|
743 |
train: bool = False,
|
744 |
params: dict = None,
|
745 |
dropout_rng: PRNGKey = None,
|
746 |
):
|
|
|
|
|
|
|
|
|
747 |
return_dict = return_dict if return_dict is not None else self.config.return_dict
|
748 |
|
749 |
# prepare encoder inputs
|
|
|
795 |
dtype=self.dtype,
|
796 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
797 |
)
|
|
|
798 |
|
799 |
def _get_encoder_module(self):
|
800 |
return self.model.encoder
|
|
|
830 |
else:
|
831 |
lm_logits = self.lm_head(hidden_states)
|
832 |
|
|
|
|
|
833 |
return FlaxSeq2SeqLMOutput(
|
834 |
logits=lm_logits,
|
835 |
decoder_hidden_states=outputs.decoder_hidden_states,
|