# coding=utf-8
# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch XLNet model.
"""
import logging
import torch
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss
from torch.nn import functional as F
from .activations import gelu_new, swish
from .configuration_xlnet import XLNetConfig
from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
from .modeling_utils import PoolerAnswerClass, PoolerEndLogits, PoolerStartLogits, PreTrainedModel, SequenceSummary
logger = logging.getLogger(__name__)
XLNET_PRETRAINED_MODEL_ARCHIVE_MAP = {
"xlnet-base-cased": "https://cdn.huggingface.co/xlnet-base-cased-pytorch_model.bin",
"xlnet-large-cased": "https://cdn.huggingface.co/xlnet-large-cased-pytorch_model.bin",
}
def build_tf_xlnet_to_pytorch_map(model, config, tf_weights=None):
""" A map of modules from TF to PyTorch.
I use a map to keep the PyTorch model as
identical to the original PyTorch model as possible.
"""
tf_to_pt_map = {}
if hasattr(model, "transformer"):
if hasattr(model, "lm_loss"):
# We will load also the output bias
tf_to_pt_map["model/lm_loss/bias"] = model.lm_loss.bias
if hasattr(model, "sequence_summary") and "model/sequnece_summary/summary/kernel" in tf_weights:
# We will load also the sequence summary
tf_to_pt_map["model/sequnece_summary/summary/kernel"] = model.sequence_summary.summary.weight
tf_to_pt_map["model/sequnece_summary/summary/bias"] = model.sequence_summary.summary.bias
if (
hasattr(model, "logits_proj")
and config.finetuning_task is not None
and "model/regression_{}/logit/kernel".format(config.finetuning_task) in tf_weights
):
tf_to_pt_map["model/regression_{}/logit/kernel".format(config.finetuning_task)] = model.logits_proj.weight
tf_to_pt_map["model/regression_{}/logit/bias".format(config.finetuning_task)] = model.logits_proj.bias
# Now load the rest of the transformer
model = model.transformer
# Embeddings and output
tf_to_pt_map.update(
{
"model/transformer/word_embedding/lookup_table": model.word_embedding.weight,
"model/transformer/mask_emb/mask_emb": model.mask_emb,
}
)
# Transformer blocks
for i, b in enumerate(model.layer):
layer_str = "model/transformer/layer_%d/" % i
tf_to_pt_map.update(
{
layer_str + "rel_attn/LayerNorm/gamma": b.rel_attn.layer_norm.weight,
layer_str + "rel_attn/LayerNorm/beta": b.rel_attn.layer_norm.bias,
layer_str + "rel_attn/o/kernel": b.rel_attn.o,
layer_str + "rel_attn/q/kernel": b.rel_attn.q,
layer_str + "rel_attn/k/kernel": b.rel_attn.k,
layer_str + "rel_attn/r/kernel": b.rel_attn.r,
layer_str + "rel_attn/v/kernel": b.rel_attn.v,
layer_str + "ff/LayerNorm/gamma": b.ff.layer_norm.weight,
layer_str + "ff/LayerNorm/beta": b.ff.layer_norm.bias,
layer_str + "ff/layer_1/kernel": b.ff.layer_1.weight,
layer_str + "ff/layer_1/bias": b.ff.layer_1.bias,
layer_str + "ff/layer_2/kernel": b.ff.layer_2.weight,
layer_str + "ff/layer_2/bias": b.ff.layer_2.bias,
}
)
# Relative positioning biases
if config.untie_r:
r_r_list = []
r_w_list = []
r_s_list = []
seg_embed_list = []
for b in model.layer:
r_r_list.append(b.rel_attn.r_r_bias)
r_w_list.append(b.rel_attn.r_w_bias)
r_s_list.append(b.rel_attn.r_s_bias)
seg_embed_list.append(b.rel_attn.seg_embed)
else:
r_r_list = [model.r_r_bias]
r_w_list = [model.r_w_bias]
r_s_list = [model.r_s_bias]
seg_embed_list = [model.seg_embed]
tf_to_pt_map.update(
{
"model/transformer/r_r_bias": r_r_list,
"model/transformer/r_w_bias": r_w_list,
"model/transformer/r_s_bias": r_s_list,
"model/transformer/seg_embed": seg_embed_list,
}
)
return tf_to_pt_map
def load_tf_weights_in_xlnet(model, config, tf_path):
""" Load tf checkpoints in a pytorch model
"""
try:
import numpy as np
import tensorflow as tf
except ImportError:
logger.error(
"Loading a TensorFlow models in PyTorch, requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions."
)
raise
# Load weights from TF model
init_vars = tf.train.list_variables(tf_path)
tf_weights = {}
for name, shape in init_vars:
logger.info("Loading TF weight {} with shape {}".format(name, shape))
array = tf.train.load_variable(tf_path, name)
tf_weights[name] = array
# Build TF to PyTorch weights loading map
tf_to_pt_map = build_tf_xlnet_to_pytorch_map(model, config, tf_weights)
for name, pointer in tf_to_pt_map.items():
logger.info("Importing {}".format(name))
if name not in tf_weights:
logger.info("{} not in tf pre-trained weights, skipping".format(name))
continue
array = tf_weights[name]
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if "kernel" in name and ("ff" in name or "summary" in name or "logit" in name):
logger.info("Transposing")
array = np.transpose(array)
if isinstance(pointer, list):
# Here we will split the TF weights
assert len(pointer) == array.shape[0]
for i, p_i in enumerate(pointer):
arr_i = array[i, ...]
try:
assert p_i.shape == arr_i.shape
except AssertionError as e:
e.args += (p_i.shape, arr_i.shape)
raise
logger.info("Initialize PyTorch weight {} for layer {}".format(name, i))
p_i.data = torch.from_numpy(arr_i)
else:
try:
assert pointer.shape == array.shape
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
logger.info("Initialize PyTorch weight {}".format(name))
pointer.data = torch.from_numpy(array)
tf_weights.pop(name, None)
tf_weights.pop(name + "/Adam", None)
tf_weights.pop(name + "/Adam_1", None)
logger.info("Weights not copied to PyTorch model: {}".format(", ".join(tf_weights.keys())))
return model
ACT2FN = {"gelu": gelu_new, "relu": torch.nn.functional.relu, "swish": swish}
XLNetLayerNorm = nn.LayerNorm
class XLNetRelativeAttention(nn.Module):
def __init__(self, config):
super().__init__()
self.output_attentions = config.output_attentions
if config.d_model % config.n_head != 0:
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention "
"heads (%d)" % (config.d_model, config.n_head)
)
self.n_head = config.n_head
self.d_head = config.d_head
self.d_model = config.d_model
self.scale = 1 / (config.d_head ** 0.5)
self.q = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
self.k = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
self.v = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
self.o = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
self.r = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
self.r_s_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
self.seg_embed = nn.Parameter(torch.FloatTensor(2, self.n_head, self.d_head))
self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
self.dropout = nn.Dropout(config.dropout)
def prune_heads(self, heads):
raise NotImplementedError
@staticmethod
def rel_shift(x, klen=-1):
"""perform relative shift to form the relative attention score."""
x_size = x.shape
x = x.reshape(x_size[1], x_size[0], x_size[2], x_size[3])
x = x[1:, ...]
x = x.reshape(x_size[0], x_size[1] - 1, x_size[2], x_size[3])
# x = x[:, 0:klen, :, :]
x = torch.index_select(x, 1, torch.arange(klen, device=x.device, dtype=torch.long))
return x
@staticmethod
def rel_shift_bnij(x, klen=-1):
x_size = x.shape
x = x.reshape(x_size[0], x_size[1], x_size[3], x_size[2])
x = x[:, :, 1:, :]
x = x.reshape(x_size[0], x_size[1], x_size[2], x_size[3] - 1)
# Note: the tensor-slice form was faster in my testing than torch.index_select
# However, tracing doesn't like the nature of the slice, and if klen changes
# during the run then it'll fail, whereas index_select will be fine.
x = torch.index_select(x, 3, torch.arange(klen, device=x.device, dtype=torch.long))
# x = x[:, :, :, :klen]
return x
def rel_attn_core(self, q_head, k_head_h, v_head_h, k_head_r, seg_mat=None, attn_mask=None, head_mask=None):
"""Core relative positional attention operations."""
# content based attention score
ac = torch.einsum("ibnd,jbnd->bnij", q_head + self.r_w_bias, k_head_h)
# position based attention score
bd = torch.einsum("ibnd,jbnd->bnij", q_head + self.r_r_bias, k_head_r)
bd = self.rel_shift_bnij(bd, klen=ac.shape[3])
# segment based attention score
if seg_mat is None:
ef = 0
else:
ef = torch.einsum("ibnd,snd->ibns", q_head + self.r_s_bias, self.seg_embed)
ef = torch.einsum("ijbs,ibns->bnij", seg_mat, ef)
# merge attention scores and perform masking
attn_score = (ac + bd + ef) * self.scale
if attn_mask is not None:
# attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
if attn_mask.dtype == torch.float16:
attn_score = attn_score - 65500 * torch.einsum("ijbn->bnij", attn_mask)
else:
attn_score = attn_score - 1e30 * torch.einsum("ijbn->bnij", attn_mask)
# attention probability
attn_prob = F.softmax(attn_score, dim=3)
attn_prob = self.dropout(attn_prob)
# Mask heads if we want to
if head_mask is not None:
attn_prob = attn_prob * torch.einsum("ijbn->bnij", head_mask)
# attention output
attn_vec = torch.einsum("bnij,jbnd->ibnd", attn_prob, v_head_h)
if self.output_attentions:
return attn_vec, torch.einsum("bnij->ijbn", attn_prob)
return attn_vec
def post_attention(self, h, attn_vec, residual=True):
"""Post-attention processing."""
# post-attention projection (back to `d_model`)
attn_out = torch.einsum("ibnd,hnd->ibh", attn_vec, self.o)
attn_out = self.dropout(attn_out)
if residual:
attn_out = attn_out + h
output = self.layer_norm(attn_out)
return output
def forward(self, h, g, attn_mask_h, attn_mask_g, r, seg_mat, mems=None, target_mapping=None, head_mask=None):
if g is not None:
# Two-stream attention with relative positional encoding.
# content based attention score
if mems is not None and mems.dim() > 1:
cat = torch.cat([mems, h], dim=0)
else:
cat = h
# content-based key head
k_head_h = torch.einsum("ibh,hnd->ibnd", cat, self.k)
# content-based value head
v_head_h = torch.einsum("ibh,hnd->ibnd", cat, self.v)
# position-based key head
k_head_r = torch.einsum("ibh,hnd->ibnd", r, self.r)
# h-stream
# content-stream query head
q_head_h = torch.einsum("ibh,hnd->ibnd", h, self.q)
# core attention ops
attn_vec_h = self.rel_attn_core(
q_head_h, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_h, head_mask=head_mask
)
if self.output_attentions:
attn_vec_h, attn_prob_h = attn_vec_h
# post processing
output_h = self.post_attention(h, attn_vec_h)
# g-stream
# query-stream query head
q_head_g = torch.einsum("ibh,hnd->ibnd", g, self.q)
# core attention ops
if target_mapping is not None:
q_head_g = torch.einsum("mbnd,mlb->lbnd", q_head_g, target_mapping)
attn_vec_g = self.rel_attn_core(
q_head_g, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_g, head_mask=head_mask
)
if self.output_attentions:
attn_vec_g, attn_prob_g = attn_vec_g
attn_vec_g = torch.einsum("lbnd,mlb->mbnd", attn_vec_g, target_mapping)
else:
attn_vec_g = self.rel_attn_core(
q_head_g, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_g, head_mask=head_mask
)
if self.output_attentions:
attn_vec_g, attn_prob_g = attn_vec_g
# post processing
output_g = self.post_attention(g, attn_vec_g)
if self.output_attentions:
attn_prob = attn_prob_h, attn_prob_g
else:
# Multi-head attention with relative positional encoding
if mems is not None and mems.dim() > 1:
cat = torch.cat([mems, h], dim=0)
else:
cat = h
# content heads
q_head_h = torch.einsum("ibh,hnd->ibnd", h, self.q)
k_head_h = torch.einsum("ibh,hnd->ibnd", cat, self.k)
v_head_h = torch.einsum("ibh,hnd->ibnd", cat, self.v)
# positional heads
k_head_r = torch.einsum("ibh,hnd->ibnd", r, self.r)
# core attention ops
attn_vec = self.rel_attn_core(
q_head_h, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_h, head_mask=head_mask
)
if self.output_attentions:
attn_vec, attn_prob = attn_vec
# post processing
output_h = self.post_attention(h, attn_vec)
output_g = None
outputs = (output_h, output_g)
if self.output_attentions:
outputs = outputs + (attn_prob,)
return outputs
class XLNetFeedForward(nn.Module):
def __init__(self, config):
super().__init__()
self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
self.layer_1 = nn.Linear(config.d_model, config.d_inner)
self.layer_2 = nn.Linear(config.d_inner, config.d_model)
self.dropout = nn.Dropout(config.dropout)
if isinstance(config.ff_activation, str):
self.activation_function = ACT2FN[config.ff_activation]
else:
self.activation_function = config.ff_activation
def forward(self, inp):
output = inp
output = self.layer_1(output)
output = self.activation_function(output)
output = self.dropout(output)
output = self.layer_2(output)
output = self.dropout(output)
output = self.layer_norm(output + inp)
return output
class XLNetLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.rel_attn = XLNetRelativeAttention(config)
self.ff = XLNetFeedForward(config)
self.dropout = nn.Dropout(config.dropout)
def forward(
self, output_h, output_g, attn_mask_h, attn_mask_g, r, seg_mat, mems=None, target_mapping=None, head_mask=None
):
outputs = self.rel_attn(
output_h,
output_g,
attn_mask_h,
attn_mask_g,
r,
seg_mat,
mems=mems,
target_mapping=target_mapping,
head_mask=head_mask,
)
output_h, output_g = outputs[:2]
if output_g is not None:
output_g = self.ff(output_g)
output_h = self.ff(output_h)
outputs = (output_h, output_g) + outputs[2:] # Add again attentions if there are there
return outputs
class XLNetPreTrainedModel(PreTrainedModel):
""" An abstract class to handle weights initialization and
a simple interface for downloading and loading pretrained models.
"""
config_class = XLNetConfig
pretrained_model_archive_map = XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
load_tf_weights = load_tf_weights_in_xlnet
base_model_prefix = "transformer"
def _init_weights(self, module):
""" Initialize the weights.
"""
if isinstance(module, (nn.Linear, nn.Embedding)):
# 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 isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, XLNetLayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
elif isinstance(module, XLNetRelativeAttention):
for param in [
module.q,
module.k,
module.v,
module.o,
module.r,
module.r_r_bias,
module.r_s_bias,
module.r_w_bias,
module.seg_embed,
]:
param.data.normal_(mean=0.0, std=self.config.initializer_range)
elif isinstance(module, XLNetModel):
module.mask_emb.data.normal_(mean=0.0, std=self.config.initializer_range)
XLNET_START_DOCSTRING = r"""
This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
usage and behavior.
Parameters:
config (:class:`~transformers.XLNetConfig`): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the configuration.
Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""
XLNET_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.BertTokenizer`.
See :func:`transformers.PreTrainedTokenizer.encode` and
:func:`transformers.PreTrainedTokenizer.encode_plus` for details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on padding token indices.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
`What are attention masks? <../glossary.html#attention-mask>`__
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
(see `mems` output below). Can be used to speed up sequential decoding. The token ids which have their mems
given to this model should not be passed as input ids as they have already been computed.
`use_cache` has to be set to `True` to make use of `mems`.
perm_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to indicate the attention pattern for each input token with values selected in ``[0, 1]``:
If ``perm_mask[k, i, j] = 0``, i attend to j in batch k;
if ``perm_mask[k, i, j] = 1``, i does not attend to j in batch k.
If None, each token attends to all the others (full bidirectional attention).
Only used during pretraining (to define factorization order) or for sequential decoding (generation).
target_mapping (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_predict, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to indicate the output tokens to use.
If ``target_mapping[k, i, j] = 1``, the i-th predict in batch k is on the j-th token.
Only used during pretraining for partial prediction or for sequential decoding (generation).
token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Segment token indices to indicate first and second portions of the inputs.
Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
corresponds to a `sentence B` token. The classifier token should be represented by a ``2``.
`What are token type IDs? <../glossary.html#token-type-ids>`_
input_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on padding token indices.
Negative of `attention_mask`, i.e. with 0 for real tokens and 1 for padding.
Kept for compatibility with the original code base.
You can only uses one of `input_mask` and `attention_mask`
Mask values selected in ``[0, 1]``:
``1`` for tokens that are MASKED, ``0`` for tokens that are NOT MASKED.
head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
Mask to nullify selected heads of the self-attention modules.
Mask values selected in ``[0, 1]``:
:obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
use_cache (:obj:`bool`):
If `use_cache` is True, `mems` are returned and can be used to speed up decoding (see `mems`). Defaults to `True`.
"""
[docs]@add_start_docstrings(
"The bare XLNet Model transformer outputting raw hidden-states without any specific head on top.",
XLNET_START_DOCSTRING,
)
class XLNetModel(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.mem_len = config.mem_len
self.reuse_len = config.reuse_len
self.d_model = config.d_model
self.same_length = config.same_length
self.attn_type = config.attn_type
self.bi_data = config.bi_data
self.clamp_len = config.clamp_len
self.n_layer = config.n_layer
self.word_embedding = nn.Embedding(config.vocab_size, config.d_model)
self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, config.d_model))
self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
self.dropout = nn.Dropout(config.dropout)
self.init_weights()
def _prune_heads(self, heads_to_prune):
raise NotImplementedError
[docs] def create_mask(self, qlen, mlen):
"""
Creates causal attention mask. Float mask where 1.0 indicates masked, 0.0 indicates not-masked.
Args:
qlen: Sequence length
mlen: Mask length
::
same_length=False: same_length=True:
<mlen > < qlen > <mlen > < qlen >
^ [0 0 0 0 0 1 1 1 1] [0 0 0 0 0 1 1 1 1]
[0 0 0 0 0 0 1 1 1] [1 0 0 0 0 0 1 1 1]
qlen [0 0 0 0 0 0 0 1 1] [1 1 0 0 0 0 0 1 1]
[0 0 0 0 0 0 0 0 1] [1 1 1 0 0 0 0 0 1]
v [0 0 0 0 0 0 0 0 0] [1 1 1 1 0 0 0 0 0]
"""
attn_mask = torch.ones([qlen, qlen])
mask_up = torch.triu(attn_mask, diagonal=1)
attn_mask_pad = torch.zeros([qlen, mlen])
ret = torch.cat([attn_mask_pad, mask_up], dim=1)
if self.same_length:
mask_lo = torch.tril(attn_mask, diagonal=-1)
ret = torch.cat([ret[:, :qlen] + mask_lo, ret[:, qlen:]], dim=1)
ret = ret.to(next(self.parameters()))
return ret
def cache_mem(self, curr_out, prev_mem):
# cache hidden states into memory.
if self.reuse_len is not None and self.reuse_len > 0:
curr_out = curr_out[: self.reuse_len]
if prev_mem is None:
new_mem = curr_out[-self.mem_len :]
else:
new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len :]
return new_mem.detach()
@staticmethod
def positional_embedding(pos_seq, inv_freq, bsz=None):
sinusoid_inp = torch.einsum("i,d->id", pos_seq, inv_freq)
pos_emb = torch.cat([torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)], dim=-1)
pos_emb = pos_emb[:, None, :]
if bsz is not None:
pos_emb = pos_emb.expand(-1, bsz, -1)
return pos_emb
def relative_positional_encoding(self, qlen, klen, bsz=None):
# create relative positional encoding.
freq_seq = torch.arange(0, self.d_model, 2.0, dtype=torch.float)
inv_freq = 1 / torch.pow(10000, (freq_seq / self.d_model))
if self.attn_type == "bi":
# beg, end = klen - 1, -qlen
beg, end = klen, -qlen
elif self.attn_type == "uni":
# beg, end = klen - 1, -1
beg, end = klen, -1
else:
raise ValueError("Unknown `attn_type` {}.".format(self.attn_type))
if self.bi_data:
fwd_pos_seq = torch.arange(beg, end, -1.0, dtype=torch.float)
bwd_pos_seq = torch.arange(-beg, -end, 1.0, dtype=torch.float)
if self.clamp_len > 0:
fwd_pos_seq = fwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)
bwd_pos_seq = bwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)
if bsz is not None:
fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz // 2)
bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq, bsz // 2)
else:
fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq)
bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq)
pos_emb = torch.cat([fwd_pos_emb, bwd_pos_emb], dim=1)
else:
fwd_pos_seq = torch.arange(beg, end, -1.0)
if self.clamp_len > 0:
fwd_pos_seq = fwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)
pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz)
pos_emb = pos_emb.to(next(self.parameters()))
return pos_emb
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
):
r"""
Return:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_predict, hidden_size)`):
Sequence of hidden-states at the last layer of the model.
`num_predict` corresponds to `target_mapping.shape[1]`. If `target_mapping` is `None`, then `num_predict` corresponds to `sequence_length`.
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetModel
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetModel.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=False)).unsqueeze(0) # Batch size 1
outputs = model(input_ids)
last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple
"""
# the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
# but we want a unified interface in the library with the batch size on the first dimension
# so we move here the first dimension (batch) to the end
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_ids = input_ids.transpose(0, 1).contiguous()
qlen, bsz = input_ids.shape[0], input_ids.shape[1]
elif inputs_embeds is not None:
inputs_embeds = inputs_embeds.transpose(0, 1).contiguous()
qlen, bsz = inputs_embeds.shape[0], inputs_embeds.shape[1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
token_type_ids = token_type_ids.transpose(0, 1).contiguous() if token_type_ids is not None else None
input_mask = input_mask.transpose(0, 1).contiguous() if input_mask is not None else None
attention_mask = attention_mask.transpose(0, 1).contiguous() if attention_mask is not None else None
perm_mask = perm_mask.permute(1, 2, 0).contiguous() if perm_mask is not None else None
target_mapping = target_mapping.permute(1, 2, 0).contiguous() if target_mapping is not None else None
mlen = mems[0].shape[0] if mems is not None and mems[0] is not None else 0
klen = mlen + qlen
dtype_float = next(self.parameters()).dtype
device = next(self.parameters()).device
# Attention mask
# causal attention mask
if self.attn_type == "uni":
attn_mask = self.create_mask(qlen, mlen)
attn_mask = attn_mask[:, :, None, None]
elif self.attn_type == "bi":
attn_mask = None
else:
raise ValueError("Unsupported attention type: {}".format(self.attn_type))
# data mask: input mask & perm mask
assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
"or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
if input_mask is None and attention_mask is not None:
input_mask = 1.0 - attention_mask
if input_mask is not None and perm_mask is not None:
data_mask = input_mask[None] + perm_mask
elif input_mask is not None and perm_mask is None:
data_mask = input_mask[None]
elif input_mask is None and perm_mask is not None:
data_mask = perm_mask
else:
data_mask = None
if data_mask is not None:
# all mems can be attended to
if mlen > 0:
mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask)
data_mask = torch.cat([mems_mask, data_mask], dim=1)
if attn_mask is None:
attn_mask = data_mask[:, :, :, None]
else:
attn_mask += data_mask[:, :, :, None]
if attn_mask is not None:
attn_mask = (attn_mask > 0).to(dtype_float)
if attn_mask is not None:
non_tgt_mask = -torch.eye(qlen).to(attn_mask)
if mlen > 0:
non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1)
non_tgt_mask = ((attn_mask + non_tgt_mask[:, :, None, None]) > 0).to(attn_mask)
else:
non_tgt_mask = None
# Word embeddings and prepare h & g hidden states
if inputs_embeds is not None:
word_emb_k = inputs_embeds
else:
word_emb_k = self.word_embedding(input_ids)
output_h = self.dropout(word_emb_k)
if target_mapping is not None:
word_emb_q = self.mask_emb.expand(target_mapping.shape[0], bsz, -1)
# else: # We removed the inp_q input which was same as target mapping
# inp_q_ext = inp_q[:, :, None]
# word_emb_q = inp_q_ext * self.mask_emb + (1 - inp_q_ext) * word_emb_k
output_g = self.dropout(word_emb_q)
else:
output_g = None
# Segment embedding
if token_type_ids is not None:
# Convert `token_type_ids` to one-hot `seg_mat`
if mlen > 0:
mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device)
cat_ids = torch.cat([mem_pad, token_type_ids], dim=0)
else:
cat_ids = token_type_ids
# `1` indicates not in the same segment [qlen x klen x bsz]
seg_mat = (token_type_ids[:, None] != cat_ids[None, :]).long()
seg_mat = F.one_hot(seg_mat, num_classes=2).to(dtype_float)
else:
seg_mat = None
# Positional encoding
pos_emb = self.relative_positional_encoding(qlen, klen, bsz=bsz)
pos_emb = self.dropout(pos_emb)
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] (a head_mask for each layer)
# and head_mask is converted to shape [num_hidden_layers x qlen x klen x bsz x n_head]
if head_mask is not None:
if head_mask.dim() == 1:
head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0).unsqueeze(0)
head_mask = head_mask.expand(self.n_layer, -1, -1, -1, -1)
elif head_mask.dim() == 2:
head_mask = head_mask.unsqueeze(1).unsqueeze(1).unsqueeze(1)
head_mask = head_mask.to(
dtype=next(self.parameters()).dtype
) # switch to fload if need + fp16 compatibility
else:
head_mask = [None] * self.n_layer
new_mems = ()
if mems is None:
mems = [None] * len(self.layer)
attentions = []
hidden_states = []
for i, layer_module in enumerate(self.layer):
if self.mem_len is not None and self.mem_len > 0 and use_cache is True:
# cache new mems
new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
if self.output_hidden_states:
hidden_states.append((output_h, output_g) if output_g is not None else output_h)
outputs = layer_module(
output_h,
output_g,
attn_mask_h=non_tgt_mask,
attn_mask_g=attn_mask,
r=pos_emb,
seg_mat=seg_mat,
mems=mems[i],
target_mapping=target_mapping,
head_mask=head_mask[i],
)
output_h, output_g = outputs[:2]
if self.output_attentions:
attentions.append(outputs[2])
# Add last hidden state
if self.output_hidden_states:
hidden_states.append((output_h, output_g) if output_g is not None else output_h)
output = self.dropout(output_g if output_g is not None else output_h)
# Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
outputs = (output.permute(1, 0, 2).contiguous(),)
if self.mem_len is not None and self.mem_len > 0 and use_cache is True:
outputs = outputs + (new_mems,)
if self.output_hidden_states:
if output_g is not None:
hidden_states = tuple(h.permute(1, 0, 2).contiguous() for hs in hidden_states for h in hs)
else:
hidden_states = tuple(hs.permute(1, 0, 2).contiguous() for hs in hidden_states)
outputs = outputs + (hidden_states,)
if self.output_attentions:
if target_mapping is not None:
# when target_mapping is provided, there are 2-tuple of attentions
attentions = tuple(
tuple(att_stream.permute(2, 3, 0, 1).contiguous() for att_stream in t) for t in attentions
)
else:
attentions = tuple(t.permute(2, 3, 0, 1).contiguous() for t in attentions)
outputs = outputs + (attentions,)
return outputs # outputs, (new_mems), (hidden_states), (attentions)
[docs]@add_start_docstrings(
"""XLNet Model with a language modeling head on top
(linear layer with weights tied to the input embeddings). """,
XLNET_START_DOCSTRING,
)
class XLNetLMHeadModel(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.attn_type = config.attn_type
self.same_length = config.same_length
self.transformer = XLNetModel(config)
self.lm_loss = nn.Linear(config.d_model, config.vocab_size, bias=True)
self.init_weights()
[docs] def get_output_embeddings(self):
return self.lm_loss
def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
# Add dummy token at the end (no attention on this one)
effective_batch_size = input_ids.shape[0]
dummy_token = torch.zeros((effective_batch_size, 1), dtype=torch.long, device=input_ids.device)
input_ids = torch.cat([input_ids, dummy_token], dim=1)
# Build permutation mask so that previous tokens don't see last token
sequence_length = input_ids.shape[1]
perm_mask = torch.zeros(
(effective_batch_size, sequence_length, sequence_length), dtype=torch.float, device=input_ids.device
)
perm_mask[:, :, -1] = 1.0
# We'll only predict the last token
target_mapping = torch.zeros(
(effective_batch_size, 1, sequence_length), dtype=torch.float, device=input_ids.device
)
target_mapping[0, 0, -1] = 1.0
inputs = {
"input_ids": input_ids,
"perm_mask": perm_mask,
"target_mapping": target_mapping,
"use_cache": kwargs["use_cache"],
}
# if past is defined in model kwargs then use it for faster decoding
if past:
inputs["mems"] = past
return inputs
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
labels=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, num_predict)`, `optional`, defaults to :obj:`None`):
Labels for masked language modeling.
`num_predict` corresponds to `target_mapping.shape[1]`. If `target_mapping` is `None`, then `num_predict` corresponds to `sequence_length`.
The labels should correspond to the masked input words that should be predicted and depends on `target_mapping`. Note in order to perform standard auto-regressive language modeling a `<mask>` token has to be added to the `input_ids` (see `prepare_inputs_for_generation` fn and examples below)
Indices are selected in ``[-100, 0, ..., config.vocab_size]``
All labels set to ``-100`` are ignored, the loss is only
computed for labels in ``[0, ..., config.vocab_size]``
Return:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when ``labels`` is provided)
Language modeling loss.
prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_predict, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
`num_predict` corresponds to `target_mapping.shape[1]`. If `target_mapping` is `None`, then `num_predict` corresponds to `sequence_length`.
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetLMHeadModel
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetLMHeadModel.from_pretrained('xlnet-large-cased')
# We show how to setup inputs to predict a next token using a bi-directional context.
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=False)).unsqueeze(0) # We will predict the masked token
perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token
target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float) # Shape [1, 1, seq_length] => let's predict one token
target_mapping[0, 0, -1] = 1.0 # Our first (and only) prediction will be the last token of the sequence (the masked token)
outputs = model(input_ids, perm_mask=perm_mask, target_mapping=target_mapping)
next_token_logits = outputs[0] # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]
# The same way can the XLNetLMHeadModel be used to be trained by standard auto-regressive language modeling.
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>", add_special_tokens=False)).unsqueeze(0) # We will predict the masked token
labels = torch.tensor(tokenizer.encode("cute", add_special_tokens=False)).unsqueeze(0)
assert labels.shape[0] == 1, 'only one word will be predicted'
perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token as is done in standard auto-regressive lm training
target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float) # Shape [1, 1, seq_length] => let's predict one token
target_mapping[0, 0, -1] = 1.0 # Our first (and only) prediction will be the last token of the sequence (the masked token)
outputs = model(input_ids, perm_mask=perm_mask, target_mapping=target_mapping, labels=labels)
loss, next_token_logits = outputs[:2] # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]
"""
transformer_outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
mems=mems,
perm_mask=perm_mask,
target_mapping=target_mapping,
token_type_ids=token_type_ids,
input_mask=input_mask,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
)
logits = self.lm_loss(transformer_outputs[0])
outputs = (logits,) + transformer_outputs[1:] # Keep mems, hidden states, attentions if there are in it
if labels is not None:
# Flatten the tokens
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
outputs = (loss,) + outputs
return outputs # return (loss), logits, (mems), (hidden states), (attentions)
[docs]@add_start_docstrings(
"""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
the pooled output) e.g. for GLUE tasks. """,
XLNET_START_DOCSTRING,
)
class XLNetForSequenceClassification(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = XLNetModel(config)
self.sequence_summary = SequenceSummary(config)
self.logits_proj = nn.Linear(config.d_model, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
labels=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`)
Labels for computing the sequence classification/regression loss.
Indices should be in ``[0, ..., config.num_labels - 1]``.
If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
Return:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
Classification (or regression if config.num_labels==1) loss.
logits (:obj:`torch.FloatTensor` of shape :obj:(batch_size, config.num_labels)`):
Classification (or regression if config.num_labels==1) scores (before SoftMax).
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetForSequenceClassification
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
loss, logits = outputs[:2]
"""
transformer_outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
mems=mems,
perm_mask=perm_mask,
target_mapping=target_mapping,
token_type_ids=token_type_ids,
input_mask=input_mask,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
)
output = transformer_outputs[0]
output = self.sequence_summary(output)
logits = self.logits_proj(output)
outputs = (logits,) + transformer_outputs[1:] # Keep mems, hidden states, attentions if there are in it
if labels is not None:
if self.num_labels == 1:
# We are doing regression
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
else:
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
outputs = (loss,) + outputs
return outputs # return (loss), logits, (mems), (hidden states), (attentions)
[docs]@add_start_docstrings(
"""XLNet Model with a token classification head on top (a linear layer on top of
the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
XLNET_START_DOCSTRING,
)
class XLNetForTokenClassification(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = XLNetModel(config)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
labels=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
Labels for computing the multiple choice classification loss.
Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
of the input tensors. (see `input_ids` above)
Return:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
Classification loss.
logits (:obj:`torch.FloatTensor` of shape :obj:(batch_size, config.num_labels)`):
Classification scores (before SoftMax).
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetForTokenClassification
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
model = XLNetForTokenClassification.from_pretrained('xlnet-large-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0) # Batch size 1
outputs = model(input_ids, labels=labels)
scores = outputs[0]
"""
outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
mems=mems,
perm_mask=perm_mask,
target_mapping=target_mapping,
token_type_ids=token_type_ids,
input_mask=input_mask,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output)
outputs = (logits,) + outputs[1:] # Keep mems, hidden states, attentions if there are in it
if labels is not None:
loss_fct = CrossEntropyLoss()
# Only keep active parts of the loss
if attention_mask is not None:
active_loss = attention_mask.view(-1) == 1
active_logits = logits.view(-1, self.num_labels)
active_labels = torch.where(
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
)
loss = loss_fct(active_logits, active_labels)
else:
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
outputs = (loss,) + outputs
return outputs # return (loss), logits, (mems), (hidden states), (attentions)
[docs]@add_start_docstrings(
"""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
the hidden-states output to compute `span start logits` and `span end logits`). """,
XLNET_START_DOCSTRING,
)
class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = XLNetModel(config)
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
start_positions=None,
end_positions=None,
):
r"""
start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
Returns:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
start_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
Span-start scores (before SoftMax).
end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
Span-end scores (before SoftMax).
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetForQuestionAnsweringSimple
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
model = XLNetForQuestionAnsweringSimple.from_pretrained('xlnet-base-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
start_positions = torch.tensor([1])
end_positions = torch.tensor([3])
outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
loss = outputs[0]
"""
outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
mems=mems,
perm_mask=perm_mask,
target_mapping=target_mapping,
token_type_ids=token_type_ids,
input_mask=input_mask,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = logits.split(1, dim=-1)
start_logits = start_logits.squeeze(-1)
end_logits = end_logits.squeeze(-1)
outputs = (start_logits, end_logits,) + outputs[2:]
if start_positions is not None and end_positions is not None:
# If we are on multi-GPU, split add a dimension
if len(start_positions.size()) > 1:
start_positions = start_positions.squeeze(-1)
if len(end_positions.size()) > 1:
end_positions = end_positions.squeeze(-1)
# sometimes the start/end positions are outside our model inputs, we ignore these terms
ignored_index = start_logits.size(1)
start_positions.clamp_(0, ignored_index)
end_positions.clamp_(0, ignored_index)
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
start_loss = loss_fct(start_logits, start_positions)
end_loss = loss_fct(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2
outputs = (total_loss,) + outputs
return outputs # (loss), start_logits, end_logits, (mems), (hidden_states), (attentions)
[docs]@add_start_docstrings(
"""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
the hidden-states output to compute `span start logits` and `span end logits`). """,
XLNET_START_DOCSTRING,
)
class XLNetForQuestionAnswering(XLNetPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.start_n_top = config.start_n_top
self.end_n_top = config.end_n_top
self.transformer = XLNetModel(config)
self.start_logits = PoolerStartLogits(config)
self.end_logits = PoolerEndLogits(config)
self.answer_class = PoolerAnswerClass(config)
self.init_weights()
[docs] @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
def forward(
self,
input_ids=None,
attention_mask=None,
mems=None,
perm_mask=None,
target_mapping=None,
token_type_ids=None,
input_mask=None,
head_mask=None,
inputs_embeds=None,
use_cache=True,
start_positions=None,
end_positions=None,
is_impossible=None,
cls_index=None,
p_mask=None,
):
r"""
start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
is_impossible (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
Labels whether a question has an answer or no answer (SQuAD 2.0)
cls_index (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
Labels for position (index) of the classification token to use as input for computing plausibility of the answer.
p_mask (``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``, `optional`, defaults to :obj:`None`):
Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...).
1.0 means token should be masked. 0.0 mean token is not masked.
Returns:
:obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned if both :obj:`start_positions` and :obj:`end_positions` are provided):
Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
start_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Log probabilities for the top config.start_n_top start token possibilities (beam-search).
start_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Indices for the top config.start_n_top start token possibilities (beam-search).
end_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
end_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
cls_logits (``torch.FloatTensor`` of shape ``(batch_size,)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Log probabilities for the ``is_impossible`` label of the answers.
mems (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
Contains pre-computed hidden-states (key and values in the attention blocks).
Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
should not be passed as input ids as they have already been computed.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (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 model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
Examples::
from transformers import XLNetTokenizer, XLNetForQuestionAnswering
import torch
tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
model = XLNetForQuestionAnswering.from_pretrained('xlnet-base-cased')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1
start_positions = torch.tensor([1])
end_positions = torch.tensor([3])
outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
loss = outputs[0]
"""
transformer_outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
mems=mems,
perm_mask=perm_mask,
target_mapping=target_mapping,
token_type_ids=token_type_ids,
input_mask=input_mask,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
)
hidden_states = transformer_outputs[0]
start_logits = self.start_logits(hidden_states, p_mask=p_mask)
outputs = transformer_outputs[1:] # Keep mems, hidden states, attentions if there are in it
if start_positions is not None and end_positions is not None:
# If we are on multi-GPU, let's remove the dimension added by batch splitting
for x in (start_positions, end_positions, cls_index, is_impossible):
if x is not None and x.dim() > 1:
x.squeeze_(-1)
# during training, compute the end logits based on the ground truth of the start position
end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)
loss_fct = CrossEntropyLoss()
start_loss = loss_fct(start_logits, start_positions)
end_loss = loss_fct(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2
if cls_index is not None and is_impossible is not None:
# Predict answerability from the representation of CLS and START
cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
loss_fct_cls = nn.BCEWithLogitsLoss()
cls_loss = loss_fct_cls(cls_logits, is_impossible)
# note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
total_loss += cls_loss * 0.5
outputs = (total_loss,) + outputs
else:
# during inference, compute the end logits based on beam search
bsz, slen, hsz = hidden_states.size()
start_log_probs = F.softmax(start_logits, dim=-1) # shape (bsz, slen)
start_top_log_probs, start_top_index = torch.topk(
start_log_probs, self.start_n_top, dim=-1
) # shape (bsz, start_n_top)
start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz) # shape (bsz, start_n_top, hsz)
start_states = torch.gather(hidden_states, -2, start_top_index_exp) # shape (bsz, start_n_top, hsz)
start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1) # shape (bsz, slen, start_n_top, hsz)
hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(
start_states
) # shape (bsz, slen, start_n_top, hsz)
p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
end_log_probs = F.softmax(end_logits, dim=1) # shape (bsz, slen, start_n_top)
end_top_log_probs, end_top_index = torch.topk(
end_log_probs, self.end_n_top, dim=1
) # shape (bsz, end_n_top, start_n_top)
end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)
start_states = torch.einsum(
"blh,bl->bh", hidden_states, start_log_probs
) # get the representation of START as weighted sum of hidden states
cls_logits = self.answer_class(
hidden_states, start_states=start_states, cls_index=cls_index
) # Shape (batch size,): one single `cls_logits` for each sample
outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits) + outputs
# return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
# or (if labels are provided) (total_loss,)
return outputs