add transformer-vae code

.gitignore

@@ -0,0 +1,3 @@
+.vscode
+venv
+*.pyc

README.md

@@ -0,0 +1,36 @@
+# Transformer-VAE (flax) (WIP)
+
+A Transformer-VAE made using flax.
+
+Done as part of Huggingface community training ([see forum post](https://discuss.huggingface.co/t/train-a-vae-to-interpolate-on-english-sentences/7548)).
+
+Builds on T5, using an autoencoder to convert it into a VAE.
+
+[See training logs.](https://wandb.ai/fraser/flax-vae)
+
+## ToDo
+
+- [ ] Basic training script working. (Fraser + Theo)
+- [ ] Add MMD loss (Theo)
+
+- [ ] Save a wikipedia sentences dataset to Huggingface (see original https://github.com/ChunyuanLI/Optimus/blob/master/data/download_datasets.md) (Mina)
+- [ ] Make a tokenizer using the OPTIMUS tokenized dataset.
+- [ ] Train on the OPTIMUS wikipedia sentences dataset.
+
+- [ ] Make Huggingface widget interpolating sentences! (???) https://github.com/huggingface/transformers/tree/master/examples/research_projects/jax-projects#how-to-build-a-demo
+
+Optional ToDos:
+
+- [ ] Add Funnel transformer encoder to FLAX (don't need weights).
+- [ ] Train a Funnel-encoder + T5-decoder transformer VAE.
+
+- [ ] Additional datasets:
+- [ ] Poetry (https://www.gwern.net/GPT-2#data-the-project-gutenberg-poetry-corpus)
+- [ ] 8-bit music (https://github.com/chrisdonahue/LakhNES)
+
+## Setup
+
+Follow all steps to install dependencies from https://cloud.google.com/tpu/docs/jax-quickstart-tpu-vm
+
+- [ ] Find dataset storage site.
+- [ ] Ask JAX team for dataset storage.

check_install.py

@@ -0,0 +1,15 @@
+from transformers import FlaxRobertaModel, RobertaTokenizerFast
+from datasets import load_dataset
+import jax
+
+dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
+
+dummy_input = next(iter(dataset))["text"]
+
+tokenizer = RobertaTokenizerFast.from_pretrained("roberta-base")
+input_ids = tokenizer(dummy_input, return_tensors="np").input_ids[:, :10]
+
+model = FlaxRobertaModel.from_pretrained("julien-c/dummy-unknown")
+
+# run a forward pass, should return an object `FlaxBaseModelOutputWithPooling`
+z = model(input_ids)

model/config.py

@@ -0,0 +1,137 @@
+import copy
+from transformers.utils import logging
+from transformers.configuration_utils import PretrainedConfig
+from transformers import AutoConfig, T5Config
+
+from model.encoders import VAE_ENCODER_MODELS
+from model.decoders import VAE_DECODER_MODELS
+from model.utils import assertEqual, assertIn
+
+logger = logging.get_logger(__name__)
+
+
+class T5VaeConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of :class:`FlaxT5VAE`.
+    It is used to instantiate a T5-VAE model according to the specified arguments, defining the model architecture.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the T5 `t5-vae-base architecture.
+
+    To be able to use `transformer.trainer.Trainer` we need some specific training logic & config in the model.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+    Arguments:
+        n_latent_tokens (:obj:`int`, `optional`, defaults to 6):
+            Number of latent tokens (must be less than seq length).
+        latent_token_size (:obj:`int`, `optional`, defaults to 32):
+            Number of dimensions to use for each latent token.
+        t5_name (:obj:`str`, `optional`, defaults to t5-base):
+            Name of the Transformer model to use as a decoder.
+        block_size (:obj:`int`, `optional`, defaults to 60):
+            NOTE: Every input sequence must be padded to be equal to this length.
+    """
+    model_type = "transformer_vae"
+    is_composition = True
+
+    def __init__(
+        self,
+        t5_model_name_or_path=None,
+        n_latent_tokens=6,  # set to -1 for full sequence
+        latent_token_size=32,
+        vae_encoder_model='',
+        vae_decoder_model='',
+        block_size=60,
+        decoder_start_token_id=0,
+        cache_dir=None,
+        tie_word_embeddings=True,
+        # T5 config
+        t5=dict(),
+        vocab_size=32128,
+        d_model=512,
+        d_kv=64,
+        d_ff=2048,
+        num_layers=6,
+        num_decoder_layers=None,
+        num_heads=8,
+        relative_attention_num_buckets=32,
+        dropout_rate=0.1,
+        layer_norm_epsilon=1e-6,
+        initializer_factor=1.0,
+        feed_forward_proj="relu",
+        is_encoder_decoder=True,
+        use_cache=True,
+        pad_token_id=0,
+        eos_token_id=1,
+        gradient_checkpointing=False,
+        # end
+        **kwargs,
+    ):
+        assertIn(vae_encoder_model, VAE_ENCODER_MODELS.keys(), "Unexpected VAE encoder.")
+        assertIn(vae_decoder_model, VAE_DECODER_MODELS.keys(), "Unexpected VAE decoder.")
+
+        super().__init__(**kwargs)
+
+        self.set_seq_size = block_size
+
+        # VAE
+        self.vae_encoder_model = vae_encoder_model
+        self.vae_decoder_model = vae_decoder_model
+
+        self.latent_token_size = latent_token_size
+        assert(n_latent_tokens <= self.set_seq_size, 'Cannot use more latent tokens than input tokens.')
+        self.n_latent_tokens = n_latent_tokens
+        self.use_cache = use_cache
+
+        # T5
+        if t5_model_name_or_path:
+            self.t5 = AutoConfig.from_pretrained(t5_model_name_or_path, cache_dir=cache_dir)
+            assertEqual(self.t5.model_type, "t5", "Need t5 model type for transformer_decoder.")
+            self.t5.decoder_start_token_id = decoder_start_token_id
+        elif t5:
+            # use for loading a config
+            self.t5 = T5Config(**t5)
+        else:
+            self.t5 = T5Config(
+                vocab_size=vocab_size,
+                d_model=d_model,
+                d_kv=d_kv,
+                d_ff=d_ff,
+                num_layers=num_layers,
+                num_decoder_layers=num_decoder_layers,
+                num_heads=num_heads,
+                relative_attention_num_buckets=relative_attention_num_buckets,
+                dropout_rate=dropout_rate,
+                layer_norm_epsilon=layer_norm_epsilon,
+                initializer_factor=initializer_factor,
+                feed_forward_proj=feed_forward_proj,
+                is_encoder_decoder=is_encoder_decoder,
+                use_cache=use_cache,
+                pad_token_id=pad_token_id,
+                eos_token_id=eos_token_id,
+                gradient_checkpointing=gradient_checkpointing,
+                **kwargs
+            )
+
+        if self.t5.d_model < self.latent_token_size:
+            raise Exception('Using larger latent token dimension then T5 hidden dimension.')
+
+        # Add t5 config options
+        self.tie_word_embeddings = tie_word_embeddings
+        self.t5.tie_word_embeddings = self.tie_word_embeddings
+        self.t5.use_cache = self.use_cache
+        self.pad_token_id = pad_token_id
+        self.eos_token_id = eos_token_id
+        self.decoder_start_token_id = self.t5.decoder_start_token_id
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default `to_dict()` from `PretrainedConfig`.
+
+        Returns:
+            :obj:`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["model_type"] = self.__class__.model_type
+        output['t5'] = self.t5.to_dict()
+        return output

model/decoders.py

@@ -0,0 +1,23 @@
+import logging
+import flax.linen as nn
+
+logger = logging.getLogger(__name__)
+
+
+class Decoder(nn.Module):
+    '''
+        Converts latent code -> transformer encoding.
+    '''
+    dim_model: int
+    n_latent_tokens: int
+
+    @nn.compact
+    def __call__(self, latent_code):  # (batch, latent_tokens_per_sequence, latent_token_dim)
+        raw_latent_tokens = nn.Dense(self.dim_model)(latent_code)
+        latent_tokens = nn.LayerNorm()(raw_latent_tokens)
+        return latent_tokens  # (batch, latent_tokens_per_sequence, dim_model)
+
+
+VAE_DECODER_MODELS = {
+    '': Decoder,
+}

model/encoders.py

@@ -0,0 +1,26 @@
+import logging
+import jax.numpy as jnp
+import flax.linen as nn
+
+logger = logging.getLogger(__name__)
+
+
+class Encoder(nn.Module):
+    '''
+        Converts N hidden tokens into N seperate latent codes.
+    '''
+    latent_token_size: int
+    n_latent_tokens: int
+
+    @nn.compact
+    def __call__(self, encoding):
+        latent_tokens = nn.Dense(self.latent_token_size)(encoding)
+        raw_latent_code = latent_tokens[:, : self.n_latent_tokens, :]
+        # TODO does this just apply tanh to each latent token? Or across the whole batch
+        latent_code = jnp.tanh(raw_latent_code)
+        return latent_code  # (batch, latent_tokens_per_sequence, latent_token_dim)
+
+
+VAE_ENCODER_MODELS = {
+    '': Encoder,
+}

model/outputs.py

@@ -0,0 +1,74 @@
+from typing import Optional, Tuple
+
+import flax
+import jaxlib.xla_extension as jax_xla
+
+from transformers.file_utils import ModelOutput
+
+
+@flax.struct.dataclass
+class TransformerVaeOutput(ModelOutput):
+    """
+    Base class for a Transformer-VAE's outputs.
+
+    Args:
+        latent_codes (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, n_latent_tokens, latent_token_size)`):
+            Latent codes representing encoded sequences.
+        remade_encoder_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, n_tokens, model_dim)`):
+            Reconstructed encoder hidden states representing sequences.
+
+    (std Seq2Seq) Args:
+        logits (:obj:`jax_xla.DeviceArray` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`tuple(tuple(jax_xla.DeviceArray))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(jax_xla.DeviceArray)` of length :obj:`config.n_layers`, with each tuple having 2
+            tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional
+            tensors of shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        last_hidden_state (:obj:`tuple(jax_xla.DeviceArray)`:
+            Last model hidden state.
+        decoder_hidden_states (:obj:`tuple(jax_xla.DeviceArray)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`jax_xla.DeviceArray` (one for the output of the embeddings + one for the output of each
+            layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(jax_xla.DeviceArray)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`jax_xla.DeviceArray` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(jax_xla.DeviceArray)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`jax_xla.DeviceArray` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`jax_xla.DeviceArray` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(jax_xla.DeviceArray)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`jax_xla.DeviceArray` (one for the output of the embeddings + one for the output of each
+            layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(jax_xla.DeviceArray)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`jax_xla.DeviceArray` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+    logits: jax_xla.DeviceArray = None
+    latent_codes: jax_xla.DeviceArray = None
+    remade_encoder_hidden_state: jax_xla.DeviceArray = None
+    # seq2seq
+    past_key_values: Optional[Tuple[Tuple[jax_xla.DeviceArray]]] = None
+    decoder_hidden_states: Optional[Tuple[jax_xla.DeviceArray]] = None
+    decoder_attentions: Optional[Tuple[jax_xla.DeviceArray]] = None
+    cross_attentions: Optional[Tuple[jax_xla.DeviceArray]] = None
+    last_hidden_state: Optional[jax_xla.DeviceArray] = None
+    encoder_last_hidden_state: Optional[jax_xla.DeviceArray] = None
+    encoder_hidden_states: Optional[Tuple[jax_xla.DeviceArray]] = None
+    encoder_attentions: Optional[Tuple[jax_xla.DeviceArray]] = None

model/t5_vae.py

@@ -0,0 +1,522 @@
+from typing import Optional, Tuple
+
+import jax
+import jax.numpy as jnp
+from jax.random import PRNGKey
+import flax.linen as nn
+from flax.core.frozen_dict import FrozenDict, unfreeze
+
+from transformers.modeling_flax_outputs import FlaxCausalLMOutputWithCrossAttentions
+from transformers.file_utils import add_start_docstrings
+from transformers.modeling_flax_utils import FlaxPreTrainedModel
+from transformers.models.t5.modeling_flax_t5 import FlaxT5ForConditionalGenerationModule
+
+from model.vae import VAE
+from model.outputs import TransformerVaeOutput
+from model.config import T5VaeConfig
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top converted into a VAE.""")
+class FlaxT5VaeForAutoencodingModule(nn.Module):
+    config: T5VaeConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def _get_encoder_module(self):
+        return self.t5.encoder
+
+    def _get_vae_encoder_module(self):
+        return self.vae.encoder
+
+    def _get_vae_decoder_module(self):
+        return self.vae.decoder
+
+    def _get_decoder_module(self):
+        return self.t5.decoder
+
+    def setup(self):
+        self.t5 = FlaxT5ForConditionalGenerationModule(self.config.t5)
+        self.vae = VAE(self.config)
+
+    def __call__(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        latent_codes=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        deterministic: bool = True,
+    ):
+        """
+            Adapted from `FlaxT5ForConditionalGenerationModule`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # Encode
+        encoder_outputs = self.t5.encoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        hidden_states = encoder_outputs[0]
+
+        # Autoencode
+        hidden_states, latent_codes = self.vae(hidden_states, latent_codes)
+        encoder_attention_mask = jnp.ones((hidden_states.shape[0], hidden_states.shape[1]))
+
+        # Decode
+        decoder_outputs = self.t5.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        sequence_output = decoder_outputs[0]
+
+        if self.t5.config.tie_word_embeddings:
+            # Rescale output before projecting on vocab
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+            sequence_output = sequence_output * (self.t5.config.d_model ** -0.5)
+
+        if self.t5.config.tie_word_embeddings:
+            shared_embedding = self.t5.shared.variables["params"]["embedding"]
+            lm_logits = self.t5.lm_head.apply({"params": {"kernel": shared_embedding.T}}, sequence_output)
+        else:
+            lm_logits = self.t5.lm_head(sequence_output)
+
+        if not return_dict:
+            return [lm_logits, latent_codes] + decoder_outputs[1:] + encoder_outputs
+
+        return TransformerVaeOutput(
+            logits=lm_logits,
+            latent_codes=latent_codes,
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+class FlaxT5VaePreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = T5VaeConfig
+    base_model_prefix = "transformer"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: T5VaeConfig,
+        input_shape: Tuple[int] = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+
+        attention_mask = jnp.ones_like(input_ids)
+        decoder_input_ids = jnp.ones_like(input_ids)
+        decoder_attention_mask = jnp.ones_like(input_ids)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        return self.module.init(
+            rngs,
+            input_ids,
+            attention_mask,
+            decoder_input_ids,
+            decoder_attention_mask,
+        )["params"]
+
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        decoder_input_ids: jnp.ndarray = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if decoder_input_ids is None:
+            raise ValueError(
+                "Make sure to provide both `input_ids` and `decoder_input_ids`. `decoder_input_ids` is not passed here."
+            )
+
+        # prepare encoder inputs
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # prepare decoder inputs
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+    def init_cache(self, batch_size, max_length, latent_codes):
+        r"""
+        Args:
+            batch_size (:obj:`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (:obj:`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+            latent_codes (:obj:`Union[FlaxBaseModelOutput, tuple(tuple(jnp.ndarray)]`):
+                ``latent_codes`` consists of compressed hidden-states at the output of the last layer of the encoder.
+                Used in the cross-attention of the decoder.
+        """
+        # init input variables to retrieve cache
+        decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        def _decoder_forward(module, decoder_input_ids, latent_codes, decoder_attention_mask, **kwargs):
+            vae_decoder_module = module._get_vae_decoder_module()
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                encoder_hidden_states=vae_decoder_module(latent_codes),
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            decoder_input_ids=decoder_input_ids,
+            latent_codes=latent_codes,
+            decoder_attention_mask=decoder_attention_mask,
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    def encode(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        raise NotImplementedError()
+
+    def decode(
+        self,
+        decoder_input_ids,
+        latent_codes,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        raise NotImplementedError()
+
+
+class FlaxT5VaeForAutoencoding(FlaxT5VaePreTrainedModel):
+    module_class = FlaxT5VaeForAutoencodingModule
+
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        '''
+            Adapted from `FlaxT5PreTrainedModel`
+        '''
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if decoder_input_ids is None:
+            raise ValueError(
+                "Make sure to provide both `input_ids` and `decoder_input_ids`. `decoder_input_ids` is not passed here."
+            )
+
+        # prepare encoder inputs
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # prepare decoder inputs
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+    def encode(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, input_ids, attention_mask, **kwargs):
+            encode_module = module._get_encoder_module()
+            vae_encoder_module = module._get_vae_encoder_module()
+            return vae_encoder_module(encode_module(input_ids, attention_mask, **kwargs)[0])
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    def decode(
+        self,
+        decoder_input_ids,
+        latent_codes,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example::
+
+            >>> model = FlaxT5VaeForAutoencoding.from_pretrained('t5-small')
+            >>> tokenizer = T5Tokenizer.from_pretrained('t5-small')
+
+            >>> text = "My friends are cool but they eat too many carbs."
+            >>> inputs = tokenizer(text, max_length=512, return_tensors='jax')
+            >>> latent_codes = model.encode(**inputs)
+
+            >>> decoder_start_token_id = model.config.decoder_start_token_id
+            >>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+            >>> outputs = model.decode(decoder_input_ids, latent_codes)
+            >>> last_decoder_hidden_states = outputs.last_hidden_state
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = latent_codes.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxT5Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, latent_codes, decoder_attention_mask, **kwargs):
+            vae_decoder_module = module._get_vae_decoder_module()
+            decoder_module = module._get_decoder_module()
+            decoder_outputs = decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                encoder_hidden_states=vae_decoder_module(latent_codes),
+                **kwargs,
+            )
+            sequence_output = decoder_outputs[0]
+
+            if self.config.tie_word_embeddings:
+                # Rescale output before projecting on vocab
+                # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+                sequence_output = sequence_output * (self.config.d_model ** -0.5)
+
+            if self.config.tie_word_embeddings:
+                shared_embedding = module.t5.shared.variables["params"]["embedding"]
+                lm_logits = module.t5.lm_head.apply({"params": {"kernel": shared_embedding.T}}, sequence_output)
+            else:
+                lm_logits = module.t5.lm_head(sequence_output)
+
+            return lm_logits, decoder_outputs
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            latent_codes=latent_codes,
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, decoder_outputs = outputs
+        else:
+            (lm_logits, decoder_outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=decoder_outputs.hidden_states,
+                attentions=decoder_outputs.attentions,
+                cross_attentions=decoder_outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + decoder_outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        max_length,
+        attention_mask: Optional[jnp.DeviceArray] = None,
+        decoder_attention_mask: Optional[jnp.DeviceArray] = None,
+        latent_codes=None,
+        **kwargs
+    ):
+        # initializing the cache
+        batch_size, seq_length = decoder_input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, latent_codes)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if decoder_attention_mask is not None:
+            extended_attention_mask = jax.lax.dynamic_update_slice(
+                extended_attention_mask, decoder_attention_mask, (0, 0)
+            )
+
+        return {
+            "past_key_values": past_key_values,
+            "latent_codes": latent_codes,
+            "encoder_attention_mask": attention_mask,
+            "decoder_attention_mask": extended_attention_mask,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        return model_kwargs

model/utils.py

@@ -0,0 +1,24 @@
+from typing import Sequence
+
+import flax.linen as nn
+
+
+class MLP(nn.Module):
+    features: Sequence[int]
+
+    @nn.compact
+    def __call__(self, x):
+        for feat in self.features[:-1]:
+            x = nn.relu(nn.Dense(feat)(x))
+        x = nn.Dense(self.features[-1])(x)
+        return x
+
+
+def assertEqual(actual, expected, msg, first="Got", second="Expected"):
+    if actual != expected:
+        raise ValueError(msg + f' {first}: "{actual}" {second}: "{expected}"')
+
+
+def assertIn(actual, expected, msg, first="Got", second="Expected one of"):
+    if actual not in expected:
+        raise ValueError(msg + f' {first}: "{actual}" {second}: {expected}')

model/vae.py

@@ -0,0 +1,30 @@
+import jax.numpy as jnp
+import flax.linen as nn
+
+from model.encoders import VAE_ENCODER_MODELS
+from model.decoders import VAE_DECODER_MODELS
+from model.config import T5VaeConfig
+
+
+class VAE(nn.Module):
+    # see https://github.com/google/flax#what-does-flax-look-like
+    """
+        An MMD-VAE used with encoder-decoder models.
+        Encodes all token encodings into a single latent & spits them back out.
+    """
+    config: T5VaeConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.encoder = VAE_ENCODER_MODELS[self.config.vae_encoder_model](self.config.latent_token_size, self.config.n_latent_tokens)
+        self.decoder = VAE_DECODER_MODELS[self.config.vae_decoder_model](self.config.t5.d_model,  self.config.n_latent_tokens)
+
+    def __call__(self, encoding=None, latent_codes=None):
+        latent_codes = self.encode(encoding)
+        return self.decode(latent_codes), latent_codes
+
+    def encode(self, encoding):
+        return self.encoder(encoding)
+
+    def decode(self, latent):
+        return self.decoder(latent)

requirements.txt

@@ -0,0 +1,3 @@
+jax
+jaxlib
+-r requirements-tpu.txt

train.py

@@ -0,0 +1,706 @@
+'''
+    Pre-training/Fine-tuning seq2seq models on autoencoding a dataset.
+
+    TODO:
+    - [x] Get this running.
+    - [x] Don't make decoder input ids.
+    - [ ] Add reg loss
+        - [x] calculate MMD loss
+        - [ ] schedule MMD loss weight
+            - [ ] Add these params to the training arguments.
+
+                reg_schedule_k (:obj:`float`, `optional`, defaults to 0.0025):
+                    Multiplied by global_step in a sigmoid, more gradually increase regulariser loss weight.
+                reg_schedule_b (:obj:`float`, `optional`, defaults to 6.25):
+                    Added to global step in sigmoid, further delays increase in regulariser loss weight.
+                use_extra_logs (:obj:`bool`, `optional`, defaults to False):
+                    Store extra logs during each training inference.
+
+            - [ ] Send the schedule time to the compute_loss method and calculate a coefficient based on that.
+'''
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    AutoTokenizer,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.models.t5.modeling_flax_t5 import shift_tokens_right
+from transformers.testing_utils import CaptureLogger
+
+from model.t5_vae import FlaxT5VaeForAutoencoding
+from model.config import T5VaeConfig
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    t5_model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The T5 model checkpoint for weights initialization."
+            "Needed when not starting from a T5-VAE model."
+        },
+    )
+    n_latent_tokens: Optional[int] = field(
+        default=6,
+        metadata={
+            "help": "Number of latent tokens (must be less than seq length)."
+        },
+    )
+    latent_token_size: Optional[int] = field(
+        default=32,
+        metadata={
+            "help": "Number of dimensions to use for each latent token."
+        },
+    )
+    config_path: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config path"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+            "The training dataset will be truncated in block of this size for training. "
+            "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    streaming: bool = field(
+        default=False, metadata={"help": "Stream the dataset."}
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    if data_args.block_size is None:
+        raise Exception('Must set block_size so we know what length of sequence to autoencode.')
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    #  Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, streaming=data_args.streaming, keep_in_memory=False
+        )
+
+        if "validation" not in dataset.keys():
+            dataset["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            dataset["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+
+    if model_args.config_path:
+        config = T5VaeConfig.from_pretrained(
+            model_args.config_path, cache_dir=model_args.cache_dir
+        )
+    elif model_args.model_name_or_path:
+        config = T5VaeConfig.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir
+        )
+    else:
+        config = T5VaeConfig(**model_args.__dict__)
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.t5_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.t5_model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxT5VaeForAutoencoding.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+        # TODO assert token embedding size == len(tokenizer)
+        assert(model.params['t5']['shared'].shape[0] == len(tokenizer), "T5 Tokenizer doesn't match T5Vae embedding size.")
+    else:
+        vocab_size = len(tokenizer)
+        config.t5.vocab_size = vocab_size
+        config.vocab_size = vocab_size
+        logger.info("Training new model from scratch.")
+        model = FlaxT5VaeForAutoencoding(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    if model_args.add_special_tokens:
+        special_tokens_dict = {'pad_token': '<PAD>', 'bos_token': '<BOS>', 'eos_token': '<EOS>'}
+        num_added_tokens = tokenizer.add_special_tokens(special_tokens_dict)
+        print('We have added', num_added_tokens, 'tokens to GPT2')
+        model.resize_token_embeddings(len(tokenizer))
+        assert tokenizer.pad_token == '<PAD>'
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    # remove dataset tasks
+    for k in dataset.keys():
+        dataset[k].info.task_templates = []
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=column_names,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if data_args.block_size > tokenizer.model_max_length:
+        logger.warning(
+            f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+            f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+        )
+    block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    pad_token_id, start_token_id = tokenizer.pad_token_id, config.decoder_start_token_id
+
+    def clip_texts(examples):
+        examples["labels"] = examples["input_ids"].copy()
+
+        for i, input_ids in enumerate(examples["input_ids"]):
+            if len(input_ids) > block_size:
+                for k in examples.keys():
+                    examples[k][i] = examples[k][i][:block_size]
+            elif len(input_ids) < block_size:
+                delta = block_size - len(input_ids)
+                examples['input_ids'][i] = examples['input_ids'][i] + [pad_token_id] * delta
+                examples['attention_mask'][i] = examples['attention_mask'][i] + [0] * delta
+                examples['labels'][i] = examples['labels'][i] + [-100] * delta
+
+        return examples
+
+    logger.info('clip_texts...')
+    clipped_lm_datasets = tokenized_datasets.map(
+        clip_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    def add_decoder_input_ids(examples):
+        arr_input_ids = jnp.array(examples["input_ids"])
+        pad = pad_token_id * jnp.ones((arr_input_ids.shape[0], 1), dtype=jnp.int32)
+        arr_pad_input_ids = jnp.concatenate((arr_input_ids, pad), axis=1)
+        examples['decoder_input_ids'] = shift_tokens_right(arr_pad_input_ids, pad_token_id, start_token_id)
+
+        arr_attention_mask = jnp.array(examples['attention_mask'])
+        ones = jnp.ones((arr_attention_mask.shape[0], 1), dtype=jnp.int32)
+        examples['decoder_attention_mask'] = jnp.concatenate((ones, arr_attention_mask), axis=1)
+
+        for k in ['decoder_input_ids', 'decoder_attention_mask']:
+            examples[k] = examples[k].tolist()
+
+        return examples
+
+    logger.info('add_decoder_input_ids...')
+    lm_datasets = clipped_lm_datasets.map(
+        add_decoder_input_ids,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
+
+    def compute_kernel(x, y):
+        x_size = x.shape[0]
+        y_size = y.shape[0]
+        dim = x.shape[1]
+        tiled_x = jnp.repeat(jnp.reshape(x, (x_size, 1, dim)), y_size, axis=1)
+        tiled_y = jnp.repeat(jnp.reshape(y, (1, y_size, dim)), x_size, axis=0)
+        return jnp.exp(-jnp.mean((tiled_x - tiled_y) ** 2, axis=2) / dim * 1.0)
+
+    def compute_mmd(x, y):
+        x_kernel = compute_kernel(x, x)
+        y_kernel = compute_kernel(y, y)
+        xy_kernel = compute_kernel(x, y)
+        return jnp.mean(x_kernel) + jnp.mean(y_kernel) - 2 * jnp.mean(xy_kernel)
+
+    def regulariser_loss(latent_codes, rng):
+        true_samples = jax.random.normal(rng, latent_codes.shape)
+        # return jax.vmap(compute_mmd)(true_samples, latent_codes)
+        return compute_mmd(true_samples, latent_codes)
+
+    def loss_fn(logits, labels, latent_codes, regulariser_rng):
+        shift_logits = logits[..., :-1, :]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(labels, logits.shape[-1]))
+        reg_loss = regulariser_loss(latent_codes.reshape(-1, latent_codes.shape[-1]), regulariser_rng)
+        return loss.mean() + reg_loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+        new_dropout_rng, regulariser_rng = jax.random.split(new_dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            outputs = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)
+            loss = loss_fn(outputs[0], labels, outputs[1], regulariser_rng)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, rng, batch):
+        labels = batch.pop("labels")
+        logits, latent_codes = model(**batch, params=params, train=False)[:2]
+        loss = loss_fn(logits, labels, latent_codes, rng)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (len(train_dataset) // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                eval_metrics = []
+                eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+                eval_steps = len(eval_dataset) // eval_batch_size
+                for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+                    # Model forward
+                    batch = next(eval_loader)
+                    metrics = p_eval_step(state.params, state.dropout_rng, batch)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                try:
+                    eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+                except OverflowError:
+                    eval_metrics["perplexity"] = float("inf")
+
+                # Print metrics and update progress bar
+                desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+                epochs.write(desc)
+                epochs.desc = desc
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    cur_step = epoch * (len(train_dataset) // train_batch_size)
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(unreplicate(state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )
+
+
+if __name__ == "__main__":
+    main()