add model

config.json

@@ -0,0 +1,36 @@
+{
+  "architectures": [
+    "CharmenElectraModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "config.CharmenElectraConfig",
+    "AutoModel": "modeling_charmen.CharmenElectraModel"
+  },
+  "classifier_dropout": null,
+  "downsampling_factor": 4,
+  "embedding_size": 768,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 2048,
+  "layer_norm_eps": 1e-12,
+  "max_block_size": 4,
+  "max_position_embeddings": 1024,
+  "model_type": "SzegedAI/charmen-electra",
+  "num_attention_heads": 8,
+  "num_hidden_layers": 12,
+  "position_embedding_type": "absolute",
+  "sampling": "fp32_gumbel",
+  "score_consensus_attn": true,
+  "summary_activation": "gelu",
+  "summary_last_dropout": 0.1,
+  "summary_type": "first",
+  "summary_use_proj": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.21.0",
+  "type_vocab_size": 2,
+  "upsample_output": true,
+  "vocab_size": 261
+}

config.py

@@ -0,0 +1,68 @@
+from transformers import PretrainedConfig
+from typing import List, Literal, Optional
+
+
+_SAMPLING_TYPE = Literal['fp32_gumbel', 'fp16_gumbel', 'multinomial']
+
+
+class CharmenElectraConfig(PretrainedConfig):
+    model_type = "SzegedAI/charmen-electra"
+    _name_or_path = "SzegedAI/charmen-electra"
+
+    def __init__(
+        self,
+        downsampling_factor: int = 4,
+        max_block_size: int = 4,
+        score_consensus_attn: bool = True,
+        upsample_output: bool = True,
+        sampling: _SAMPLING_TYPE = 'fp32_gumbel',
+        attention_probs_dropout_prob: float = 0.1,
+        embedding_size: int = 768,
+        hidden_act: str = "gelu",
+        hidden_dropout_prob: float = 0.1,
+        hidden_size: int = 512,
+        initializer_range: float = 0.02,
+        intermediate_size: int = 2048,
+        layer_norm_eps: float = 1e-12,
+        max_position_embeddings: int = 1024,
+        model_type: str = "electra",
+        num_attention_heads: int = 8,
+        num_hidden_layers: int = 12,
+        pad_token_id: int = 0,
+        position_embedding_type: str = "absolute",
+        summary_activation: str = "gelu",
+        summary_last_dropout: float = 0.1,
+        summary_type: str = "first",
+        summary_use_proj: bool = True,
+        type_vocab_size: int = 2,
+        vocab_size: int = 261,
+        classifier_dropout: Optional[float] = None,
+        **kwargs,
+    ):
+        self.downsampling_factor = downsampling_factor
+        self.max_block_size = max_block_size
+        self.score_consensus_attn = score_consensus_attn
+        self.upsample_output = upsample_output
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.embedding_size = embedding_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.hidden_size = hidden_size
+        self.initializer_range = initializer_range
+        self.intermediate_size = intermediate_size
+        self.layer_norm_eps = layer_norm_eps
+        self.max_position_embeddings = max_position_embeddings
+        self.model_type = model_type
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.pad_token_id = pad_token_id
+        self.position_embedding_type = position_embedding_type
+        self.summary_activation = summary_activation
+        self.summary_last_dropout = summary_last_dropout
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.type_vocab_size = type_vocab_size
+        self.vocab_size = vocab_size
+        self.sampling = sampling
+        self.classifier_dropout = classifier_dropout
+        super().__init__(**kwargs)

gbst.py

@@ -0,0 +1,405 @@
+import math
+from math import gcd
+import functools
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+
+from einops import rearrange, reduce, repeat
+from einops.layers.torch import Rearrange
+from transformers.modeling_utils import PreTrainedModel
+
+
+def exists(val):
+    return val is not None
+
+
+def lcm(*numbers):
+    return int(functools.reduce(lambda x, y: int((x * y) / gcd(x, y)), numbers, 1))
+
+
+def masked_mean(tensor, mask, dim = -1):
+    diff_len = len(tensor.shape) - len(mask.shape)
+    mask = mask[(..., *((None,) * diff_len))]
+    tensor.masked_fill_(~mask, 0.)
+
+    total_el = mask.sum(dim = dim)
+    mean = tensor.sum(dim = dim) / total_el.clamp(min = 1.)
+    mean.masked_fill_(total_el == 0, 0.)
+    return mean
+
+
+def next_divisible_length(seqlen, multiple):
+    return math.ceil(seqlen / multiple) * multiple
+
+
+def pad_to_multiple(tensor, multiple, *, seq_dim, dim = -1, value = 0.):
+    seqlen = tensor.shape[seq_dim]
+    length = next_divisible_length(seqlen, multiple)
+    if length == seqlen:
+        return tensor
+    remainder = length - seqlen
+    pad_offset = (0,) * (-1 - dim) * 2
+    return F.pad(tensor, (*pad_offset, 0, remainder), value = value)
+
+
+# helper classes
+class Pad(nn.Module):
+    def __init__(self, padding, value = 0.):
+        super().__init__()
+        self.padding = padding
+        self.value = value
+
+    def forward(self, x):
+        return F.pad(x, self.padding, value = self.value)
+
+
+class DepthwiseConv1d(nn.Module):
+    def __init__(self, dim_in, dim_out, kernel_size):
+        super().__init__()
+        self.conv = nn.Conv1d(dim_in, dim_out, kernel_size, groups = dim_in)
+        self.proj_out = nn.Conv1d(dim_out, dim_out, 1)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return self.proj_out(x)
+
+
+# main class
+class GBST(PreTrainedModel):
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def __init__(
+        self,
+        *,
+        num_tokens,
+        dim,
+        max_block_size = None,
+        blocks = None,
+        downsample_factor = 4,
+        score_consensus_attn = True,
+        return_without_downsample = True,
+            config = None
+    ):
+        super(GBST, self).__init__(config=config)
+        assert exists(max_block_size) ^ exists(blocks), 'either max_block_size or blocks are given on initialization'
+        self.word_embeddings = nn.Embedding(num_tokens, dim)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, dim)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, dim)
+
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+        self.return_without_downsample = return_without_downsample
+
+        if exists(blocks):
+            assert isinstance(blocks, tuple), 'blocks must be a tuple of block sizes'
+            self.blocks = tuple(map(lambda el: el if isinstance(el, tuple) else (el, 0), blocks))
+            assert all([(offset < block_size) for block_size, offset in self.blocks]), 'offset must be always smaller than the block size'
+
+            max_block_size = max(list(map(lambda t: t[0], self.blocks)))
+        else:
+            self.blocks = tuple(map(lambda el: (el, 0), range(1, max_block_size + 1)))
+
+        self.pos_conv = nn.Sequential(
+            Pad((0, 0, 0, max_block_size - 1)),
+            Rearrange('b n d -> b d n'),
+            DepthwiseConv1d(dim, dim, kernel_size = max_block_size),
+            Rearrange('b d n -> b n d')
+        )
+
+        self.score_fn = nn.Sequential(
+            nn.Linear(dim, 1),
+            Rearrange('... () -> ...')
+        )
+
+        self.score_consensus_attn = score_consensus_attn
+
+        assert downsample_factor <= max_block_size, 'final downsample factor should be less than the maximum block size'
+
+        self.block_pad_multiple = lcm(*[block_size for block_size, _ in self.blocks])
+        self.downsample_factor = downsample_factor
+
+    def forward(self, input_ids, attention_mask=None, position_ids=None, token_type_ids=None, inputs_embeds=None):
+        b, n, block_mult, ds_factor, device = *input_ids.shape, self.block_pad_multiple, self.downsample_factor, input_ids.device
+        m = next_divisible_length(n, ds_factor)
+
+        # get character token embeddings
+
+        input_ids = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        seq_len = input_ids.size()[1]
+        position_ids = self.position_ids[:, :seq_len]
+        position_embeddings = self.position_embeddings(position_ids)
+
+        input_ids = input_ids + token_type_embeddings + position_embeddings
+        # do a conv to generate the positions for the tokens
+
+        input_ids = self.pos_conv(input_ids)
+
+        # pad both sequence and attention_mask to length visibile by all block sizes from 0 to max block size
+
+        input_ids = pad_to_multiple(input_ids, block_mult, seq_dim=1, dim=-2)
+
+        if exists(attention_mask):
+            attention_mask = pad_to_multiple(attention_mask, block_mult, seq_dim=1, dim=-1, value=False)
+
+        # compute representations for all blocks by mean pooling
+
+        block_masks = []
+        block_reprs = []
+
+        for block_size, offset in self.blocks:
+            # clone the input sequence as well as the attention_mask, in order to pad for offsets
+
+            block_x = input_ids.clone()
+
+            if exists(attention_mask):
+                block_mask = attention_mask.clone()
+
+            # pad for offsets, if needed
+
+            need_padding = offset > 0
+
+            if need_padding:
+                left_offset, right_offset = (block_size - offset), offset
+                block_x = F.pad(block_x, (0, 0, left_offset, right_offset), value = 0.)
+
+                if exists(attention_mask):
+                    block_mask = F.pad(block_mask, (left_offset, right_offset), value = False)
+
+            # group input sequence into blocks
+
+            blocks = rearrange(block_x, 'b (n m) d -> b n m d', m = block_size)
+
+            # either mean pool the blocks, or do a masked mean
+
+            if exists(attention_mask):
+                mask_blocks = rearrange(block_mask, 'b (n m) -> b n m', m = block_size)
+                block_repr = masked_mean(blocks, mask_blocks, dim = -2)
+            else:
+                block_repr = blocks.mean(dim = -2)
+
+            # append the block representations, as well as the pooled block masks
+
+            block_repr = repeat(block_repr, 'b n d -> b (n m) d', m = block_size)
+
+            if need_padding:
+                block_repr = block_repr[:, left_offset:-right_offset]
+
+            block_reprs.append(block_repr)
+
+            if exists(attention_mask):
+                mask_blocks = torch.any(mask_blocks, dim = -1)
+                mask_blocks = repeat(mask_blocks, 'b n -> b (n m)', m = block_size)
+
+                if need_padding:
+                    mask_blocks = mask_blocks[:, left_offset:-right_offset]
+
+                block_masks.append(mask_blocks)
+
+        # stack all the block representations
+
+        block_reprs = torch.stack(block_reprs, dim = 2)
+
+        # calculate scores and softmax across the block size dimension
+
+        scores = self.score_fn(block_reprs)
+
+        if exists(attention_mask):
+            block_masks = torch.stack(block_masks, dim = 2)
+            max_neg_value = -torch.finfo(scores.dtype).max
+            scores = scores.masked_fill(~block_masks, max_neg_value)
+
+        scores = scores.softmax(dim = 2)
+
+        # do the cheap consensus attention, eq (5) in paper
+
+        if self.score_consensus_attn:
+            score_sim = einsum('b i d, b j d -> b i j', scores, scores)
+
+            if exists(attention_mask):
+                cross_mask = rearrange(attention_mask, 'b i -> b i ()') * rearrange(attention_mask, 'b j -> b () j')
+                max_neg_value = -torch.finfo(score_sim.dtype).max
+                score_sim = score_sim.masked_fill(~cross_mask, max_neg_value)
+
+            score_attn = score_sim.softmax(dim=-1)
+            scores = einsum('b i j, b j m -> b i m', score_attn, scores)
+
+        # multiply the block representations by the position-wise scores
+
+        scores = rearrange(scores, 'b n m -> b n m ()')
+        input_ids = (block_reprs * scores).sum(dim=2)
+
+        # truncate to length divisible by downsample factor
+
+        input_ids = input_ids[:, :m]
+
+        original = None
+        if self.return_without_downsample:
+            original = torch.clone(input_ids)
+
+        input_ids, attention_mask = self.down_sample(input_ids, attention_mask, ds_factor)
+
+        return input_ids, attention_mask, original
+
+    @staticmethod
+    def down_sample(input_ids, attention_mask, ds_factor):
+        n = input_ids.shape[1]
+        m = next_divisible_length(n, ds_factor)
+        if exists(attention_mask):
+            attention_mask = attention_mask[:, :m]
+
+        # final mean pooling downsample
+        input_ids = rearrange(input_ids, 'b (n m) d -> b n m d', m=ds_factor)
+
+        if exists(attention_mask):
+            attention_mask = rearrange(attention_mask, 'b (n m) -> b n m', m=ds_factor)
+            input_ids = masked_mean(input_ids, attention_mask, dim=2)
+            attention_mask = torch.any(attention_mask, dim=-1)
+        else:
+            input_ids = input_ids.mean(dim=-2)
+        return input_ids, attention_mask
+
+    def block_score(self, input_ids, attention_mask=None, position_ids=None, token_type_ids=None, inputs_embeds=None):
+        b, n, block_mult, ds_factor, device = *input_ids.shape, self.block_pad_multiple, self.downsample_factor, input_ids.device
+        m = next_divisible_length(n, ds_factor)
+
+        # get character token embeddings
+
+        input_ids = self.word_embeddings(input_ids)
+
+        # do a conv to generate the positions for the tokens
+
+        input_ids = self.pos_conv(input_ids)
+
+        # pad both sequence and attention_mask to length visibile by all block sizes from 0 to max block size
+
+        input_ids = pad_to_multiple(input_ids, block_mult, seq_dim=1, dim=-2)
+
+        if exists(attention_mask):
+            attention_mask = pad_to_multiple(attention_mask, block_mult, seq_dim=1, dim=-1, value=False)
+
+        # compute representations for all blocks by mean pooling
+
+        block_masks = []
+        block_reprs = []
+
+        for block_size, offset in self.blocks:
+            # clone the input sequence as well as the attention_mask, in order to pad for offsets
+
+            block_x = input_ids.clone()
+
+            if exists(attention_mask):
+                block_mask = attention_mask.clone()
+
+            # pad for offsets, if needed
+
+            need_padding = offset > 0
+
+            if need_padding:
+                left_offset, right_offset = (block_size - offset), offset
+                block_x = F.pad(block_x, (0, 0, left_offset, right_offset), value = 0.)
+
+                if exists(attention_mask):
+                    block_mask = F.pad(block_mask, (left_offset, right_offset), value = False)
+
+            # group input sequence into blocks
+
+            blocks = rearrange(block_x, 'b (n m) d -> b n m d', m = block_size)
+
+            # either mean pool the blocks, or do a masked mean
+
+            if exists(attention_mask):
+                mask_blocks = rearrange(block_mask, 'b (n m) -> b n m', m = block_size)
+                block_repr = masked_mean(blocks, mask_blocks, dim = -2)
+            else:
+                block_repr = blocks.mean(dim = -2)
+
+            # append the block representations, as well as the pooled block masks
+
+            block_repr = repeat(block_repr, 'b n d -> b (n m) d', m = block_size)
+
+            if need_padding:
+                block_repr = block_repr[:, left_offset:-right_offset]
+
+            block_reprs.append(block_repr)
+
+            if exists(attention_mask):
+                mask_blocks = torch.any(mask_blocks, dim = -1)
+                mask_blocks = repeat(mask_blocks, 'b n -> b (n m)', m = block_size)
+
+                if need_padding:
+                    mask_blocks = mask_blocks[:, left_offset:-right_offset]
+
+                block_masks.append(mask_blocks)
+
+        # stack all the block representations
+
+        block_reprs = torch.stack(block_reprs, dim = 2)
+
+        # calculate scores and softmax across the block size dimension
+
+        scores = self.score_fn(block_reprs)
+
+        if exists(attention_mask):
+            block_masks = torch.stack(block_masks, dim = 2)
+            max_neg_value = -torch.finfo(scores.dtype).max
+            scores = scores.masked_fill(~block_masks, max_neg_value)
+
+        scores = scores.softmax(dim = 2)
+
+        # do the cheap consensus attention, eq (5) in paper
+
+        if self.score_consensus_attn:
+            score_sim = einsum('b i d, b j d -> b i j', scores, scores)
+
+            if exists(attention_mask):
+                cross_mask = rearrange(attention_mask, 'b i -> b i ()') * rearrange(attention_mask, 'b j -> b () j')
+                max_neg_value = -torch.finfo(score_sim.dtype).max
+                score_sim = score_sim.masked_fill(~cross_mask, max_neg_value)
+
+            score_attn = score_sim.softmax(dim=-1)
+            scores = einsum('b i j, b j m -> b i m', score_attn, scores)
+
+        # multiply the block representations by the position-wise scores
+
+        scores = rearrange(scores, 'b n m -> b n m ()')
+        input_ids = (block_reprs * scores).sum(dim=2)
+
+        # truncate to length divisible by downsample factor
+
+        input_ids = input_ids[:, :m]
+
+        if exists(attention_mask):
+            attention_mask = attention_mask[:, :m]
+
+        original = None
+        if self.return_without_downsample:
+            original = torch.clone(input_ids)
+
+        # final mean pooling downsample
+        input_ids = rearrange(input_ids, 'b (n m) d -> b n m d', m=ds_factor)
+
+        if exists(attention_mask):
+            attention_mask = rearrange(attention_mask, 'b (n m) -> b n m', m=ds_factor)
+            input_ids = masked_mean(input_ids, attention_mask, dim=2)
+            attention_mask = torch.any(attention_mask, dim=-1)
+        else:
+            input_ids = input_ids.mean(dim=-2)
+
+        return scores

modeling_charmen.py

@@ -0,0 +1,410 @@
+from transformers.models.electra.modeling_electra import ElectraPreTrainedModel, ElectraEncoder, ElectraLayer, \
+    ModelOutput, ElectraForSequenceClassification, SequenceClassifierOutput, ElectraForTokenClassification, \
+    ElectraForMultipleChoice
+from .config import CharmenElectraConfig
+from .gbst import GBST
+import torch.nn as nn
+import copy
+import torch
+from torch import Tensor
+from dataclasses import dataclass
+from typing import Optional, Tuple
+from typing import OrderedDict as OrderDictType
+from collections import OrderedDict
+from transformers.activations import get_activation
+
+
+@dataclass
+class CharmenElectraModelOutput(ModelOutput):
+    """
+    Output type of :class:`~.CharmenElectraModel`.
+    """
+    downsampled_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    upsampled_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class CharmenElectraModel(ElectraPreTrainedModel):
+    config_class = CharmenElectraConfig
+
+    def __init__(self, config: CharmenElectraConfig, compatibility_with_transformers=False):
+        super().__init__(config)
+        self.embeddings: GBST = GBST(
+            num_tokens=config.vocab_size,
+            # number of tokens, should be 256 for byte encoding (+ 1 special token for padding in this example)
+            dim=config.embedding_size,  # dimension of token and intra-block positional embedding
+            max_block_size=config.max_block_size,  # maximum block size
+            downsample_factor=config.downsampling_factor,
+            # the final downsample factor by which the sequence length will decrease by
+            score_consensus_attn=config.score_consensus_attn,
+            config=config
+            # whether to do the cheap score consensus (aka attention) as in eq. 5 in the paper
+        )
+        self.compatibility_with_transformers = compatibility_with_transformers
+
+        if config.embedding_size != config.hidden_size:
+            self.embeddings_project = nn.Linear(config.embedding_size, config.hidden_size)
+
+        self.upsampling = nn.Upsample(scale_factor=config.downsampling_factor, mode='nearest')
+        self.upsampling_convolution = nn.Conv1d(in_channels=config.hidden_size * 2,
+                                                out_channels=config.hidden_size,
+                                                kernel_size=(config.downsampling_factor*2-1,),
+                                                padding='same',
+                                                dilation=(1,))
+        self.upsample_output = config.upsample_output
+
+        # config.num_hidden_layers = config.num_hidden_layers - 2
+
+        cfg = copy.deepcopy(config)
+        cfg.num_hidden_layers = config.num_hidden_layers - 2
+        self.encoder = ElectraEncoder(cfg)
+
+        # frame_hidden_size
+        self.encoder_first_layer = ElectraLayer(config)
+        self.encoder_last_layer = ElectraLayer(config)
+
+        self.config = config
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=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.use_return_dict
+
+        if input_ids.shape.__len__() == 1:
+            input_ids = input_ids.view(1, -1)
+            attention_mask = attention_mask.view(1, -1)
+            token_type_ids = token_type_ids.view(1, -1)
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        unscaled_attention_mask = torch.clone(attention_mask)
+
+        _, _, unscaled_hidden_states = self.embeddings(
+            input_ids=input_ids, attention_mask=attention_mask,
+            position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+
+        if hasattr(self, "embeddings_project"):
+            unscaled_hidden_states = self.embeddings_project(unscaled_hidden_states)
+
+        extended_unscaled_attention_mask = self.get_extended_attention_mask(unscaled_attention_mask, input_shape,
+                                                                            device)
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        unscaled_hidden_states = self.encoder_first_layer(unscaled_hidden_states, extended_unscaled_attention_mask,
+                                                          None, None, None, None, False)[0]
+
+        hidden_states, attention_mask = self.embeddings.down_sample(unscaled_hidden_states, unscaled_attention_mask,
+                                                                    self.config.downsampling_factor)
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        encoder_output = self.encoder(
+            hidden_states,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        downsampled_hidden_states = encoder_output[0]
+        hidden_states = encoder_output[0]
+
+        # upsampling
+        upsampled = self.upsampling(hidden_states.permute(0, 2, 1)).permute(0, 2, 1)
+        hidden_states = torch.cat([unscaled_hidden_states, upsampled], dim=-1)
+        # padded_hidden_states = F.pad(hidden_states.permute(0, 2, 1), (3, 3))
+        hidden_states = self.upsampling_convolution(hidden_states.permute(0, 2, 1)).permute(0, 2, 1)
+
+        hidden_states = self.encoder_last_layer(hidden_states, extended_unscaled_attention_mask,
+                                                None, None, None, None, False)
+
+        upsampled_output = hidden_states[0]
+
+        return CharmenElectraModelOutput(
+            downsampled_hidden_states=downsampled_hidden_states,
+            upsampled_hidden_states=upsampled_output
+        )
+
+    def load_state_dict(self, state_dict: OrderDictType[str, Tensor], strict: bool = True):
+        model = OrderedDict()
+        prefix = "discriminator.electra."
+
+        for key, value in state_dict.items():
+            if key.startswith(prefix):
+                model[key[len(prefix):]] = value
+
+        super(CharmenElectraModel, self).load_state_dict(model, strict)
+
+
+class CharmenElectraClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config: CharmenElectraConfig):
+        super().__init__()
+        self.config = config
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+        self.ds_factor = config.downsampling_factor
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = get_activation(self.config.summary_activation)(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+class CharmenElectraForSequenceClassification(ElectraForSequenceClassification):
+    config_class = CharmenElectraConfig
+
+    def __init__(self, config: CharmenElectraConfig, class_weight=None, label_smoothing=0.0):
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.config = config
+        self.model = CharmenElectraModel(config, compatibility_with_transformers=True)
+        self.classifier = CharmenElectraClassificationHead(config)
+        self.cls_loss_fct = torch.nn.CrossEntropyLoss(weight=class_weight, label_smoothing=label_smoothing)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        output_discriminator: CharmenElectraModelOutput = self.model(input_ids, attention_mask, token_type_ids)
+
+        if self.carmen_config.upsample_output:
+            cls = self.classifier(output_discriminator.upsampled_hidden_states)
+        else:
+            cls = self.classifier(output_discriminator.downsampled_hidden_states)
+        cls_loss = self.cls_loss_fct(cls, labels)
+
+        return SequenceClassifierOutput(
+            loss=cls_loss,
+            logits=cls,
+            hidden_states=output_discriminator.downsampled_hidden_states,
+            attentions=None,
+        )
+
+    def load_state_dict(self, state_dict: OrderDictType[str, Tensor], strict: bool = True):
+        model = OrderedDict()
+        prefix = "discriminator.electra."
+
+        for key, value in state_dict.items():
+            if key.startswith(prefix):
+                model[key[len(prefix):]] = value
+
+        self.model.load_state_dict(state_dict=model, strict=strict)
+
+
+class CharmenElectraForTokenClassification(ElectraForTokenClassification):
+    def __init__(self, config: CharmenElectraConfig, class_weight=None, label_smoothing=0.0):
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.carmen_config = config
+        self.model = CharmenElectraModel(config, compatibility_with_transformers=True)
+
+        classifier_dropout = (
+            config.discriminator.classifier_dropout if config.discriminator.classifier_dropout is not None else config.discriminator.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.discriminator.hidden_size, config.num_labels)
+
+        self.cls_loss_fct = torch.nn.CrossEntropyLoss(weight=class_weight, label_smoothing=label_smoothing)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        output_discriminator: CharmenElectraModelOutput = self.model(
+            input_ids, attention_mask, token_type_ids)
+
+        discriminator_sequence_output = self.dropout(output_discriminator.upsampled_hidden_states)
+        logits = self.classifier(discriminator_sequence_output)
+
+        if labels is not None:
+            cls_loss = self.cls_loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
+        else:
+            cls_loss = None
+
+        return SequenceClassifierOutput(
+            loss=cls_loss,
+            logits=logits,
+            hidden_states=output_discriminator.upsampled_hidden_states,
+            attentions=None,
+        )
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.electra.get_input_embeddings()
+
+    def load_state_dict(self, state_dict: OrderDictType[str, Tensor], strict: bool = True):
+        model = OrderedDict()
+        prefix = "discriminator.electra."
+
+        for key, value in state_dict.items():
+            if key.startswith(prefix):
+                model[key[len(prefix):]] = value
+
+        self.model.load_state_dict(state_dict=model, strict=strict)
+
+
+class Pooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class CharmenElectraForMultipleChoice(ElectraForMultipleChoice):
+    def __init__(self, config: CharmenElectraConfig, class_weight=None, label_smoothing=0.0):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+        self.carmen_config = config
+        self.model = CharmenElectraModel(config, compatibility_with_transformers=True)
+        self.pooler = Pooler(config)
+
+        classifier_dropout = (
+            config.classifier_dropout if config.discriminator.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.cls_loss_fct = torch.nn.CrossEntropyLoss(weight=class_weight, label_smoothing=label_smoothing)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+
+        output_discriminator: CharmenElectraModelOutput = self.model(
+            input_ids, attention_mask, token_type_ids)
+
+        if self.carmen_config.upsample_output:
+            pooled_output = self.pooler(output_discriminator.upsampled_hidden_states)
+        else:
+            pooled_output = self.pooler(output_discriminator.downsampled_hidden_states)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        cls_loss = self.cls_loss_fct(reshaped_logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=cls_loss,
+            logits=reshaped_logits,
+            hidden_states=output_discriminator.downsampled_hidden_states,
+            attentions=None,
+        )
+
+    def load_state_dict(self, state_dict: OrderDictType[str, Tensor], strict: bool = True):
+        model = OrderedDict()
+        prefix = "discriminator.electra."
+
+        for key, value in state_dict.items():
+            if key.startswith(prefix):
+                model[key[len(prefix):]] = value
+
+        self.model.load_state_dict(state_dict=model, strict=strict)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7074667cdc918bf66a2b408b6e879995964891452d4dd598f0b42fbbdc0ee60b
+size 173978597