Spaces:
Running
Running
fix(model): use correct params
Browse files
dalle_mini/configuration_bart.py
CHANGED
@@ -21,16 +21,11 @@ from transformers.utils import logging
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
24 |
-
BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
25 |
-
"facebook/bart-large": "https://huggingface.co/facebook/bart-large/resolve/main/config.json",
|
26 |
-
# See all BART models at https://huggingface.co/models?filter=bart
|
27 |
-
}
|
28 |
|
29 |
-
|
30 |
-
class BartConfig(PretrainedConfig):
|
31 |
r"""
|
32 |
-
This is the configuration class to store the configuration of a
|
33 |
-
instantiate a
|
34 |
configuration with the defaults will yield a similar configuration to that of the BART `facebook/bart-large
|
35 |
<https://huggingface.co/facebook/bart-large>`__ architecture.
|
36 |
|
@@ -39,7 +34,7 @@ class BartConfig(PretrainedConfig):
|
|
39 |
|
40 |
|
41 |
Args:
|
42 |
-
|
43 |
Vocabulary size of the BART model. Defines the number of different tokens that can be represented by the
|
44 |
:obj:`inputs_ids` passed when calling :class:`~transformers.BartModel` or
|
45 |
:class:`~transformers.TFBartModel`.
|
@@ -90,30 +85,18 @@ class BartConfig(PretrainedConfig):
|
|
90 |
forced_eos_token_id (:obj:`int`, `optional`, defaults to 2):
|
91 |
The id of the token to force as the last generated token when :obj:`max_length` is reached. Usually set to
|
92 |
:obj:`eos_token_id`.
|
93 |
-
|
94 |
-
Example::
|
95 |
-
|
96 |
-
>>> from transformers import BartModel, BartConfig
|
97 |
-
|
98 |
-
>>> # Initializing a BART facebook/bart-large style configuration
|
99 |
-
>>> configuration = BartConfig()
|
100 |
-
|
101 |
-
>>> # Initializing a model from the facebook/bart-large style configuration
|
102 |
-
>>> model = BartModel(configuration)
|
103 |
-
|
104 |
-
>>> # Accessing the model configuration
|
105 |
-
>>> configuration = model.config
|
106 |
"""
|
107 |
-
model_type = "
|
108 |
keys_to_ignore_at_inference = ["past_key_values"]
|
109 |
attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
|
110 |
|
111 |
def __init__(
|
112 |
self,
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
|
|
117 |
encoder_layers=12,
|
118 |
encoder_ffn_dim=4096,
|
119 |
encoder_attention_heads=16,
|
@@ -133,19 +116,16 @@ class BartConfig(PretrainedConfig):
|
|
133 |
gradient_checkpointing=False,
|
134 |
use_cache=True,
|
135 |
num_labels=3,
|
136 |
-
pad_token_id=1,
|
137 |
-
bos_token_id=0,
|
138 |
-
eos_token_id=2,
|
139 |
is_encoder_decoder=True,
|
140 |
-
|
141 |
-
|
142 |
-
tie_word_embeddings=False, # don't tie for scaling reasons
|
143 |
**kwargs,
|
144 |
):
|
145 |
-
self.
|
146 |
-
self.
|
147 |
-
self.
|
148 |
-
self.
|
|
|
149 |
self.d_model = d_model
|
150 |
self.encoder_ffn_dim = encoder_ffn_dim
|
151 |
self.encoder_layers = encoder_layers
|
@@ -165,12 +145,15 @@ class BartConfig(PretrainedConfig):
|
|
165 |
self.num_hidden_layers = encoder_layers
|
166 |
self.gradient_checkpointing = gradient_checkpointing
|
167 |
self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
|
|
|
|
|
|
|
168 |
|
169 |
super().__init__(
|
170 |
num_labels=num_labels,
|
171 |
-
pad_token_id=
|
172 |
-
bos_token_id=
|
173 |
-
eos_token_id=
|
174 |
is_encoder_decoder=is_encoder_decoder,
|
175 |
decoder_start_token_id=decoder_start_token_id,
|
176 |
forced_eos_token_id=forced_eos_token_id,
|
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
|
|
|
|
|
|
|
|
24 |
|
25 |
+
class DalleBartConfig(PretrainedConfig):
|
|
|
26 |
r"""
|
27 |
+
This is the configuration class to store the configuration of a `DalleBartModel`. It is used to
|
28 |
+
instantiate a DalleBart model according to the specified arguments, defining the model architecture. Instantiating a
|
29 |
configuration with the defaults will yield a similar configuration to that of the BART `facebook/bart-large
|
30 |
<https://huggingface.co/facebook/bart-large>`__ architecture.
|
31 |
|
|
|
34 |
|
35 |
|
36 |
Args:
|
37 |
+
encoder_vocab_size (:obj:`int`, `optional`, defaults to 50265):
|
38 |
Vocabulary size of the BART model. Defines the number of different tokens that can be represented by the
|
39 |
:obj:`inputs_ids` passed when calling :class:`~transformers.BartModel` or
|
40 |
:class:`~transformers.TFBartModel`.
|
|
|
85 |
forced_eos_token_id (:obj:`int`, `optional`, defaults to 2):
|
86 |
The id of the token to force as the last generated token when :obj:`max_length` is reached. Usually set to
|
87 |
:obj:`eos_token_id`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
"""
|
89 |
+
model_type = "dallebart"
|
90 |
keys_to_ignore_at_inference = ["past_key_values"]
|
91 |
attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
|
92 |
|
93 |
def __init__(
|
94 |
self,
|
95 |
+
normalize_text=False,
|
96 |
+
encoder_vocab_size=50264,
|
97 |
+
image_vocab_size=16384, # encoded image token space
|
98 |
+
image_length=256, # number of encoded tokens
|
99 |
+
max_text_length=64, # max number of text tokens
|
100 |
encoder_layers=12,
|
101 |
encoder_ffn_dim=4096,
|
102 |
encoder_attention_heads=16,
|
|
|
116 |
gradient_checkpointing=False,
|
117 |
use_cache=True,
|
118 |
num_labels=3,
|
|
|
|
|
|
|
119 |
is_encoder_decoder=True,
|
120 |
+
forced_eos_token_id=None,
|
121 |
+
tie_word_embeddings=False, # don't tie for scaling reasons and due to different modalities and sizes
|
|
|
122 |
**kwargs,
|
123 |
):
|
124 |
+
self.normalize_text = normalize_text
|
125 |
+
self.encoder_vocab_size = encoder_vocab_size
|
126 |
+
self.decoder_vocab_size = image_vocab_size
|
127 |
+
self.image_length = image_length
|
128 |
+
self.max_text_length = max_text_length
|
129 |
self.d_model = d_model
|
130 |
self.encoder_ffn_dim = encoder_ffn_dim
|
131 |
self.encoder_layers = encoder_layers
|
|
|
145 |
self.num_hidden_layers = encoder_layers
|
146 |
self.gradient_checkpointing = gradient_checkpointing
|
147 |
self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
|
148 |
+
self.decoder_start_token_id = image_vocab_size, # BOS appended to vocab
|
149 |
+
self.min_length = image_length + 1
|
150 |
+
self.max_length = image_length + 1
|
151 |
|
152 |
super().__init__(
|
153 |
num_labels=num_labels,
|
154 |
+
pad_token_id=image_vocab_size + 1, # needed to avoid errors during generation (converted to jnp.array)
|
155 |
+
bos_token_id=image_vocab_size + 1, # set to unreachable values
|
156 |
+
eos_token_id=image_vocab_size + 1,
|
157 |
is_encoder_decoder=is_encoder_decoder,
|
158 |
decoder_start_token_id=decoder_start_token_id,
|
159 |
forced_eos_token_id=forced_eos_token_id,
|
dalle_mini/modeling_bart_flax.py
CHANGED
@@ -93,7 +93,7 @@ class FlaxBartAttention(nn.Module):
|
|
93 |
|
94 |
if self.causal:
|
95 |
self.causal_mask = make_causal_mask(
|
96 |
-
jnp.ones((1,
|
97 |
)
|
98 |
|
99 |
def _split_heads(self, hidden_states):
|
@@ -431,11 +431,10 @@ class FlaxBartEncoder(nn.Module):
|
|
431 |
|
432 |
embed_dim = self.config.d_model
|
433 |
self.padding_idx = self.config.pad_token_id
|
434 |
-
self.max_source_positions = self.config.max_position_embeddings
|
435 |
self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
|
436 |
|
437 |
self.embed_tokens = nn.Embed(
|
438 |
-
self.config.
|
439 |
embed_dim,
|
440 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
441 |
)
|
@@ -444,7 +443,7 @@ class FlaxBartEncoder(nn.Module):
|
|
444 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
445 |
self.offset = 0
|
446 |
self.embed_positions = nn.Embed(
|
447 |
-
self.config.
|
448 |
embed_dim,
|
449 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
450 |
)
|
@@ -489,11 +488,10 @@ class FlaxBartDecoder(nn.Module):
|
|
489 |
|
490 |
embed_dim = self.config.d_model
|
491 |
self.padding_idx = self.config.pad_token_id
|
492 |
-
self.max_target_positions = self.config.max_position_embeddings
|
493 |
self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
|
494 |
|
495 |
self.embed_tokens = nn.Embed(
|
496 |
-
self.config.
|
497 |
embed_dim,
|
498 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
499 |
)
|
@@ -502,7 +500,7 @@ class FlaxBartDecoder(nn.Module):
|
|
502 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
503 |
self.offset = 0
|
504 |
self.embed_positions = nn.Embed(
|
505 |
-
self.config.
|
506 |
embed_dim,
|
507 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
508 |
)
|
@@ -802,11 +800,14 @@ class FlaxBartForConditionalGenerationModule(nn.Module):
|
|
802 |
def setup(self):
|
803 |
self.model = FlaxBartModule(config=self.config, dtype=self.dtype)
|
804 |
self.lm_head = nn.Dense(
|
805 |
-
self.config.
|
806 |
use_bias=False,
|
807 |
dtype=self.dtype,
|
808 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
809 |
)
|
|
|
|
|
|
|
810 |
|
811 |
def _get_encoder_module(self):
|
812 |
return self.model.encoder
|
|
|
93 |
|
94 |
if self.causal:
|
95 |
self.causal_mask = make_causal_mask(
|
96 |
+
jnp.ones((1, embed_dim), dtype="bool"), dtype="bool"
|
97 |
)
|
98 |
|
99 |
def _split_heads(self, hidden_states):
|
|
|
431 |
|
432 |
embed_dim = self.config.d_model
|
433 |
self.padding_idx = self.config.pad_token_id
|
|
|
434 |
self.embed_scale = math.sqrt(embed_dim) if self.config.scale_embedding else 1.0
|
435 |
|
436 |
self.embed_tokens = nn.Embed(
|
437 |
+
self.config.encoder_vocab_size,
|
438 |
embed_dim,
|
439 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
440 |
)
|
|
|
443 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
444 |
self.offset = 0
|
445 |
self.embed_positions = nn.Embed(
|
446 |
+
self.config.max_text_length + self.offset,
|
447 |
embed_dim,
|
448 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
449 |
)
|
|
|
488 |
|
489 |
embed_dim = self.config.d_model
|
490 |
self.padding_idx = self.config.pad_token_id
|
|
|
491 |
self.embed_scale = math.sqrt(self.config.d_model) if self.config.scale_embedding else 1.0
|
492 |
|
493 |
self.embed_tokens = nn.Embed(
|
494 |
+
self.config.image_vocab_size + 1, # image vocab size + 1 for BOS
|
495 |
embed_dim,
|
496 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
497 |
)
|
|
|
500 |
# and adjust num_embeddings appropriately. Other models don't have this hack
|
501 |
self.offset = 0
|
502 |
self.embed_positions = nn.Embed(
|
503 |
+
self.config.image_length + 1 + self.offset, # image length + 1 for BOS
|
504 |
embed_dim,
|
505 |
embedding_init=jax.nn.initializers.normal(self.config.init_std),
|
506 |
)
|
|
|
800 |
def setup(self):
|
801 |
self.model = FlaxBartModule(config=self.config, dtype=self.dtype)
|
802 |
self.lm_head = nn.Dense(
|
803 |
+
self.config.image_vocab_size + 1, # image vocab size + 1 for BOS
|
804 |
use_bias=False,
|
805 |
dtype=self.dtype,
|
806 |
kernel_init=jax.nn.initializers.normal(self.config.init_std),
|
807 |
)
|
808 |
+
self.final_logits_bias = self.param(
|
809 |
+
"final_logits_bias", self.bias_init, (1, self.config.image_vocab_size + 1)
|
810 |
+
)
|
811 |
|
812 |
def _get_encoder_module(self):
|
813 |
return self.model.encoder
|