import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class MPCNN(nn.Module):
def __init__(self, embedding, n_word_dim, n_holistic_filters, n_per_dim_filters, filter_widths, hidden_layer_units, num_classes, dropout):
super(MPCNN, self).__init__()
self.embedding = embedding
self.n_word_dim = n_word_dim
self.n_holistic_filters = n_holistic_filters
self.n_per_dim_filters = n_per_dim_filters
self.filter_widths = filter_widths
holistic_conv_layers = []
per_dim_conv_layers = []
self.in_channels = n_word_dim
for ws in filter_widths:
if np.isinf(ws):
continue
holistic_conv_layers.append(nn.Sequential(
nn.Conv1d(self.in_channels, n_holistic_filters, ws),
nn.Tanh()
))
per_dim_conv_layers.append(nn.Sequential(
nn.Conv1d(self.in_channels, self.in_channels * n_per_dim_filters, ws, groups=self.in_channels),
nn.Tanh()
))
self.holistic_conv_layers = nn.ModuleList(holistic_conv_layers)
self.per_dim_conv_layers = nn.ModuleList(per_dim_conv_layers)
# compute number of inputs to first hidden layer
COMP_1_COMPONENTS_HOLISTIC, COMP_1_COMPONENTS_PER_DIM, COMP_2_COMPONENTS = 2 + n_holistic_filters, 2 + self.in_channels, 2
n_feat_h = 3 * len(self.filter_widths) * COMP_2_COMPONENTS
n_feat_v = (
# comparison units from holistic conv for min, max, mean pooling for non-infinite widths
3 * ((len(self.filter_widths) - 1) ** 2) * COMP_1_COMPONENTS_HOLISTIC +
# comparison units from holistic conv for min, max, mean pooling for infinite widths
3 * 3 +
# comparison units from per-dim conv
2 * (len(self.filter_widths) - 1) * n_per_dim_filters * COMP_1_COMPONENTS_PER_DIM
)
n_feat = n_feat_h + n_feat_v
self.final_layers = nn.Sequential(
nn.Linear(n_feat, hidden_layer_units),
nn.Tanh(),
nn.Dropout(dropout),
nn.Linear(hidden_layer_units, num_classes),
nn.LogSoftmax(1)
)
def _get_blocks_for_sentence(self, sent):
block_a = {}
block_b = {}
for ws in self.filter_widths:
if np.isinf(ws):
sent_flattened, sent_flattened_size = sent.contiguous().view(sent.size(0), 1, -1), sent.size(1) * sent.size(2)
block_a[ws] = {
'max': F.max_pool1d(sent_flattened, sent_flattened_size).view(sent.size(0), -1),
'min': F.max_pool1d(-1 * sent_flattened, sent_flattened_size).view(sent.size(0), -1),
'mean': F.avg_pool1d(sent_flattened, sent_flattened_size).view(sent.size(0), -1)
}
continue
holistic_conv_out = self.holistic_conv_layers[ws - 1](sent)
block_a[ws] = {
'max': F.max_pool1d(holistic_conv_out, holistic_conv_out.size(2)).contiguous().view(-1, self.n_holistic_filters),
'min': F.max_pool1d(-1 * holistic_conv_out, holistic_conv_out.size(2)).contiguous().view(-1, self.n_holistic_filters),
'mean': F.avg_pool1d(holistic_conv_out, holistic_conv_out.size(2)).contiguous().view(-1, self.n_holistic_filters)
}
per_dim_conv_out = self.per_dim_conv_layers[ws - 1](sent)
block_b[ws] = {
'max': F.max_pool1d(per_dim_conv_out, per_dim_conv_out.size(2)).contiguous().view(-1, self.in_channels, self.n_per_dim_filters),
'min': F.max_pool1d(-1 * per_dim_conv_out, per_dim_conv_out.size(2)).contiguous().view(-1, self.in_channels, self.n_per_dim_filters)
}
return block_a, block_b
def _algo_1_horiz_comp(self, sent1_block_a, sent2_block_a):
comparison_feats = []
for pool in ('max', 'min', 'mean'):
for ws in self.filter_widths:
x1 = sent1_block_a[ws][pool]
x2 = sent2_block_a[ws][pool]
batch_size = x1.size()[0]
comparison_feats.append(F.cosine_similarity(x1, x2).contiguous().view(batch_size, 1))
comparison_feats.append(F.pairwise_distance(x1, x2))
return torch.cat(comparison_feats, dim=1)
def _algo_2_vert_comp(self, sent1_block_a, sent2_block_a, sent1_block_b, sent2_block_b):
comparison_feats = []
ws_no_inf = [w for w in self.filter_widths if not np.isinf(w)]
for pool in ('max', 'min', 'mean'):
for ws1 in self.filter_widths:
x1 = sent1_block_a[ws1][pool]
batch_size = x1.size()[0]
for ws2 in self.filter_widths:
x2 = sent2_block_a[ws2][pool]
if (not np.isinf(ws1) and not np.isinf(ws2)) or (np.isinf(ws1) and np.isinf(ws2)):
comparison_feats.append(F.cosine_similarity(x1, x2).contiguous().view(batch_size, 1))
comparison_feats.append(F.pairwise_distance(x1, x2))
comparison_feats.append(torch.abs(x1 - x2))
for pool in ('max', 'min'):
for ws in ws_no_inf:
oG_1B = sent1_block_b[ws][pool]
oG_2B = sent2_block_b[ws][pool]
for i in range(0, self.n_per_dim_filters):
x1 = oG_1B[:, :, i]
x2 = oG_2B[:, :, i]
batch_size = x1.size()[0]
comparison_feats.append(F.cosine_similarity(x1, x2).contiguous().view(batch_size, 1))
comparison_feats.append(F.pairwise_distance(x1, x2))
comparison_feats.append(torch.abs(x1 - x2))
return torch.cat(comparison_feats, dim=1)
def forward(self, batch):
sent1 = self.embedding(batch.sentence_a).transpose(1, 2)
sent2 = self.embedding(batch.sentence_b).transpose(1, 2)
# Sentence modeling module
sent1_block_a, sent1_block_b = self._get_blocks_for_sentence(sent1)
sent2_block_a, sent2_block_b = self._get_blocks_for_sentence(sent2)
# Similarity measurement layer
feat_h = self._algo_1_horiz_comp(sent1_block_a, sent2_block_a)
feat_v = self._algo_2_vert_comp(sent1_block_a, sent2_block_a, sent1_block_b, sent2_block_b)
combined_feats = [feat_h, feat_v]
feat_all = torch.cat(combined_feats, dim=1)
preds = self.final_layers(feat_all)
return preds
