torch-rechub 0.0.3__py3-none-any.whl → 0.0.5__py3-none-any.whl

torch_rechub/models/matching/narm.py

@@ -1,76 +1,82 @@
-"""
-Date: created on 06/09/2022
-References:
-    paper: Neural Attentive Session-based Recommendation
-    url: http://arxiv.org/abs/1711.04725
-    official Theano implementation: https://github.com/lijingsdu/sessionRec_NARM
-    another Pytorch implementation: https://github.com/Wang-Shuo/Neural-Attentive-Session-Based-Recommendation-PyTorch
-Authors: Bo Kang, klinux@live.com
-"""
-
-import torch
-import torch.nn as nn
-import torch.nn.utils.rnn as rnn_utils
-from torch import sigmoid
-from torch.nn import GRU, Embedding, Dropout, Parameter
-
-
-class NARM(nn.Module):
-    def __init__(self, item_history_feature, hidden_dim, emb_dropout_p, session_rep_dropout_p):
-        super(NARM, self).__init__()        
-        
-        # item embedding layer
-        self.item_history_feature = item_history_feature
-        self.item_emb = Embedding(item_history_feature.vocab_size, item_history_feature.embed_dim, padding_idx=0)
-
-        # embedding dropout layer
-        self.emb_dropout = Dropout(emb_dropout_p)        
-
-        # gru unit
-        self.gru = GRU(input_size=item_history_feature.embed_dim, hidden_size=hidden_dim)
-
-        # attention projection matrices
-        self.a_1, self.a_2 = Parameter(torch.randn(hidden_dim, hidden_dim)), Parameter(torch.randn(hidden_dim, hidden_dim))
-
-        # attention context vector
-        self.v = Parameter(torch.randn(hidden_dim, 1))
-
-        # session representation dropout layer
-        self.session_rep_dropout = Dropout(session_rep_dropout_p)
-
-        # bilinear projection matrix
-        self.b = Parameter(torch.randn(item_history_feature.embed_dim, hidden_dim * 2))
-
-    def forward(self, input_dict):        
-        # Eq. 1-4, index item embeddings and pass through gru
-        ## Fetch the embeddings for items in the session        
-        input = input_dict[self.item_history_feature.name]
-        value_mask = (input != 0)
-        value_counts = value_mask.sum(dim=1, keepdim=False).to("cpu").detach()
-        embs = rnn_utils.pack_padded_sequence(self.emb_dropout(self.item_emb(input)), value_counts, batch_first=True, enforce_sorted=False)
-
-        ## compute hidden states at each time step
-        h, h_t = self.gru(embs)
-        h_t = h_t.permute(1, 0, 2)
-        h, _ = rnn_utils.pad_packed_sequence(h, batch_first=True)
-
-        # Eq. 5, set last hidden state of gru as the output of the global encoder
-        c_g = h_t.squeeze(1)
-
-        # Eq. 8, compute similarity between final hidden state and previous hidden states
-        q = sigmoid(h_t @ self.a_1.T + h @ self.a_2.T) @ self.v
-        
-        # Eq. 7, compute attention
-        alpha = torch.exp(q) * value_mask.unsqueeze(-1)
-        alpha /= alpha.sum(dim=1, keepdim=True)
-
-        # Eq. 6, compute the output of the local encoder
-        c_l = (alpha * h).sum(1)
-
-        # Eq. 9, compute session representation by concatenating user sequential behavior (global) and main purpose in the current session (local)
-        c = self.session_rep_dropout(torch.hstack((c_g, c_l)))
-
-        # Eq. 10, compute bilinear similarity between current session and each candidate items
-        s = c @ self.b.T @ self.item_emb.weight.T
-        
-        return s
+"""
+Date: created on 06/09/2022
+References:
+    paper: Neural Attentive Session-based Recommendation
+    url: http://arxiv.org/abs/1711.04725
+    official Theano implementation: https://github.com/lijingsdu/sessionRec_NARM
+    another Pytorch implementation: https://github.com/Wang-Shuo/Neural-Attentive-Session-Based-Recommendation-PyTorch
+Authors: Bo Kang, klinux@live.com
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.utils.rnn as rnn_utils
+from torch import sigmoid
+from torch.nn import GRU, Dropout, Embedding, Parameter
+
+
+class NARM(nn.Module):
+
+    def __init__(self, item_history_feature, hidden_dim, emb_dropout_p, session_rep_dropout_p):
+        super(NARM, self).__init__()
+
+        # item embedding layer
+        self.item_history_feature = item_history_feature
+        self.item_emb = Embedding(item_history_feature.vocab_size, item_history_feature.embed_dim, padding_idx=0)
+
+        # embedding dropout layer
+        self.emb_dropout = Dropout(emb_dropout_p)
+
+        # gru unit
+        self.gru = GRU(input_size=item_history_feature.embed_dim, hidden_size=hidden_dim)
+
+        # attention projection matrices
+        self.a_1, self.a_2 = Parameter(torch.randn(hidden_dim, hidden_dim)), Parameter(torch.randn(hidden_dim, hidden_dim))
+
+        # attention context vector
+        self.v = Parameter(torch.randn(hidden_dim, 1))
+
+        # session representation dropout layer
+        self.session_rep_dropout = Dropout(session_rep_dropout_p)
+
+        # bilinear projection matrix
+        self.b = Parameter(torch.randn(item_history_feature.embed_dim, hidden_dim * 2))
+
+    def forward(self, input_dict):
+        # Eq. 1-4, index item embeddings and pass through gru
+        # # Fetch the embeddings for items in the session
+        input = input_dict[self.item_history_feature.name]
+        value_mask = (input != 0)
+        value_counts = value_mask.sum(dim=1, keepdim=False).to("cpu").detach()
+        embs = rnn_utils.pack_padded_sequence(self.emb_dropout(self.item_emb(input)), value_counts, batch_first=True, enforce_sorted=False)
+
+        # # compute hidden states at each time step
+        h, h_t = self.gru(embs)
+        h_t = h_t.permute(1, 0, 2)
+        h, _ = rnn_utils.pad_packed_sequence(h, batch_first=True)
+
+        # Eq. 5, set last hidden state of gru as the output of the global
+        # encoder
+        c_g = h_t.squeeze(1)
+
+        # Eq. 8, compute similarity between final hidden state and previous
+        # hidden states
+        q = sigmoid(h_t @ self.a_1.T + h @ self.a_2.T) @ self.v
+
+        # Eq. 7, compute attention
+        alpha = torch.exp(q) * value_mask.unsqueeze(-1)
+        alpha /= alpha.sum(dim=1, keepdim=True)
+
+        # Eq. 6, compute the output of the local encoder
+        c_l = (alpha * h).sum(1)
+
+        # Eq. 9, compute session representation by concatenating user
+        # sequential behavior (global) and main purpose in the current session
+        # (local)
+        c = self.session_rep_dropout(torch.hstack((c_g, c_l)))
+
+        # Eq. 10, compute bilinear similarity between current session and each
+        # candidate items
+        s = c @ self.b.T @ self.item_emb.weight.T
+
+        return s

torch_rechub/models/matching/sasrec.py

@@ -1,140 +1,143 @@
-"""
-Date: create on 2022/5/8, update on 2022/5/8
-References:
-    paper: (ICDM'2018) Self-attentive sequential recommendation
-    url: https://arxiv.org/pdf/1808.09781.pdf
-    code: https://github.com/kang205/SASRec
-Authors: Yuchen Wang, 615922749@qq.com
-"""
-import numpy as np
-import torch
-import torch.nn as nn
-
-from torch_rechub.basic.features import DenseFeature, SparseFeature, SequenceFeature
-from torch_rechub.basic.layers import EmbeddingLayer, MLP
-
-
-class SASRec(torch.nn.Module):
-    """SASRec: Self-Attentive Sequential Recommendation
-    Args:
-        features (list): the list of `Feature Class`. In sasrec, the features list needs to have three elements in order: user historical behavior sequence features, positive sample sequence, and negative sample sequence.
-        max_len: The length of the sequence feature.
-        num_blocks: The number of stacks of attention modules.
-        num_heads: The number of heads in MultiheadAttention.
-
-    """
-    def __init__(self,
-                 features,
-                 max_len=50,
-                 dropout_rate=0.5,
-                 num_blocks=2,
-                 num_heads=1,
-                 ):
-        super(SASRec, self).__init__()
-
-        self.features = features
-
-        self.item_num = self.features[0].vocab_size
-        self.embed_dim = self.features[0].embed_dim
-
-        self.item_emb = EmbeddingLayer(self.features)
-        self.position_emb = torch.nn.Embedding(max_len, self.embed_dim)
-        self.emb_dropout = torch.nn.Dropout(p=dropout_rate)
-
-        self.attention_layernorms = torch.nn.ModuleList()
-        self.attention_layers = torch.nn.ModuleList()
-        self.forward_layernorms = torch.nn.ModuleList()
-        self.forward_layers = torch.nn.ModuleList()
-        self.last_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
-
-        for _ in range(num_blocks):
-            new_attn_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
-            self.attention_layernorms.append(new_attn_layernorm)
-
-            new_attn_layer = torch.nn.MultiheadAttention(self.embed_dim,
-                                                         num_heads,
-                                                         dropout_rate)
-            self.attention_layers.append(new_attn_layer)
-
-            new_fwd_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
-            self.forward_layernorms.append(new_fwd_layernorm)
-
-            new_fwd_layer = PointWiseFeedForward(self.embed_dim, dropout_rate)
-            self.forward_layers.append(new_fwd_layer)
-
-    def seq_forward(self, x, embed_x_feature):
-        x = x['seq']
-
-        embed_x_feature *= self.features[0].embed_dim ** 0.5
-        embed_x_feature = embed_x_feature.squeeze() # (bacth_size, max_len, embed_dim)
-
-        positions = np.tile(np.array(range(x.shape[1])), [x.shape[0], 1])
-
-        embed_x_feature += self.position_emb(torch.LongTensor(positions))
-        embed_x_feature = self.emb_dropout(embed_x_feature)
-
-        timeline_mask = torch.BoolTensor(x == 0)
-        embed_x_feature *= ~timeline_mask.unsqueeze(-1)
-
-        attention_mask = ~torch.tril(torch.ones((embed_x_feature.shape[1], embed_x_feature.shape[1]), dtype=torch.bool))
-
-        for i in range(len(self.attention_layers)):
-            embed_x_feature = torch.transpose(embed_x_feature, 0, 1)
-            Q = self.attention_layernorms[i](embed_x_feature)
-            mha_outputs, _ = self.attention_layers[i](Q, embed_x_feature, embed_x_feature,
-                                                      attn_mask=attention_mask)
-
-            embed_x_feature = Q + mha_outputs
-            embed_x_feature = torch.transpose(embed_x_feature, 0, 1)
-
-            embed_x_feature = self.forward_layernorms[i](embed_x_feature)
-            embed_x_feature = self.forward_layers[i](embed_x_feature)
-            embed_x_feature *= ~timeline_mask.unsqueeze(-1)
-
-        seq_output = self.last_layernorm(embed_x_feature)
-
-        return seq_output
-
-    def forward(self, x):
-        embedding = self.item_emb(x, self.features) # (batch_size, 3, max_len, embed_dim)
-        seq_embed, pos_embed, neg_embed = embedding[:, 0], embedding[:, 1], embedding[:, 2] # (batch_size, max_len, embed_dim)
-
-        seq_output = self.seq_forward(x, seq_embed) # (batch_size, max_len, embed_dim)
-
-        pos_logits = (seq_output * pos_embed).sum(dim=-1)
-        neg_logits = (seq_output * neg_embed).sum(dim=-1) # (batch_size, max_len)
-
-        return pos_logits, neg_logits
-
-
-class PointWiseFeedForward(torch.nn.Module):
-    def __init__(self, hidden_units, dropout_rate):
-        super(PointWiseFeedForward, self).__init__()
-
-        self.conv1 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
-        self.dropout1 = torch.nn.Dropout(p=dropout_rate)
-        self.relu = torch.nn.ReLU()
-        self.conv2 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
-        self.dropout2 = torch.nn.Dropout(p=dropout_rate)
-
-    def forward(self, inputs):
-        outputs = self.dropout2(self.conv2(self.relu(self.dropout1(self.conv1(inputs.transpose(-1, -2))))))
-        outputs = outputs.transpose(-1, -2)
-        outputs += inputs
-        return outputs
-
-
-if __name__ == '__main__':
-    seq = SequenceFeature('seq', vocab_size=17, embed_dim=7, pooling='concat')
-    pos = SequenceFeature('pos', vocab_size=17, embed_dim=7, pooling='concat', shared_with='seq')
-    neg = SequenceFeature('neg', vocab_size=17, embed_dim=7, pooling='concat', shared_with='seq')
-
-    seq = [seq, pos, neg]
-
-    hist_seq = torch.tensor([[1, 2, 3, 4], [2, 3, 7, 8]])
-    pos_seq = hist_seq
-    neg_seq = hist_seq
-
-    x = {'seq': hist_seq, 'pos': pos_seq, 'neg': neg_seq}
-    model = SASRec(features=seq)
-    print('out', model(x))
+"""
+Date: create on 2022/5/8, update on 2022/5/8
+References:
+    paper: (ICDM'2018) Self-attentive sequential recommendation
+    url: https://arxiv.org/pdf/1808.09781.pdf
+    code: https://github.com/kang205/SASRec
+Authors: Yuchen Wang, 615922749@qq.com
+"""
+import numpy as np
+import torch
+import torch.nn as nn
+
+from torch_rechub.basic.features import DenseFeature, SequenceFeature, SparseFeature
+from torch_rechub.basic.layers import MLP, EmbeddingLayer
+
+
+class SASRec(torch.nn.Module):
+    """SASRec: Self-Attentive Sequential Recommendation
+    Args:
+        features (list): the list of `Feature Class`. In sasrec, the features list needs to have three elements in order: user historical behavior sequence features, positive sample sequence, and negative sample sequence.
+        max_len: The length of the sequence feature.
+        num_blocks: The number of stacks of attention modules.
+        num_heads: The number of heads in MultiheadAttention.
+
+    """
+
+    def __init__(
+        self,
+        features,
+        max_len=50,
+        dropout_rate=0.5,
+        num_blocks=2,
+        num_heads=1,
+    ):
+        super(SASRec, self).__init__()
+
+        self.features = features
+
+        self.item_num = self.features[0].vocab_size
+        self.embed_dim = self.features[0].embed_dim
+
+        self.item_emb = EmbeddingLayer(self.features)
+        self.position_emb = torch.nn.Embedding(max_len, self.embed_dim)
+        self.emb_dropout = torch.nn.Dropout(p=dropout_rate)
+
+        self.attention_layernorms = torch.nn.ModuleList()
+        self.attention_layers = torch.nn.ModuleList()
+        self.forward_layernorms = torch.nn.ModuleList()
+        self.forward_layers = torch.nn.ModuleList()
+        self.last_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
+
+        for _ in range(num_blocks):
+            new_attn_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
+            self.attention_layernorms.append(new_attn_layernorm)
+
+            new_attn_layer = torch.nn.MultiheadAttention(self.embed_dim, num_heads, dropout_rate)
+            self.attention_layers.append(new_attn_layer)
+
+            new_fwd_layernorm = torch.nn.LayerNorm(self.embed_dim, eps=1e-8)
+            self.forward_layernorms.append(new_fwd_layernorm)
+
+            new_fwd_layer = PointWiseFeedForward(self.embed_dim, dropout_rate)
+            self.forward_layers.append(new_fwd_layer)
+
+    def seq_forward(self, x, embed_x_feature):
+        x = x['seq']
+
+        embed_x_feature *= self.features[0].embed_dim**0.5
+        embed_x_feature = embed_x_feature.squeeze()  # (bacth_size, max_len, embed_dim)
+
+        positions = np.tile(np.array(range(x.shape[1])), [x.shape[0], 1])
+
+        embed_x_feature += self.position_emb(torch.LongTensor(positions))
+        embed_x_feature = self.emb_dropout(embed_x_feature)
+
+        timeline_mask = torch.BoolTensor(x == 0)
+        embed_x_feature *= ~timeline_mask.unsqueeze(-1)
+
+        attention_mask = ~torch.tril(torch.ones((embed_x_feature.shape[1], embed_x_feature.shape[1]), dtype=torch.bool))
+
+        for i in range(len(self.attention_layers)):
+            embed_x_feature = torch.transpose(embed_x_feature, 0, 1)
+            Q = self.attention_layernorms[i](embed_x_feature)
+            mha_outputs, _ = self.attention_layers[i](Q, embed_x_feature, embed_x_feature, attn_mask=attention_mask)
+
+            embed_x_feature = Q + mha_outputs
+            embed_x_feature = torch.transpose(embed_x_feature, 0, 1)
+
+            embed_x_feature = self.forward_layernorms[i](embed_x_feature)
+            embed_x_feature = self.forward_layers[i](embed_x_feature)
+            embed_x_feature *= ~timeline_mask.unsqueeze(-1)
+
+        seq_output = self.last_layernorm(embed_x_feature)
+
+        return seq_output
+
+    def forward(self, x):
+        # (batch_size, 3, max_len, embed_dim)
+        embedding = self.item_emb(x, self.features)
+        # (batch_size, max_len, embed_dim)
+        seq_embed, pos_embed, neg_embed = embedding[:, 0], embedding[:, 1], embedding[:, 2]
+
+        # (batch_size, max_len, embed_dim)
+        seq_output = self.seq_forward(x, seq_embed)
+
+        pos_logits = (seq_output * pos_embed).sum(dim=-1)
+        neg_logits = (seq_output * neg_embed).sum(dim=-1)  # (batch_size, max_len)
+
+        return pos_logits, neg_logits
+
+
+class PointWiseFeedForward(torch.nn.Module):
+
+    def __init__(self, hidden_units, dropout_rate):
+        super(PointWiseFeedForward, self).__init__()
+
+        self.conv1 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
+        self.dropout1 = torch.nn.Dropout(p=dropout_rate)
+        self.relu = torch.nn.ReLU()
+        self.conv2 = torch.nn.Conv1d(hidden_units, hidden_units, kernel_size=1)
+        self.dropout2 = torch.nn.Dropout(p=dropout_rate)
+
+    def forward(self, inputs):
+        outputs = self.dropout2(self.conv2(self.relu(self.dropout1(self.conv1(inputs.transpose(-1, -2))))))
+        outputs = outputs.transpose(-1, -2)
+        outputs += inputs
+        return outputs
+
+
+if __name__ == '__main__':
+    seq = SequenceFeature('seq', vocab_size=17, embed_dim=7, pooling='concat')
+    pos = SequenceFeature('pos', vocab_size=17, embed_dim=7, pooling='concat', shared_with='seq')
+    neg = SequenceFeature('neg', vocab_size=17, embed_dim=7, pooling='concat', shared_with='seq')
+
+    seq = [seq, pos, neg]
+
+    hist_seq = torch.tensor([[1, 2, 3, 4], [2, 3, 7, 8]])
+    pos_seq = hist_seq
+    neg_seq = hist_seq
+
+    x = {'seq': hist_seq, 'pos': pos_seq, 'neg': neg_seq}
+    model = SASRec(features=seq)
+    print('out', model(x))