File size: 10,356 Bytes
f953fd7
 
 
 
 
 
 
96a7d84
 
f953fd7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
96a7d84
 
 
 
 
 
 
 
 
 
 
 
 
 
f953fd7
 
 
 
96a7d84
 
 
 
 
 
 
 
 
 
 
 
 
 
f953fd7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
from typing import Any
import torch
from torch import nn
from torch.utils.data import Dataset, DataLoader
import numpy as np
from os import listdir
from os.path import isfile, join
import concurrent
import itertools

if __package__ == None or __package__ == "":
    from utils import tag_training_data, get_upenn_tags_dict, parse_tags    
else:
    from .utils import tag_training_data, get_upenn_tags_dict, parse_tags

# Model Type 1: LSTM with 1-logit lookahead.
class SegmentorDataset(Dataset):
    def __init__(self, datapoints):
        self.datapoints = [(torch.from_numpy(k).float(), torch.tensor([t]).float()) for k, t in datapoints]
    
    def __len__(self):
        return len(self.datapoints)

    def __getitem__(self, idx):
        return self.datapoints[idx][0], self.datapoints[idx][1]
    
class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, device=None):
        super(RNN, self).__init__()

        if device == None:
            if torch.cuda.is_available():
                self.device = "cuda"
            else:
                self.device = "cpu"
        else:
            self.device = device
        
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.rnn = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True)

        self.fc = nn.Linear(hidden_size, 1)

    def forward(self, x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size, device=self.device)
        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size, device=self.device)
        out, _ = self.rnn(x, (h0, c0))

        out = out[:, -1, :]

        out = self.fc(out)

        return out

# Model 2: Bidirectional LSTM with entire sequence context (hopefully)
class SegmentorDatasetDirectTag(Dataset):
    def __init__(self, document_root: str):
        self.tags_dict = get_upenn_tags_dict()
        self.datapoints = []
        self.eye = np.eye(len(self.tags_dict))

        files = listdir(document_root)
        for f in files:
            if f.endswith(".txt"):
                fname = join(document_root, f)
                print(f"Loaded datafile: {fname}")
                reconstructed_tags = tag_training_data(fname)
                input, tag = parse_tags(reconstructed_tags)
                self.datapoints.append((
                    np.array(input),
                    np.array(tag)
                ))
        
    def __len__(self):
        return len(self.datapoints)

    def __getitem__(self, idx):
        item = self.datapoints[idx]
        return torch.from_numpy(self.eye[item[0]]).float(), torch.from_numpy(item[1]).float()

# The same dataset without one-hot embedding of the input.
class SegmentorDatasetNonEmbed(Dataset):
    @staticmethod
    def read_file(f: str, document_root: str):
        if f.endswith(".txt"):
            fname = join(document_root, f)
            print(f"Loaded datafile: {fname}")
            reconstructed_tags = tag_training_data(fname)
            input, tag = parse_tags(reconstructed_tags)
            return [(
                np.array(input),
                np.array(tag)
            )]
        else:
            return []
        
    def __init__(self, document_root: str):
        self.datapoints = []

        files = listdir(document_root)
        with concurrent.futures.ProcessPoolExecutor() as pool:
            out = pool.map(SegmentorDatasetNonEmbed.read_file, files, itertools.repeat(document_root))
        
        self.datapoints = list(itertools.chain.from_iterable(out))
        # for f in files:
        #     if f.endswith(".txt"):
        #         fname = join(document_root, f)
        #         print(f"Loaded datafile: {fname}")
        #         reconstructed_tags = tag_training_data(fname)
        #         input, tag = parse_tags(reconstructed_tags)
        #         self.datapoints.append((
        #             np.array(input),
        #             np.array(tag)
        #         ))
        
    def __len__(self):
        return len(self.datapoints)

    def __getitem__(self, idx):
        item = self.datapoints[idx]
        return torch.from_numpy(item[0]).int(), torch.from_numpy(item[1]).float()

class BidirLSTMSegmenter(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, device = None):
        super(BidirLSTMSegmenter, self).__init__()

        if device == None:
            if torch.cuda.is_available():
                self.device = "cuda"
            else:
                self.device = "cpu"
        else:
            self.device = device
        
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.rnn = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, bidirectional=True, device = self.device)

        self.fc = nn.Linear(2*hidden_size, 1, device = self.device)
        self.final = nn.Sigmoid()
    
    def forward(self, x):
        h0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size, device=self.device)
        c0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size, device=self.device)
        out, _ = self.rnn(x, (h0, c0))

        # out_fced = [torch.zeros((out.shape[0], out.shape[1]), device=device)]
        # # Shape of out: [batch, seq_length, 256 (num_directions * hidden_size)]
        # for i in range(out.shape[1]):
        #     out_fced[:, i] = self.fc(out[:, i, :])[0]

        out_fced = self.fc(out)[:, :, 0]
        
        # Shape of out: 

        return self.final(out_fced)

class BidirLSTMSegmenterWithEmbedding(nn.Module):
    def __init__(self, input_size, embedding_size, hidden_size, num_layers, device = None):
        super(BidirLSTMSegmenterWithEmbedding, self).__init__()

        if device == None:
            if torch.cuda.is_available():
                self.device = "cuda"
            else:
                self.device = "cpu"
        else:
            self.device = device
        
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.embedding_size = embedding_size

        self.embedding = nn.Embedding(input_size, embedding_dim=embedding_size, device = self.device)
        self.rnn = nn.LSTM(embedding_size, hidden_size, num_layers, batch_first=True, bidirectional=True, device = self.device)

        self.fc = nn.Linear(2*hidden_size, 1, device = self.device)
        self.final = nn.Sigmoid()
    
    def forward(self, x):
        h0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size, device=self.device)
        c0 = torch.zeros(self.num_layers * 2, x.size(0), self.hidden_size, device=self.device)
        embedded = self.embedding(x)
        out, _ = self.rnn(embedded, (h0, c0))

        # out_fced = [torch.zeros((out.shape[0], out.shape[1]), device=device)]
        # # Shape of out: [batch, seq_length, 256 (num_directions * hidden_size)]
        # for i in range(out.shape[1]):
        #     out_fced[:, i] = self.fc(out[:, i, :])[0]

        out_fced = self.fc(out)[:, :, 0]
        
        # Shape of out: 

        return self.final(out_fced)

def collate_fn_padd(batch):
    '''
    Padds batch of variable length

    note: it converts things ToTensor manually here since the ToTensor transform
    assume it takes in images rather than arbitrary tensors.
    '''
    ## get sequence lengths
    inputs = [i[0] for i in batch]
    tags = [i[1] for i in batch]

    padded_input = torch.nn.utils.rnn.pad_sequence(inputs, batch_first=True)
    combined_outputs = torch.nn.utils.rnn.pad_sequence(tags, batch_first=True)

    ## compute mask
    return (padded_input, combined_outputs)

def get_dataloader(dataset: SegmentorDataset, batch_size):
    return DataLoader(dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn_padd)

def train_model(model: RNN,
    dataset,
    lr = 1e-3,
    num_epochs = 3,
    batch_size = 100,
):
    train_loader = get_dataloader(dataset, batch_size=batch_size)
    
    n_total_steps = len(train_loader)
    criterion = nn.MSELoss()
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
    device = model.device

    for epoch in range(num_epochs):
        for i, (input, tags) in enumerate(train_loader):
            input = input.to(device)
            tags = tags.to(device)

            outputs = model(input)
            loss = criterion(outputs, tags)

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        
            if i%100 == 0:
                print(f"Epoch [{epoch+1}/{num_epochs}], Step [{i+1}/{n_total_steps}], Loss [{loss.item():.4f}]")

def train_bidirlstm_model(model: BidirLSTMSegmenter,
    dataset: SegmentorDatasetDirectTag,
    lr = 1e-3,
    num_epochs = 3,
    batch_size = 1,
):
    train_loader = get_dataloader(dataset, batch_size=batch_size)
    
    n_total_steps = len(train_loader)
    criterion = nn.BCELoss()
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
    device = model.device

    for epoch in range(num_epochs):
        for i, (input, tags) in enumerate(train_loader):
            input = input.to(device)
            tags = tags.to(device)

            optimizer.zero_grad()

            outputs = model(input)

            loss = criterion(outputs, tags)
            
            loss.backward()
            optimizer.step()
        
            if i%10 == 0:
                print(f"Epoch [{epoch+1}/{num_epochs}], Step [{i+1}/{n_total_steps}], Loss [{loss.item():.4f}]")

def train_bidirlstm_embedding_model(model: BidirLSTMSegmenterWithEmbedding,
    dataset: SegmentorDatasetNonEmbed,
    lr = 1e-3,
    num_epochs = 3,
    batch_size = 1,
):
    train_loader = get_dataloader(dataset, batch_size=batch_size)
    
    n_total_steps = len(train_loader)
    criterion = nn.BCELoss()
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
    device = model.device

    for epoch in range(num_epochs):
        for i, (input, tags) in enumerate(train_loader):
            input = input.to(device)
            tags = tags.to(device)

            optimizer.zero_grad()

            outputs = model(input)

            loss = criterion(outputs, tags)
            
            loss.backward()
            optimizer.step()
        
            if i%10 == 0:
                print(f"Epoch [{epoch+1}/{num_epochs}], Step [{i+1}/{n_total_steps}], Loss [{loss.item():.4f}]")