Merge 9b221a0 into 03558a8

CHANGELOG.rst

@@ -1,6 +1,23 @@
 [Unreleased]
 ------------
 
+Added
+^^^^^^^
+* `@senwu`_: Add ``SparseLSTM`` support to save LSTM memory for large application.
+  (`#175 <https://github.com/HazyResearch/fonduer/pull/175>`_)
+
+.. note::
+    With the SparseLSTM discriminative model, we save memory for the origin LSTM
+    model. In Fonduer v0.3.5, SparseLSTM is as follows:
+
+    .. code:: python
+
+        disc_model = SparseLSTM()
+        disc_model.train(
+            (train_cands, train_feature), train_marginals, n_epochs=5, lr=0.001
+        )
+
+
 Fixed
 ^^^^^
 * `@senwu`_: Use the latest MeTaL syntax and fix flake8 issues.

docs/user/learning.rst

@@ -54,6 +54,24 @@ The learning parameters of different models are described below::
       LogisticRegression:
         # bias term
         bias: False
+      # Sparse LSTM model
+      SparseLSTM:
+        # Word embedding dimension size
+        emb_dim: 100
+        # The number of features in the LSTM hidden state
+        hidden_dim: 100
+        # Use attention or not (Options: True or False)
+        attention: True
+        # Dropout parameter
+        dropout: 0.1
+        # Use bidirectional LSTM or not (Options: True or False)
+        bidirectional: True
+        # Prefered host device (Options: CPU or GPU)
+        host_device: "CPU"
+        # Maximum sentence length of LSTM input
+        max_sentence_length: 100
+        # bias term
+        bias: False
       # Sparse Logistic Regression model
       SparseLogisticRegression:
         # bias term

src/fonduer/learning/__init__.py

@@ -1,7 +1,8 @@
 from fonduer.learning.disc_models.logistic_regression import LogisticRegression
 from fonduer.learning.disc_models.lstm import LSTM
+from fonduer.learning.disc_models.sparse_lstm import SparseLSTM
 from fonduer.learning.disc_models.sparse_logistic_regression import (
     SparseLogisticRegression,
 )
 
-__all__ = ["LogisticRegression", "LSTM", "SparseLogisticRegression"]
+__all__ = ["LogisticRegression", "LSTM", "SparseLogisticRegression", "SparseLSTM"]

src/fonduer/learning/disc_models/lstm.py

@@ -26,7 +26,7 @@ def forward(self, x, f):
         """Forward function.
 
         :param x: The sequence input (batch) of the model.
-        :type x: torch.Tensor of shape (sequence_len * batch_size)
+        :type x: list of torch.Tensor of shape (sequence_len * batch_size)
         :param f: The feature input of the model.
         :type f: torch.Tensor of shape (batch_size * feature_size)
         :return: The output of LSTM layer.
@@ -214,8 +214,10 @@ def _calc_logits(self, X, batch_size=None):
 
             # TODO: optimize this
             sequences = []
+            # For loop each relation arity
             for i in range(len(C[0])):
                 sequence = []
+                # Generate sequence for the batch
                 for j in range(batch_st, batch_ed):
                     sequence.append(C[j][i])
                 x, x_mask = pad_batch(sequence, self.settings["max_sentence_length"])

src/fonduer/learning/disc_models/sparse_logistic_regression.py

@@ -19,8 +19,10 @@ def forward(self, x, w):
         """
         Run forward pass.
 
-        :param x: The input (batch) of the model.
+        :param x: The input feature (batch) of the model.
         :type x: torch.Tensor of shape (batch_size, num_classes)
+        :param w: The input feature weight (batch) of the model.
+        :type w: torch.Tensor of shape (batch_size, num_classes)
         :return: The output of sparse Logistic Regression layer.
         :rtype: torch.Tensor of shape (batch_size, num_classes)
         """
@@ -41,8 +43,7 @@ def _check_input(self, X):
     def _preprocess_data(self, X, Y=None, idxs=None, train=False):
         """
         Preprocess the data:
-        1. Convert sparse matrix to dense matrix.
-        2. Select subset of the input if idxs exists.
+        1. Select subset of the input if idxs exists.
 
         :param X: The input data of the model.
         :type X: pair with candidates and corresponding features

src/fonduer/learning/disc_models/sparse_lstm.py

@@ -0,0 +1,294 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+from fonduer.learning.disc_learning import NoiseAwareModel
+from fonduer.learning.disc_models.layers.rnn import RNN
+from fonduer.learning.disc_models.layers.sparse_linear import SparseLinear
+from fonduer.learning.disc_models.utils import (
+    SymbolTable,
+    mark_sentence,
+    mention_to_tokens,
+    pad_batch,
+)
+from fonduer.utils.config import get_config
+
+
+class SparseLSTM(NoiseAwareModel):
+    """
+    LSTM model.
+
+    :param name: User-defined name of the model
+    :type name: str
+    """
+
+    def forward(self, x, x_idx, f, w):
+        """Forward function.
+
+        :param x: The sequence input (batch) of the model.
+        :type x: list of torch.Tensor of shape (sequence_len * batch_size)
+        :param x_idx: The feature indices for lstm output (batch) of the model.
+        :type x_idx: torch.Tensor of shape (lstm_out_size * batch_size)
+        :param f: The feature input of the model.
+        :type f: torch.Tensor of shape (batch_size * feature_size)
+        :param w: The input feature weight (batch) of the model.
+        :type w: torch.Tensor of shape (batch_size, num_classes)
+        :return: The output of LSTM layer.
+        :rtype: torch.Tensor of shape (batch_size, num_classes)
+        """
+
+        batch_size = len(f)
+
+        outputs = (
+            torch.Tensor([]).cuda()
+            if self.settings["host_device"] in self._gpu
+            else torch.Tensor([])
+        )
+
+        # Calculate textual features from LSTMs
+        for i in range(len(x)):
+            state_word = self.lstms[0].init_hidden(batch_size)
+            output = self.lstms[0].forward(x[i][0], x[i][1], state_word)
+            outputs = torch.cat((outputs, output), 1)
+
+        # Concatenate textual features with multi-modal features
+        feaures = torch.cat((x_idx, f), 1)
+        weights = torch.cat((outputs, w), 1)
+
+        return self.sparse_linear(feaures, weights)
+
+    def _check_input(self, X):
+        """Check input format.
+
+        :param X: The input data of the model.
+        :type X: (candidates, features) pair
+        :return: True if valid, otherwise False.
+        :rtype: bool
+        """
+        return isinstance(X, tuple)
+
+    def _preprocess_data(self, X, Y=None, idxs=None, train=False):
+        """
+        Preprocess the data:
+        1. Make sentence with mention into sequence data for LSTM.
+        2. Select subset of the input if idxs exists.
+
+        :param X: The input data of the model.
+        :type X: pair with candidates and corresponding features
+        :param Y: The labels of input data (optional).
+        :type Y: list of floats if num_classes = 2
+            otherwise num_classes-length numpy array
+        :param idxs: The selected indexs of input data.
+        :type idxs: list or numpy.array
+        :param train: An indicator for word dictionary to extend new words.
+        :type train: bool
+        :return: Preprocessed data.
+        :rtype: list of (candidate, features) pairs
+        """
+
+        C, F = X
+
+        # Create word dictionary for LSTM
+        if not hasattr(self, "word_dict"):
+            self.word_dict = SymbolTable()
+            arity = len(C[0])
+            # Add paddings into word dictionary
+            for i in range(arity):
+                # TODO: optimize this
+                list(map(self.word_dict.get, ["~~[[" + str(i), str(i) + "]]~~"]))
+
+        # Make sequence input for LSTM from candidates
+        seq_data = []
+        for candidate in C:
+            cand_idx = []
+            for i in range(len(candidate)):
+                # Add mark for each mention in the original sentence
+                args = [
+                    (
+                        candidate[i].span.get_word_start_index(),
+                        candidate[i].span.get_word_end_index(),
+                        i,
+                    )
+                ]
+                s = mark_sentence(mention_to_tokens(candidate[i]), args)
+                f = self.word_dict.get if train else self.word_dict.lookup
+                cand_idx.append(list(map(f, s)))
+            seq_data.append(cand_idx)
+
+        # Generate proprcessed the input
+        if idxs is None:
+            if Y is not None:
+                return (
+                    [
+                        (
+                            seq_data[i],
+                            F.indices[F.indptr[i] : F.indptr[i + 1]],
+                            F.data[F.indptr[i] : F.indptr[i + 1]],
+                        )
+                        for i in range(len(C))
+                    ],
+                    Y,
+                )
+            else:
+                return [
+                    (
+                        seq_data[i],
+                        F.indices[F.indptr[i] : F.indptr[i + 1]],
+                        F.data[F.indptr[i] : F.indptr[i + 1]],
+                    )
+                    for i in range(len(C))
+                ]
+        if Y is not None:
+            return (
+                [
+                    (
+                        seq_data[i],
+                        F.indices[F.indptr[i] : F.indptr[i + 1]],
+                        F.data[F.indptr[i] : F.indptr[i + 1]],
+                    )
+                    for i in idxs
+                ],
+                Y[idxs],
+            )
+        else:
+            return [
+                (
+                    seq_data[i],
+                    F.indices[F.indptr[i] : F.indptr[i + 1]],
+                    F.data[F.indptr[i] : F.indptr[i + 1]],
+                )
+                for i in idxs
+            ]
+
+    def _update_settings(self, X):
+        """
+        Update the model argument.
+
+        :param X: The input data of the model.
+        :type X: list of (candidate, features) pairs
+        """
+
+        self.logger.info("Load defalut parameters for Sparse LSTM")
+        config = get_config()["learning"]["SparseLSTM"]
+
+        for key in config.keys():
+            if key not in self.settings:
+                self.settings[key] = config[key]
+
+        self.settings["relation_arity"] = len(X[0][0])
+        self.settings["lstm_dim"] = (
+            len(X[0][0])
+            * self.settings["hidden_dim"]
+            * (2 if self.settings["bidirectional"] else 1)
+        )
+
+        # Add one feature for padding vector (all 0s)
+        self.settings["input_dim"] = (
+            X[1].shape[1]
+            + len(X[0][0])
+            * self.settings["hidden_dim"]
+            * (2 if self.settings["bidirectional"] else 1)
+            + 1
+        )
+
+    def _build_model(self):
+        """
+        Build the model.
+        """
+        # Set up LSTM modules
+        self.lstms = nn.ModuleList(
+            [
+                RNN(
+                    num_classes=0,
+                    num_tokens=self.word_dict.s,
+                    emb_size=self.settings["emb_dim"],
+                    lstm_hidden=self.settings["hidden_dim"],
+                    attention=self.settings["attention"],
+                    dropout=self.settings["dropout"],
+                    bidirectional=self.settings["bidirectional"],
+                    use_cuda=self.settings["host_device"] in self._gpu,
+                )
+            ]
+            * self.settings["relation_arity"]
+        )
+
+        if "input_dim" not in self.settings:
+            raise ValueError("Model parameter input_dim cannot be None.")
+
+        cardinality = self.cardinality if self.cardinality > 2 else 1
+
+        # Set up final linear layer
+        self.sparse_linear = SparseLinear(
+            self.settings["input_dim"], cardinality, self.settings["bias"]
+        )
+
+    def _calc_logits(self, X, batch_size=None):
+        """
+        Calculate the logits.
+
+        :param X: The input data of the model.
+        :type X: list of (candidate, features) pairs
+        :param batch_size: The batch size.
+        :type batch_size: int
+        :return: The output logits of model.
+        :rtype: torch.Tensor of shape (batch_size, num_classes) if num_classes > 2
+            otherwise shape (batch_size, 1)
+        """
+
+        # Generate LSTM input
+        C = np.array(list(zip(*X))[0])
+
+        # Check LSTM input dimension size matches the number of lstms in the model
+        assert len(C[0]) == len(self.lstms)
+
+        # Generate sparse multi-modal feature input
+        F = (
+            np.array(list(zip(*X))[1]) + self.settings["lstm_dim"] + 1
+        )  # Correct the index since 0 is the padding and placeholder for lstm feature
+        V = np.array(list(zip(*X))[2])
+
+        outputs = (
+            torch.Tensor([]).cuda()
+            if self.settings["host_device"] in self._gpu
+            else torch.Tensor([])
+        )
+
+        n = len(F)
+        if batch_size is None:
+            batch_size = n
+        for batch_st in range(0, n, batch_size):
+            batch_ed = batch_st + batch_size if batch_st + batch_size <= n else n
+
+            # TODO: optimize this
+            sequences = []
+            # For loop each relation arity
+            for i in range(len(C[0])):
+                sequence = []
+                # Generate sequence for the batch
+                for j in range(batch_st, batch_ed):
+                    sequence.append(C[j][i])
+                x, x_mask = pad_batch(sequence, self.settings["max_sentence_length"])
+                if self.settings["host_device"] in self._gpu:
+                    x = x.cuda()
+                    x_mask = x_mask.cuda()
+                sequences.append((x, x_mask))
+
+            lstm_weight_indices = torch.as_tensor(
+                np.arange(1, self.settings["lstm_dim"] + 1)
+            ).repeat(batch_ed - batch_st, 1)
+
+            features, _ = pad_batch(F[batch_st:batch_ed], 0)
+            values, _ = pad_batch(V[batch_st:batch_ed], 0, type="float")
+
+            if self.settings["host_device"] in self._gpu:
+                lstm_weight_indices = lstm_weight_indices.cuda()
+                features = features.cuda()
+                values = values.cuda()
+
+            output = self.forward(sequences, lstm_weight_indices, features, values)
+            if self.cardinality == 2:
+                outputs = torch.cat((outputs, output.view(-1)), 0)
+            else:
+                outputs = torch.cat((outputs, output), 0)
+
+        return outputs