Implementation of the language models¶

In [ ]:

from fastai.gen_doc.nbdoc import *
from fastai.text import * 
from fastai.text.models import * 

text.models module fully implements the encoder for an AWD-LSTM, the transformer model and the transformer XL model. They can then plugged in with a decoder to make a language model, or some classifying layers to make a text classifier.

Language model modules¶

In [ ]:

show_doc(AWD_LSTM, title_level=3)

`class` `AWD_LSTM`[source]

AWD_LSTM(vocab_sz:int, emb_sz:int, n_hid:int, n_layers:int, pad_token:int=*1, hidden_p:float=0.2, input_p:float=0.6, embed_p:float=0.1, weight_p:float=0.5, qrnn:bool=False*) :: Module

AWD-LSTM/QRNN inspired by https://arxiv.org/abs/1708.02182.

The main idea of the article is to use a RNN with dropout everywhere, but in an intelligent way. There is a difference with the usual dropout, which is why you’ll see a RNNDropout module: we zero things, as is usual in dropout, but we always zero the same thing according to the sequence dimension (which is the first dimension in pytorch). This ensures consistency when updating the hidden state through the whole sentences/articles.

This being given, there are a total four different dropouts in the encoder of the AWD-LSTM:

the first one, embedding dropout, is applied when we look the ids of our tokens inside the embedding matrix (to transform them from numbers to a vector of float). We zero some lines of it, so random ids are sent to a vector of zeros instead of being sent to their embedding vector. This is the embed_p parameter.
the second one, input dropout, is applied to the result of the embedding with dropout. We forget random pieces of the embedding matrix (but as stated in the last paragraph, the same ones in the sequence dimension). This is the input_p parameter.
the third one is the weight dropout. It’s the trickiest to implement as we randomly replace by 0s some weights of the hidden-to-hidden matrix inside the RNN: this needs to be done in a way that ensure the gradients are still computed and the initial weights still updated. This is the weight_p parameter.
the fourth one is the hidden dropout. It’s applied to the output of one of the layers of the RNN before it’s used as input of the next layer (again same coordinates are zeroed in the sequence dimension). It isn’t applied to the last output (which will get its own dropout in the decoder).This is the hidden_p parameter.

The other attributes are vocab_sz for the number of tokens in your vocabulary, emb_sz for the embedding size, n_hid for the hidden size of your inner LSTMs (or QRNNs), n_layers the number of layers and pad_token for the index of an eventual padding token (1 by default in fastai).

The flag qrnn=True replace the inner LSTMs by QRNNs.

In [ ]:

show_doc(AWD_LSTM.reset)

`reset`[source]

reset()

Reset the hidden states.

In [ ]:

show_doc(Transformer, title_level=3)

`class` `Transformer`[source]

Transformer(vocab_sz:int, ctx_len:int, n_layers:int, n_heads:int, d_model:int, d_head:int, d_inner:int, resid_p:float=*0.0, attn_p:float=0.0, ff_p:float=0.0, embed_p:float=0.0, bias:bool=True, scale:bool=True, act:Activation=<Activation.ReLU: 1>, double_drop:bool=True, attn_cls:Callable='MultiHeadAttention', learned_pos_enc:bool=True, mask:bool=True*) :: Module

Transformer model: https://arxiv.org/abs/1706.03762.

The main idea of this article is to use regular neural net for NLP instead of an RNN, but with lots of attention layers. Intuitively, those attention layers tell the model to pay more interest to this or that world when trying to predict its output.

It starts from embeddings from vocab_sz (number of tokens) to d_model (which is basically the hidden size throughout the model), and it will look at inputs of size batch_size by ctx_len (for context length). We add a positional encoding to the embeddings (since a regular neural net has no idea of the order of words), either learned or coming from PositionalEncoding depending on learned_pos_enc. We then have a dropout of embed_p followed by n_layers blocks of MultiHeadAttention followed by feed_forward.

In the attention we use n_heads with each a hidden state of d_head (will default to d_model//n_heads). If mask=True, a mask will make sure no attention is paid to future tokens (which would be cheating when training a language model). If scale=True, the attention scores are scaled by a factor 1 / math.sqrt(d_head). A dropout of attn_p is applied to the attention scores, then the final result get applied a dropout of resid_p before being summed to the original input (residual connection before the layer norm).

In feed forward, we have two linear layers from d_model to d_inner and then back. Those have bias if that flag is True and a dropout of ff_p is applied, after each if double_drop=True, or just at the end otherwise. act is used in the middle as a non-linearity.

In [ ]:

show_doc(TransformerXL, title_level=3)

`class` `TransformerXL`[source]

TransformerXL(vocab_sz:int, ctx_len:int, n_layers:int, n_heads:int, d_model:int, d_head:int, d_inner:int, resid_p:float=*0.0, attn_p:float=0.0, ff_p:float=0.0, embed_p:float=0.0, bias:bool=False, scale:bool=True, act:Activation=<Activation.ReLU: 1>, double_drop:bool=True, attn_cls:Callable='MultiHeadRelativeAttention', learned_pos_enc:bool=False, mask:bool=True, mem_len:int=0*) :: Module

TransformerXL model: https://arxiv.org/abs/1901.02860.

TransformerXL is a transformer architecture with a sort of hidden state formed by the results of the intermediate layers on previous tokens. Its size is determined by mem_len. By using this context, those models are capable of learning longer dependencies and can also be used for faster text generation at inference: a regular transformer model would have to reexamine the whole of sequence of indexes generated so far, whereas we can feed the new tokens one by one to a transformer XL (like we do with a regular RNN).

In [ ]:

show_doc(TransformerXL.reset)

`reset`[source]

reset()

Reset the internal memory.

Decoders¶

In [ ]:

show_doc(LinearDecoder, title_level=3)

`class` `LinearDecoder`[source]

LinearDecoder(n_out:int, n_hid:int, output_p:float, tie_encoder:Module=*None, bias:bool=True*) :: Module

To go on top of a RNNCore module and create a Language Model.

Create a the decoder to go on top of an RNNCore encoder and create a language model. n_hid is the dimension of the last hidden state of the encoder, n_out the size of the output. Dropout of output_p is applied. If a tie_encoder is passed, it will be used for the weights of the linear layer, that will have bias or not.

In [ ]:

show_doc(PoolingLinearClassifier, title_level=3)

`class` `PoolingLinearClassifier`[source]

PoolingLinearClassifier(layers:Collection[int], drops:Collection[float]) :: Module

Create a linear classifier with pooling.

The last output, MaxPooling of all the outputs and AvgPooling of all the outputs are concatenated, then blocks of bn_drop_lin are stacked, according to the values in layers and drops.

In [ ]:

show_doc(PoolingLinearClassifier.pool)

`pool`[source]

pool(x:Tensor, bs:int, is_max:bool)

Pool the tensor along the seq_len dimension.

The input tensor x (of batch size bs) is pooled along the batch dimension. is_max decides if we do an AvgPooling or a MaxPooling.

Basic NLP modules¶

On top of the pytorch or the fastai layers, the language models use some custom layers specific to NLP.

In [ ]:

show_doc(EmbeddingDropout, title_level=3)

`class` `EmbeddingDropout`[source]

EmbeddingDropout(emb:Module, embed_p:float) :: Module

Apply dropout with probabily embed_p to an embedding layer emb.

Each row of the embedding matrix has a probability embed_p of being replaced by zeros while the others are rescaled accordingly.

In [ ]:

enc = nn.Embedding(100, 7, padding_idx=1)
enc_dp = EmbeddingDropout(enc, 0.5)
tst_input = torch.randint(0,100,(8,))
enc_dp(tst_input)

Out[ ]:

tensor([[-0.7379, -1.3970, -0.4075, -0.1676,  2.0396,  3.2226,  0.7128],
        [-0.0000,  0.0000,  0.0000, -0.0000, -0.0000,  0.0000,  0.0000],
        [-3.2579,  2.2972, -1.8704, -0.4090,  2.6477, -1.5015,  0.7158],
        [ 2.1455,  1.0571, -0.6086,  3.5700,  2.6271, -3.1353,  0.7277],
        [-3.7003, -1.8846,  0.2029, -0.6839,  0.2968, -2.0199,  1.3127],
        [-0.0000,  0.0000, -0.0000, -0.0000,  0.0000,  0.0000, -0.0000],
        [-0.0051,  2.7428,  3.0068,  0.6242,  1.2747,  0.9262,  0.4070],
        [ 1.9312,  3.0524, -1.2806,  1.5910, -2.1789, -0.1636, -3.4924]],
       grad_fn=<EmbeddingBackward>)

In [ ]:

show_doc(RNNDropout, title_level=3)

`class` `RNNDropout`[source]

RNNDropout(p:float=*0.5*) :: Module

Dropout with probability p that is consistent on the seq_len dimension.

In [ ]:

dp = RNNDropout(0.3)
tst_input = torch.randn(3,3,7)
tst_input, dp(tst_input)

Out[ ]:

(tensor([[[-2.1156,  0.9734,  0.2428,  0.9396,  0.4072, -0.8197,  0.3718],
          [ 0.4838,  1.3077, -0.8239, -0.6557,  1.3938,  0.6086, -0.2622],
          [ 0.2372, -0.1627,  0.3117, -0.4811, -1.0841, -0.5207, -0.5131]],
 
         [[-0.6924,  0.4122,  0.2517, -1.0120,  0.6808,  0.8800, -0.7463],
          [-0.9498,  0.7655,  0.7471, -0.2767,  1.2155, -0.1042, -2.1443],
          [-1.2342,  1.9187, -0.8481, -0.4115, -1.3223,  1.4266, -1.4150]],
 
         [[ 0.1539,  0.3142,  0.2158,  1.1411,  0.1316,  0.6158, -1.5078],
          [-1.0177, -0.9230,  0.9994,  0.1140,  0.7432,  0.4353,  0.0096],
          [-0.8231,  1.0086,  1.7685,  0.3304, -0.0896, -1.0513, -1.3017]]]),
 tensor([[[-3.0223,  1.3905,  0.0000,  0.0000,  0.5818, -0.0000,  0.5312],
          [ 0.6911,  1.8681, -0.0000, -0.0000,  1.9911,  0.0000, -0.3745],
          [ 0.3389, -0.2324,  0.0000, -0.0000, -1.5487, -0.0000, -0.7331]],
 
         [[-0.9892,  0.5889,  0.3596, -1.4458,  0.9725,  1.2571, -0.0000],
          [-1.3569,  1.0936,  1.0673, -0.3953,  1.7364, -0.1489, -0.0000],
          [-1.7631,  2.7410, -1.2116, -0.5879, -1.8889,  2.0380, -0.0000]],
 
         [[ 0.0000,  0.4489,  0.0000,  1.6301,  0.1880,  0.8797, -2.1539],
          [-0.0000, -1.3186,  0.0000,  0.1628,  1.0617,  0.6218,  0.0137],
          [-0.0000,  1.4408,  0.0000,  0.4720, -0.1280, -1.5019, -1.8595]]]))

In [ ]:

show_doc(WeightDropout, title_level=3)

`class` `WeightDropout`[source]

WeightDropout(module:Module, weight_p:float, layer_names:StrList=*['weight_hh_l0']*) :: Module

A module that warps another layer in which some weights will be replaced by 0 during training.

Applies dropout of probability weight_p to the layers in layer_names of module in training mode. A copy of those weights is kept so that the dropout mask can change at every batch.

In [ ]:

module = nn.LSTM(5, 2)
dp_module = WeightDropout(module, 0.4)
getattr(dp_module.module, 'weight_hh_l0')

Out[ ]:

Parameter containing:
tensor([[-0.0702,  0.5725],
        [-0.3910,  0.6512],
        [-0.2203, -0.4315],
        [ 0.2750, -0.2917],
        [-0.4890, -0.3094],
        [ 0.4638, -0.3807],
        [-0.2290, -0.6964],
        [ 0.1224,  0.4043]], requires_grad=True)

It's at the beginning of a forward pass that the dropout is applied to the weights.

In [ ]:

tst_input = torch.randn(4,20,5)
h = (torch.zeros(1,20,2), torch.zeros(1,20,2))
x,h = dp_module(tst_input,h)
getattr(dp_module.module, 'weight_hh_l0')

Out[ ]:

tensor([[-0.0000,  0.0000],
        [-0.6517,  0.0000],
        [-0.0000, -0.7191],
        [ 0.4583, -0.0000],
        [-0.0000, -0.0000],
        [ 0.7730, -0.6345],
        [-0.0000, -1.1607],
        [ 0.2040,  0.6739]], grad_fn=<MulBackward0>)

In [ ]:

show_doc(PositionalEncoding, title_level=3)

`class` `PositionalEncoding`[source]

PositionalEncoding(d:int) :: Module

Encode the position with a sinusoid.

In [ ]:

show_doc(DecoderLayer, title_level=3)

`class` `DecoderLayer`[source]

DecoderLayer(n_heads:int, d_model:int, d_head:int, d_inner:int, resid_p:float=*0.0, attn_p:float=0.0, ff_p:float=0.0, bias:bool=True, scale:bool=True, act:Activation=<Activation.ReLU: 1>, double_drop:bool=True, attn_cls:Callable='MultiHeadAttention'*) :: Module

Basic block of a Transformer model.

In [ ]:

show_doc(MultiHeadAttention, title_level=3)

`class` `MultiHeadAttention`[source]

MultiHeadAttention(n_heads:int, d_model:int, d_head:int=*None, resid_p:float=0.0, attn_p:float=0.0, bias:bool=True, scale:bool=True*) :: Module

MutiHeadAttention.

In [ ]:

show_doc(MultiHeadRelativeAttention, title_level=3)

`class` `MultiHeadRelativeAttention`[source]

MultiHeadRelativeAttention(n_heads:int, d_model:int, d_head:int, resid_p:float=*0.0, attn_p:float=0.0, bias:bool=True, scale:bool=True*) :: MultiHeadAttention

MutiHeadAttention with relative positional encoding.

In [ ]:

show_doc(SequentialRNN, title_level=3)

`class` `SequentialRNN`[source]

SequentialRNN(***args**) :: Sequential

A sequential module that passes the reset call to its children.

In [ ]:

show_doc(SequentialRNN.reset)

`reset`[source]

reset()

Call the reset function of self.children (if they have one).

In [ ]:

show_doc(dropout_mask)

`dropout_mask`[source]

dropout_mask(x:Tensor, sz:Collection[int], p:float)

Return a dropout mask of the same type as x, size sz, with probability p to cancel an element.

In [ ]:

tst_input = torch.randn(3,3,7)
dropout_mask(tst_input, (3,7), 0.3)

Out[ ]:

tensor([[0.0000, 1.4286, 1.4286, 0.0000, 1.4286, 1.4286, 0.0000],
        [1.4286, 1.4286, 1.4286, 0.0000, 1.4286, 0.0000, 0.0000],
        [1.4286, 0.0000, 1.4286, 0.0000, 0.0000, 0.0000, 1.4286]])

Such a mask is then expanded in the sequence length dimension and multiplied by the input to do an RNNDropout.

In [ ]:

show_doc(feed_forward)

`feed_forward`[source]

feed_forward(d_model:int, d_ff:int, ff_p:float=*0.0, act:Activation=<Activation.ReLU: 1>, double_drop:bool=True*)

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

In [ ]:

show_doc(WeightDropout.forward)

`forward`[source]

forward(***args**:ArgStar)

Defines the computation performed at every call. Should be overridden by all subclasses.

.. note:: Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

In [ ]:

show_doc(EmbeddingDropout.forward)

`forward`[source]

forward(words:LongTensor, scale:Optional[float]=*None*) → Tensor

Defines the computation performed at every call. Should be overridden by all subclasses.

.. note:: Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

In [ ]:

show_doc(RNNDropout.forward)

`forward`[source]

forward(x:Tensor) → Tensor

Defines the computation performed at every call. Should be overridden by all subclasses.

.. note:: Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

In [ ]:

show_doc(WeightDropout.reset)

`reset`[source]

reset()

In [ ]:

show_doc(PoolingLinearClassifier.forward)

`forward`[source]

forward(input:Tuple[Tensor, Tensor]) → Tuple[Tensor, Tensor, Tensor]

Defines the computation performed at every call. Should be overridden by all subclasses.

.. note:: Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

In [ ]:

show_doc(LinearDecoder.forward)

`forward`[source]

forward(input:Tuple[Tensor, Tensor]) → Tuple[Tensor, Tensor, Tensor]

Defines the computation performed at every call. Should be overridden by all subclasses.

.. note:: Although the recipe for forward pass needs to be defined within this function, one should call the :class:Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Implementation of the language models¶

Language model modules¶

class AWD_LSTM[source]

reset[source]

class Transformer[source]

class TransformerXL[source]

reset[source]

Decoders¶

class LinearDecoder[source]

class PoolingLinearClassifier[source]

pool[source]

Basic NLP modules¶

class EmbeddingDropout[source]

class RNNDropout[source]

class WeightDropout[source]

class PositionalEncoding[source]

class DecoderLayer[source]

class MultiHeadAttention[source]

class MultiHeadRelativeAttention[source]

class SequentialRNN[source]

reset[source]

dropout_mask[source]

feed_forward[source]

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

forward[source]

forward[source]

forward[source]

reset[source]

forward[source]

forward[source]

New Methods - Please document or move to the undocumented section¶

`class` `AWD_LSTM`[source]

`reset`[source]

`class` `Transformer`[source]

`class` `TransformerXL`[source]

`reset`[source]

`class` `LinearDecoder`[source]

`class` `PoolingLinearClassifier`[source]

`pool`[source]

`class` `EmbeddingDropout`[source]

`class` `RNNDropout`[source]

`class` `WeightDropout`[source]

`class` `PositionalEncoding`[source]

`class` `DecoderLayer`[source]

`class` `MultiHeadAttention`[source]

`class` `MultiHeadRelativeAttention`[source]

`class` `SequentialRNN`[source]

`reset`[source]

`dropout_mask`[source]

`feed_forward`[source]

`forward`[source]

`forward`[source]

`forward`[source]

`reset`[source]

`forward`[source]

`forward`[source]