Added MLP; FIX: Respect bias argument

deepxml/libs/parameters.py

@@ -329,6 +329,12 @@ def _construct(self):
             '--validate',
             action='store_true',
             help='Validate or just train')
+        self.parser.add_argument(
+            '--bias',
+            action='store',
+            default=True,
+            type=bool,
+            help='Use bias term or not!')
         self.parser.add_argument(
             '--shuffle',
             action='store',

deepxml/models/mlp.py

@@ -0,0 +1,77 @@
+import torch
+import torch.nn as nn
+
+
+__author__ = 'KD'
+
+
+class MLP(nn.Module):
+    """
+    A multi-layer perceptron with flexibility for non-liearity
+    * no non-linearity after last layer
+    * support for 2D or 3D inputs
+
+    Parameters:
+    -----------
+    input_size: int
+        input size of embeddings
+    hidden_size: int or list of ints or str (comma separated)
+        e.g., 512: a single hidden layer with 512 neurons
+              "512": a single hidden layer with 512 neurons
+              "512,300": 512 -> nnl -> 300
+              [512, 300]: 512 -> nnl -> 300
+        dimensionality of layers in MLP
+    nnl: str, optional, default='relu'
+        which non-linearity to use
+    device: str, default="cuda:0"
+        keep on this device
+    """
+    def __init__(self, input_size, hidden_size, nnl='relu', device="cuda:0"):
+        super(MLP, self).__init__()
+        hidden_size = self.parse_hidden_size(hidden_size)
+        assert len(hidden_size) >= 1, "Should contain atleast 1 hidden layer"
+        hidden_size = [input_size] + hidden_size
+        self.device = torch.device(device)
+        layers = []
+        for i, (i_s, o_s) in enumerate(zip(hidden_size[:-1], hidden_size[1:])):
+            layers.append(nn.Linear(i_s, o_s, bias=True))
+            if i < len(hidden_size) - 2:
+                layers.append(self._get_nnl(nnl))
+        self.transform = torch.nn.Sequential(*layers)
+
+    def parse_hidden_size(self, hidden_size):
+        if isinstance(hidden_size, int):
+            return [hidden_size]
+        elif isinstance(hidden_size, str):
+            _hidden_size = []
+            for item in hidden_size.split(","):
+                _hidden_size.append(int(item))
+            return _hidden_size
+        elif isinstance(hidden_size, list):
+            return hidden_size
+        else:
+            raise NotImplementedError("hidden_size must be a int, str or list")
+
+    def _get_nnl(self, nnl):
+        if nnl == 'sigmoid':
+            return torch.nn.Sigmoid()
+        elif nnl == 'relu':
+            return torch.nn.ReLU()
+        elif nnl == 'gelu':
+            return torch.nn.GELU()
+        elif nnl == 'tanh':
+            return torch.nn.Tanh()
+        else:
+            raise NotImplementedError(f"{nnl} not implemented!")
+
+    def forward(self, x):
+        return self.transform(x)
+
+    def to(self):
+        """Transfer to device
+        """
+        super().to(self.device)
+
+    @property
+    def sparse(self):
+        return False

deepxml/models/network.py

@@ -1,6 +1,5 @@
 import torch
 import torch.nn as nn
-import numpy as np
 import math
 import os
 import models.transform_layer as transform_layer
@@ -154,6 +153,7 @@ def __init__(self, params):
         trans_config_coarse = transform_config_dict['transform_coarse']
         self.representation_dims = int(
             transform_config_dict['representation_dims'])
+        self._bias = params.bias
         super(DeepXMLf, self).__init__(trans_config_coarse)
         if params.freeze_intermediate:
             print("Freezing intermediate model parameters!")
@@ -226,7 +226,7 @@ def forward(self, batch_data, bypass_coarse=False):
 
     def _construct_classifier(self):
         if self.num_clf_partitions > 1:  # Run the distributed version
-            _bias = [True for _ in range(self.num_clf_partitions)]
+            _bias = [self._bias for _ in range(self.num_clf_partitions)]
             _clf_devices = ["cuda:{}".format(
                 idx) for idx in range(self.num_clf_partitions)]
             return linear_layer.ParallelLinear(
@@ -239,7 +239,7 @@ def _construct_classifier(self):
             return linear_layer.Linear(
                 input_size=self.representation_dims,
                 output_size=self.num_labels,  # last one is padding index
-                bias=True
+                bias=self._bias
             )
 
     def get_token_embeddings(self):
@@ -297,6 +297,7 @@ def __init__(self, params):
         trans_config_coarse = transform_config_dict['transform_coarse']
         self.representation_dims = int(
             transform_config_dict['representation_dims'])
+        self._bias = params.bias
         super(DeepXMLs, self).__init__(trans_config_coarse)
         if params.freeze_intermediate:
             print("Freezing intermediate model parameters!")
@@ -383,7 +384,7 @@ def _construct_classifier(self):
             # last one is padding index for each partition
             _num_labels = self.num_labels + offset
             _padding_idx = [None for _ in range(self.num_clf_partitions)]
-            _bias = [True for _ in range(self.num_clf_partitions)]
+            _bias = [self._bias for _ in range(self.num_clf_partitions)]
             _clf_devices = ["cuda:{}".format(
                 idx) for idx in range(self.num_clf_partitions)]
             return linear_layer.ParallelSparseLinear(
@@ -399,7 +400,7 @@ def _construct_classifier(self):
                 input_size=self.representation_dims,
                 output_size=self.num_labels + offset,
                 padding_idx=self.label_padding_index,
-                bias=True)
+                bias=self._bias)
 
     def to(self):
         """Send layers to respective devices

deepxml/models/transform_layer.py

@@ -1,10 +1,9 @@
-import sys
 import re
 import torch.nn as nn
 import models.residual_layer as residual_layer
 import models.astec as astec
 import json
-from collections import OrderedDict
+import models.mlp as mlp
 
 
 class _Identity(nn.Module):
@@ -38,7 +37,8 @@ def initialize(self, *args, **kwargs):
     'residual': residual_layer.Residual,
     'identity': Identity,
     '_identity': _Identity,
-    'astec': astec.Astec
+    'astec': astec.Astec,
+    'mlp': mlp.MLP
 }