# recbole.model.layers¶

Common Layers in recommender system

class recbole.model.layers.AttLayer(in_dim, att_dim)[source]

Bases: torch.nn.modules.module.Module

Calculate the attention signal(weight) according the input tensor.

Parameters

infeatures (torch.FloatTensor) – A 3D input tensor with shape of[batch_size, M, embed_dim].

Returns

Attention weight of input. shape of [batch_size, M].

Return type

torch.FloatTensor

forward(infeatures)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.BaseFactorizationMachine(reduce_sum=True)[source]

Bases: torch.nn.modules.module.Module

Calculate FM result over the embeddings

Parameters

reduce_sum – bool, whether to sum the result, default is True.

Input:

input_x: tensor, A 3D tensor with shape:(batch_size,field_size,embed_dim).

Output

output: tensor, A 3D tensor with shape: (batch_size,1) or (batch_size, embed_dim).

forward(input_x)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.BiGNNLayer(in_dim, out_dim)[source]

Bases: torch.nn.modules.module.Module

Propagate a layer of Bi-interaction GNN

$output = (L+I)EW_1 + LE \otimes EW_2$
forward(lap_matrix, eye_matrix, features)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.CNNLayers(channels, kernels, strides, activation='relu', init_method=None)[source]

Bases: torch.nn.modules.module.Module

Parameters
• channels (-) – a list contains the channels of each layer in cnn layers

• kernel (-) – a list contains the kernels of each layer in cnn layers

• strides (-) – a list contains the channels of each layer in cnn layers

• activation (-) – activation function after each layer in mlp layers. Default: ‘relu’ candidates: ‘sigmoid’, ‘tanh’, ‘relu’, ‘leekyrelu’, ‘none’

Shape:
• Input: $$(N, C_{in}, H_{in}, W_{in})$$

• Output: $$(N, C_{out}, H_{out}, W_{out})$$ where

$H_{out} = \left\lfloor\frac{H_{in} + 2 \times \text{padding}[0] - \text{dilation}[0] \times (\text{kernel\_size}[0] - 1) - 1}{\text{stride}[0]} + 1\right\rfloor$
$W_{out} = \left\lfloor\frac{W_{in} + 2 \times \text{padding}[1] - \text{dilation}[1] \times (\text{kernel\_size}[1] - 1) - 1}{\text{stride}[1]} + 1\right\rfloor$

Examples:

>>> m = CNNLayers([1, 32, 32], [2,2], [2,2], 'relu')
>>> input = torch.randn(128, 1, 64, 64)
>>> output = m(input)
>>> print(output.size())
>>> torch.Size([128, 32, 16, 16])
forward(input_feature)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

init_weights(module)[source]
class recbole.model.layers.ContextSeqEmbAbstractLayer[source]

Bases: torch.nn.modules.module.Module

For Deep Interest Network and feature-rich sequential recommender systems, return features embedding matrices.

embed_float_fields(float_fields, type, embed=True)[source]

Get the embedding of float fields. In the following three functions(“embed_float_fields” “embed_token_fields” “embed_token_seq_fields”) when the type is user, [batch_size, max_item_length] should be recognised as [batch_size]

Parameters
• float_fields (torch.Tensor) – [batch_size, max_item_length, num_float_field]

• type (str) – user or item

• embed (bool) – embed or not

Returns

float fields embedding. [batch_size, max_item_length, num_float_field, embed_dim]

Return type

torch.Tensor

embed_input_fields(user_idx, item_idx)[source]

Get the embedding of user_idx and item_idx

Parameters
• user_idx (torch.Tensor) – interaction[‘user_id’]

• item_idx (torch.Tensor) – interaction[‘item_id_list’]

Returns

embedding of user feature and item feature

Return type

dict

embed_token_fields(token_fields, type)[source]

Get the embedding of toekn fields

Parameters
• token_fields (torch.Tensor) – input, [batch_size, max_item_length, num_token_field]

• type (str) – user or item

Returns

token fields embedding, [batch_size, max_item_length, num_token_field, embed_dim]

Return type

torch.Tensor

embed_token_seq_fields(token_seq_fields, type)[source]

Get the embedding of token_seq fields.

Parameters
• token_seq_fields (torch.Tensor) – input, [batch_size, max_item_length, seq_len]`

• type (str) – user or item

• mode (str) – mean/max/sum

Returns

result [batch_size, max_item_length, num_token_seq_field, embed_dim]

Return type

torch.Tensor

forward(user_idx, item_idx)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

get_embedding()[source]

get embedding of all features.

get_fields_name_dim()[source]

get user feature field and item feature field.

class recbole.model.layers.ContextSeqEmbLayer(dataset, embedding_size, pooling_mode, device)[source]

For Deep Interest Network, return all features (including user features and item features) embedding matrices.

class recbole.model.layers.Dice(emb_size)[source]

Bases: torch.nn.modules.module.Module

Dice activation function

$f(s)=p(s) \cdot s+(1-p(s)) \cdot \alpha s$
$p(s)=\frac{1} {1 + e^{-\frac{s-E[s]} {\sqrt {Var[s] + \epsilon}}}}$
forward(score)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.FMEmbedding(field_dims, offsets, embed_dim)[source]

Bases: torch.nn.modules.module.Module

Embedding for token fields.

Parameters
• field_dims – list, the number of tokens in each token fields

• offsets – list, the dimension offset of each token field

• embed_dim – int, the dimension of output embedding vectors

Input:

input_x: tensor, A 3D tensor with shape:(batch_size,field_size).

Returns

tensor, A 3D tensor with shape: (batch_size,field_size,embed_dim).

Return type

output

forward(input_x)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.FMFirstOrderLinear(config, dataset, output_dim=1)[source]

Bases: torch.nn.modules.module.Module

Calculate the first order score of the input features. This class is a member of ContextRecommender, you can call it easily when inherit ContextRecommender.

embed_float_fields(float_fields, embed=True)[source]

Calculate the first order score of float feature columns

Parameters

float_fields (torch.FloatTensor) – The input tensor. shape of [batch_size, num_float_field]

Returns

The first order score of float feature columns

Return type

torch.FloatTensor

embed_token_fields(token_fields)[source]

Calculate the first order score of token feature columns

Parameters

token_fields (torch.LongTensor) – The input tensor. shape of [batch_size, num_token_field]

Returns

The first order score of token feature columns

Return type

torch.FloatTensor

embed_token_seq_fields(token_seq_fields)[source]

Calculate the first order score of token sequence feature columns

Parameters

token_seq_fields (torch.LongTensor) – The input tensor. shape of [batch_size, seq_len]

Returns

The first order score of token sequence feature columns

Return type

torch.FloatTensor

forward(interaction)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.FeatureSeqEmbLayer(dataset, embedding_size, selected_features, pooling_mode, device)[source]

For feature-rich sequential recommenders, return item features embedding matrices according to selected features.

class recbole.model.layers.FeedForward(hidden_size, inner_size, hidden_dropout_prob, hidden_act, layer_norm_eps)[source]

Bases: torch.nn.modules.module.Module

Point-wise feed-forward layer is implemented by two dense layers.

Parameters

input_tensor (torch.Tensor) – the input of the point-wise feed-forward layer

Returns

the output of the point-wise feed-forward layer

Return type

hidden_states (torch.Tensor)

forward(input_tensor)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

gelu(x)[source]

Implementation of the gelu activation function.

For information: OpenAI GPT’s gelu is slightly different (and gives slightly different results):

0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))

Also see https://arxiv.org/abs/1606.08415

get_hidden_act(act)[source]
swish(x)[source]
class recbole.model.layers.MLPLayers(layers, dropout=0, activation='relu', bn=False, init_method=None)[source]

Bases: torch.nn.modules.module.Module

Parameters
• layers (-) – a list contains the size of each layer in mlp layers

• dropout (-) – probability of an element to be zeroed. Default: 0

• activation (-) – activation function after each layer in mlp layers. Default: ‘relu’ candidates: ‘sigmoid’, ‘tanh’, ‘relu’, ‘leekyrelu’, ‘none’

Shape:

• Input: ($$N$$, *, $$H_{in}$$) where * means any number of additional dimensions $$H_{in}$$ must equal to the first value in layers

• Output: ($$N$$, *, $$H_{out}$$) where $$H_{out}$$ equals to the last value in layers

Examples:

>>> m = MLPLayers([64, 32, 16], 0.2, 'relu')
>>> input = torch.randn(128, 64)
>>> output = m(input)
>>> print(output.size())
>>> torch.Size([128, 16])
forward(input_feature)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

init_weights(module)[source]

Bases: torch.nn.modules.module.Module

Multi-head Self-attention layers, a attention score dropout layer is introduced.

Parameters
• input_tensor (torch.Tensor) – the input of the multi-head self-attention layer

Returns

the output of the multi-head self-attention layer

Return type

hidden_states (torch.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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

transpose_for_scores(x)[source]
class recbole.model.layers.SequenceAttLayer(mask_mat, att_hidden_size=80, 40, activation='sigmoid', softmax_stag=False, return_seq_weight=True)[source]

Bases: torch.nn.modules.module.Module

Attention Layer. Get the representation of each user in the batch.

Parameters
• queries (torch.Tensor) – candidate ads, [B, H], H means embedding_size * feat_num

• keys (torch.Tensor) – user_hist, [B, T, H]

• keys_length (torch.Tensor) – mask, [B]

Returns

result

Return type

torch.Tensor

forward(queries, keys, keys_length)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.TransformerEncoder(n_layers=2, n_heads=2, hidden_size=64, inner_size=256, hidden_dropout_prob=0.5, attn_dropout_prob=0.5, hidden_act='gelu', layer_norm_eps=1e-12)[source]

Bases: torch.nn.modules.module.Module

One TransformerEncoder consists of several TransformerLayers.

• n_layers(num): num of transformer layers in transformer encoder. Default: 2

• hidden_size(num): the input and output hidden size. Default: 64

• inner_size(num): the dimensionality in feed-forward layer. Default: 256

• hidden_dropout_prob(float): probability of an element to be zeroed. Default: 0.5

• attn_dropout_prob(float): probability of an attention score to be zeroed. Default: 0.5

• hidden_act(str): activation function in feed-forward layer. Default: ‘gelu’

candidates: ‘gelu’, ‘relu’, ‘swish’, ‘tanh’, ‘sigmoid’

• layer_norm_eps(float): a value added to the denominator for numerical stability. Default: 1e-12

Parameters
• hidden_states (torch.Tensor) – the input of the TrandformerEncoder

• output_all_encoded_layers (Bool) – whether output all transformer layers’ output

Returns

if output_all_encoded_layers is True, return a list consists of all transformer layers’ output, otherwise return a list only consists of the output of last transformer layer.

Return type

all_encoder_layers (list)

class recbole.model.layers.TransformerLayer(n_heads, hidden_size, intermediate_size, hidden_dropout_prob, attn_dropout_prob, hidden_act, layer_norm_eps)[source]

Bases: torch.nn.modules.module.Module

One transformer layer consists of a multi-head self-attention layer and a point-wise feed-forward layer.

Parameters
• hidden_states (torch.Tensor) – the input of the multi-head self-attention sublayer

Returns

the output of the point-wise feed-forward sublayer, is the output of the transformer layer

Return type

feedforward_output (torch.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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class recbole.model.layers.VanillaAttention(hidden_dim, attn_dim)[source]

Bases: torch.nn.modules.module.Module

Vanilla attention layer is implemented by linear layer.

Parameters

input_tensor (torch.Tensor) – the input of the attention layer

Returns

the outputs of the attention layer weights (torch.Tensor): the attention weights

Return type

hidden_states (torch.Tensor)

forward(input_tensor)[source]

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 Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

recbole.model.layers.activation_layer(activation_name='relu', emb_dim=None)[source]

Construct activation layers

Parameters
• activation_name – str, name of activation function

• emb_dim – int, used for Dice activation

Returns

activation layer

Return type

activation