Spaces:
Running
Running
teepoat
commited on
Initial commit
Browse files- main.py +60 -0
- models/seq2seq/__init__.py +0 -0
- models/seq2seq/attention.py +36 -0
- models/seq2seq/chat_dataset.py +97 -0
- models/seq2seq/custom_types.py +23 -0
- models/seq2seq/model.py +208 -0
- models/seq2seq/searchers.py +27 -0
- models/seq2seq/vocab.py +43 -0
main.py
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Final
|
2 |
+
from telegram import Update
|
3 |
+
from telegram.ext import Application, MessageHandler, filters, ContextTypes
|
4 |
+
from typing import Optional
|
5 |
+
import random
|
6 |
+
import os
|
7 |
+
from dotenv import load_dotenv
|
8 |
+
from models.seq2seq.model import Seq2SeqChatbot
|
9 |
+
import torch
|
10 |
+
|
11 |
+
|
12 |
+
load_dotenv()
|
13 |
+
|
14 |
+
TOKEN: Final = os.environ.get("TOKEN")
|
15 |
+
BOT_USERNAME: Final = os.environ.get("BOT_USERNAME")
|
16 |
+
CHAT_ID: Final = int(os.environ.get("CHAT_ID"))
|
17 |
+
|
18 |
+
CHECKPOINT_PATH: Final = "models/seq2seq/checkpoint/150_checkpoint.tar"
|
19 |
+
|
20 |
+
torch.manual_seed(0)
|
21 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
22 |
+
|
23 |
+
chatbot = Seq2SeqChatbot(500, 2, 2, 0.1, device)
|
24 |
+
chatbot.load_checkpoint(CHECKPOINT_PATH)
|
25 |
+
chatbot.eval_mode()
|
26 |
+
|
27 |
+
def handle_response(text: str) -> Optional[str]:
|
28 |
+
response_chance = 1.0
|
29 |
+
if random.random() < response_chance:
|
30 |
+
return chatbot(text)
|
31 |
+
return None
|
32 |
+
|
33 |
+
|
34 |
+
async def handle_message(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
35 |
+
if update.message.chat_id == CHAT_ID:
|
36 |
+
text: str = update.message.text.replace(BOT_USERNAME, '').strip().lower()
|
37 |
+
response: Optional[str] = handle_response(text)
|
38 |
+
if response:
|
39 |
+
await context.bot.sendMessage(update.message.chat_id, response)
|
40 |
+
|
41 |
+
|
42 |
+
async def error(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
43 |
+
print(f"{update.message.from_user.username} in {update.message.chat.type} "
|
44 |
+
f"chat caused error \"{context.error}\"\n"
|
45 |
+
f"{update}\"")
|
46 |
+
|
47 |
+
def main() -> None:
|
48 |
+
"""Run the bot."""
|
49 |
+
application = Application.builder().token(TOKEN).build()
|
50 |
+
|
51 |
+
application.add_handler(MessageHandler(filters.TEXT, handle_message))
|
52 |
+
application.add_error_handler(error)
|
53 |
+
|
54 |
+
application.run_polling(allowed_updates=Update.ALL_TYPES)
|
55 |
+
|
56 |
+
|
57 |
+
if __name__ == '__main__':
|
58 |
+
print("Running main...")
|
59 |
+
# print(chatbot("test"))
|
60 |
+
main()
|
models/seq2seq/__init__.py
ADDED
File without changes
|
models/seq2seq/attention.py
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
import torch.nn as nn
|
4 |
+
from .custom_types import Method
|
5 |
+
|
6 |
+
class LuongAttention(nn.Module):
|
7 |
+
def __init__(self, method: Method, hidden_size: int):
|
8 |
+
super().__init__()
|
9 |
+
self.hidden_size = hidden_size
|
10 |
+
if not isinstance(method, Method):
|
11 |
+
raise ValueError(method, f"should be a member of `Method` enum")
|
12 |
+
match method:
|
13 |
+
case Method.DOT:
|
14 |
+
self.method = self.dot
|
15 |
+
case Method.GENERAL:
|
16 |
+
self.method = self.general
|
17 |
+
self.Wa = nn.Linear(hidden_size, hidden_size)
|
18 |
+
case Method.CONCAT:
|
19 |
+
self.method = self.concat
|
20 |
+
self.Wa = nn.Linear(hidden_size * 2, hidden_size)
|
21 |
+
self.Va = nn.Parameter(torch.FloatTensor(1, hidden_size))
|
22 |
+
|
23 |
+
def dot(self, hidden, encoder_outputs):
|
24 |
+
return torch.sum(hidden * encoder_outputs, dim=2)
|
25 |
+
|
26 |
+
def general(self, hidden, encoder_outputs):
|
27 |
+
return torch.sum(hidden * self.Wa(encoder_outputs), dim=2)
|
28 |
+
|
29 |
+
def concat(self, hidden, encoder_outputs):
|
30 |
+
hidden = hidden.permute(1, 0, 2)
|
31 |
+
energy = self.Wa(torch.cat((hidden.permute(1, 0, 2).expand(-1, encoder_outputs.size(1), -1), encoder_outputs), 2)).tanh()
|
32 |
+
return torch.sum(self.Va * energy, dim=2)
|
33 |
+
|
34 |
+
def forward(self, hidden, encoder_outputs):
|
35 |
+
attn_weights = self.method(hidden, encoder_outputs)
|
36 |
+
return F.softmax(attn_weights, dim=1).unsqueeze(1)
|
models/seq2seq/chat_dataset.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.utils.data as data
|
3 |
+
from typing import List, Union, Tuple
|
4 |
+
from collections import OrderedDict
|
5 |
+
from .vocab import Vocab
|
6 |
+
from .custom_types import Message, MessageId, Conversation
|
7 |
+
from torch.nn.utils.rnn import pad_sequence
|
8 |
+
from .custom_types import Token
|
9 |
+
import re
|
10 |
+
import json
|
11 |
+
|
12 |
+
class ChatDataset(data.Dataset):
|
13 |
+
def __init__(self, path: str, max_message_count: int = None, batch_size=5):
|
14 |
+
super().__init__()
|
15 |
+
self.path = path
|
16 |
+
self.batch_size = batch_size
|
17 |
+
self.messages: OrderedDict[MessageId, Message] = self.__load_messages_from_json(path, max_message_count)
|
18 |
+
self.conversations: List[Conversation] = ChatDataset.__conversations_from_messages(self.messages)
|
19 |
+
self.vocab = Vocab(list(self.messages.values())) # TODO: try changing this cast to something more applicable
|
20 |
+
|
21 |
+
self.batches_X, self.batches_y, self.lengths, self.mask = self.__batches_from_conversations()
|
22 |
+
|
23 |
+
self.length = len(self.batches_X)
|
24 |
+
|
25 |
+
def __batches_from_conversations(self) -> Tuple[List[torch.LongTensor], List[torch.LongTensor], List[torch.LongTensor], List[torch.BoolTensor]]: # Shape of tensor in batch: (batch_size, max_len_in_batch)
|
26 |
+
conversations = sorted(self.conversations, key=lambda x: len(x[0])) # Sort by input sequence length
|
27 |
+
batches_X: List[torch.LongTensor] = list()
|
28 |
+
batches_y: List[torch.LongTensor] = list()
|
29 |
+
lengths: List[torch.LongTensor] = list()
|
30 |
+
mask: List[torch.BoolTensor] = list()
|
31 |
+
for i in range(0, len(conversations), self.batch_size):
|
32 |
+
batches_X.append(pad_sequence([self.vocab.sentence_indices(conversations[i+j][0] + ["<eos>"]) for j in range(self.batch_size) if i+j < len(conversations)], batch_first=True, padding_value=0))
|
33 |
+
batches_y.append(pad_sequence([self.vocab.sentence_indices(conversations[i+j][1] + ["<eos>"]) for j in range(self.batch_size) if i+j < len(conversations)], batch_first=True, padding_value=0))
|
34 |
+
lengths.append(torch.tensor([len(conversations[i+j][0]) for j in range(self.batch_size) if i+j < len(conversations)]))
|
35 |
+
mask.append(batches_y[-1] != Token.PAD_TOKEN.value)
|
36 |
+
return batches_X, batches_y, lengths, mask
|
37 |
+
|
38 |
+
@classmethod
|
39 |
+
def __load_messages_from_json(cls, path: str, max_message_count: int = None) -> OrderedDict[MessageId, Message]:
|
40 |
+
messages: OrderedDict[MessageId, Message] = OrderedDict()
|
41 |
+
with open(path, "r", encoding="utf-8") as file:
|
42 |
+
chat_json = json.load(file)
|
43 |
+
for i, message in enumerate(chat_json["messages"]):
|
44 |
+
if max_message_count and i == max_message_count:
|
45 |
+
break
|
46 |
+
if message["type"] != "message":
|
47 |
+
continue
|
48 |
+
new_message = {
|
49 |
+
"id": message["id"],
|
50 |
+
"text": cls.__normalize(message["text"])
|
51 |
+
}
|
52 |
+
if not new_message["text"]: # Check for empty message
|
53 |
+
continue
|
54 |
+
if "reply_to_message_id" in message.keys():
|
55 |
+
new_message["reply_to_id"] = message["reply_to_message_id"]
|
56 |
+
|
57 |
+
messages[new_message["id"]] = new_message
|
58 |
+
return messages
|
59 |
+
|
60 |
+
@classmethod
|
61 |
+
def __conversations_from_messages(cls, messages: OrderedDict[MessageId, Message]) -> List[Conversation]:
|
62 |
+
# Search for message with `id` in the last `current_id` messages
|
63 |
+
def _get_message_by_id(current_id: int, id: int) -> Message:
|
64 |
+
for i in range(current_id - 1, -1, -1):
|
65 |
+
if messages[i]["id"] == id:
|
66 |
+
return messages[i]
|
67 |
+
return None
|
68 |
+
|
69 |
+
conversations: List[Conversation] = []
|
70 |
+
|
71 |
+
messages_values = list(messages.values()) # TODO: try changing this cast to something more applicable
|
72 |
+
for i in range(len(messages) - 1): # There's no answer for last message so add -1
|
73 |
+
prev_message = messages_values[i]
|
74 |
+
if "reply_to_id" in messages_values[i].keys(): # Message is answer for message with `id` of `reply_to_id`
|
75 |
+
try:
|
76 |
+
prev_message = messages[messages_values[i]["reply_to_id"]]
|
77 |
+
except KeyError:
|
78 |
+
continue
|
79 |
+
conversations.append((prev_message["text"], messages_values[i+1]["text"]))
|
80 |
+
return conversations
|
81 |
+
|
82 |
+
@classmethod
|
83 |
+
def __normalize(cls, text: Union[str, List]) -> List[str]:
|
84 |
+
if isinstance(text, List):
|
85 |
+
text = " ".join([word for word in text if isinstance(word, str)])
|
86 |
+
text = text.lower().strip()
|
87 |
+
text = re.sub(r"([.!?])", r" \1 ", text)
|
88 |
+
text = re.sub(r"ё", r"е", text)
|
89 |
+
text = re.sub(r"[^а-яА-я.!?]+", r" ", text)
|
90 |
+
text = re.sub(r"\s+", r" ", text).strip()
|
91 |
+
return text.split()
|
92 |
+
|
93 |
+
def __getitem__(self, item):
|
94 |
+
return self.batches_X[item], self.batches_y[item], self.lengths[item], self.mask[item]
|
95 |
+
|
96 |
+
def __len__(self):
|
97 |
+
return self.length
|
models/seq2seq/custom_types.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, TypedDict, NotRequired, Tuple
|
2 |
+
from enum import Enum, auto
|
3 |
+
|
4 |
+
MessageId = int
|
5 |
+
MessageText = List[str]
|
6 |
+
Conversation = Tuple[MessageText]
|
7 |
+
|
8 |
+
|
9 |
+
class Message(TypedDict):
|
10 |
+
id: MessageId
|
11 |
+
text: MessageText
|
12 |
+
reply_to_id: NotRequired[int]
|
13 |
+
|
14 |
+
class Method(Enum):
|
15 |
+
DOT = auto()
|
16 |
+
GENERAL = auto()
|
17 |
+
CONCAT = auto()
|
18 |
+
|
19 |
+
class Token(Enum):
|
20 |
+
PAD_TOKEN = 0
|
21 |
+
BOS_TOKEN = 1
|
22 |
+
EOS_TOKEN = 2
|
23 |
+
UNK_TOKEN = 3
|
models/seq2seq/model.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.optim as optim
|
6 |
+
from .chat_dataset import ChatDataset
|
7 |
+
from .attention import LuongAttention
|
8 |
+
from .custom_types import Method, Token
|
9 |
+
from .vocab import Vocab
|
10 |
+
from .searchers import GreedySearch
|
11 |
+
import os
|
12 |
+
import random
|
13 |
+
from tqdm import tqdm
|
14 |
+
|
15 |
+
|
16 |
+
class Seq2SeqEncoder(nn.Module):
|
17 |
+
def __init__(self, input_size: int, hidden_size: int, num_layers: int, embedding: nn.Embedding):
|
18 |
+
super().__init__()
|
19 |
+
self.input_size = input_size
|
20 |
+
self.hidden_size = hidden_size
|
21 |
+
self.num_layers = num_layers
|
22 |
+
|
23 |
+
self.embedding = embedding
|
24 |
+
self.rnn = nn.GRU(input_size, hidden_size, num_layers=num_layers, bidirectional=True, batch_first=True) # batch_first is True, because I don't approve self-harm
|
25 |
+
|
26 |
+
def forward(self, x, lengths):
|
27 |
+
x = self.embedding(x) # Output shape: (batch_size, max_len_in_batch, hidden_size)
|
28 |
+
packed_embedded = pack_padded_sequence(x, lengths.cpu(), batch_first=True, enforce_sorted=False)
|
29 |
+
outputs, hidden = self.rnn(packed_embedded)
|
30 |
+
outputs, _ = pad_packed_sequence(outputs, batch_first=True)
|
31 |
+
return outputs[:, :, :self.hidden_size] + outputs[:, :, self.hidden_size:], hidden
|
32 |
+
|
33 |
+
|
34 |
+
class Seq2SeqDecoder(nn.Module):
|
35 |
+
def __init__(self, input_size: int, hidden_size: int, output_size: int, num_layers: int, attn, embedding: nn.Embedding, dropout: int = 0.1):
|
36 |
+
super().__init__()
|
37 |
+
self.input_size = input_size
|
38 |
+
self.hidden_size = hidden_size
|
39 |
+
self.output_size = output_size
|
40 |
+
self.num_layers = num_layers
|
41 |
+
|
42 |
+
self.attn = attn
|
43 |
+
self.embedding = embedding
|
44 |
+
self.embedding_dropout = nn.Dropout(dropout)
|
45 |
+
self.rnn = nn.GRU(input_size, hidden_size, num_layers=num_layers, batch_first=True)
|
46 |
+
|
47 |
+
self.concat = nn.Linear(hidden_size * 2, hidden_size)
|
48 |
+
self.out = nn.Linear(hidden_size, output_size)
|
49 |
+
|
50 |
+
def forward(self, x, last_hidden, encoder_outputs):
|
51 |
+
embedded = self.embedding(x)
|
52 |
+
embedded = self.embedding_dropout(embedded)
|
53 |
+
decoder_outputs, hidden = self.rnn(embedded, last_hidden)
|
54 |
+
attn_weights = self.attn(decoder_outputs, encoder_outputs)
|
55 |
+
|
56 |
+
context = attn_weights.bmm(encoder_outputs).squeeze(1)
|
57 |
+
|
58 |
+
concat_input = torch.cat((decoder_outputs.squeeze(1), context), 1)
|
59 |
+
concat_output = torch.tanh(self.concat(concat_input))
|
60 |
+
output = self.out(concat_output)
|
61 |
+
|
62 |
+
output = F.softmax(output, dim=1)
|
63 |
+
return output, hidden
|
64 |
+
|
65 |
+
|
66 |
+
class Seq2SeqChatbot(nn.Module):
|
67 |
+
def __init__(self, hidden_size: int, vocab_size: int, encoder_num_layers: int, decoder_num_layers: int, decoder_embedding_dropout: float, device: torch.device):
|
68 |
+
super().__init__()
|
69 |
+
self.hidden_size = hidden_size
|
70 |
+
self.encoder_num_layers = encoder_num_layers
|
71 |
+
self.decoder_num_layers = decoder_num_layers
|
72 |
+
self.decoder_embedding_dropout = decoder_embedding_dropout
|
73 |
+
self.vocab_size = vocab_size
|
74 |
+
self.epoch = 0
|
75 |
+
|
76 |
+
self.device = device
|
77 |
+
self.vocab = Vocab([])
|
78 |
+
self.embedding = nn.Embedding(vocab_size, hidden_size)
|
79 |
+
self.attn = LuongAttention(Method.DOT, hidden_size)
|
80 |
+
self.encoder = Seq2SeqEncoder(hidden_size, hidden_size, encoder_num_layers, self.embedding)
|
81 |
+
self.decoder = Seq2SeqDecoder(hidden_size, hidden_size, vocab_size, decoder_num_layers, self.attn, self.embedding, decoder_embedding_dropout)
|
82 |
+
self.encoder_optimizer = optim.Adam(self.encoder.parameters())
|
83 |
+
self.decoder_optimizer = optim.Adam(self.decoder.parameters())
|
84 |
+
self.searcher = GreedySearch(self.encoder, self.decoder, self.embedding, device)
|
85 |
+
self.to(device)
|
86 |
+
self.eval_mode()
|
87 |
+
|
88 |
+
def train(self, epochs, train_data, teacher_forcing_ratio, device, save_dir, model_name, clip, save_every):
|
89 |
+
def maskNLLLoss(inp, target, mask):
|
90 |
+
crossEntropy = -torch.log(torch.gather(inp, 1, target.view(-1, 1)).squeeze(1))
|
91 |
+
loss = crossEntropy.masked_select(mask).mean()
|
92 |
+
loss = loss.to(device)
|
93 |
+
return loss
|
94 |
+
|
95 |
+
epoch_progress = tqdm(range(self.epoch, self.epoch + epochs), desc="Training", unit="epoch", leave=True)
|
96 |
+
epoch_progress.set_description(f"maskNLLLoss: None")
|
97 |
+
|
98 |
+
for epoch in epoch_progress:
|
99 |
+
for x_train, y_train, x_lengths, y_mask in train_data:
|
100 |
+
self.encoder_optimizer.zero_grad()
|
101 |
+
self.decoder_optimizer.zero_grad()
|
102 |
+
# Squeeze because batches are made in dataset and DataLoader is only for shuffling
|
103 |
+
x_train = x_train.squeeze(0).to(device)
|
104 |
+
y_train = y_train.squeeze(0).to(device)
|
105 |
+
x_lengths = x_lengths.squeeze(0) # Lengths are computed on CPU
|
106 |
+
y_mask = y_mask.squeeze(0).to(device)
|
107 |
+
|
108 |
+
encoder_outputs, hidden = self.encoder(x_train, x_lengths) # Output shape: (batch_size, max_len_in_batch, hidden_size)
|
109 |
+
hidden = hidden[:self.decoder_num_layers]
|
110 |
+
loss = 0
|
111 |
+
decoder_input = torch.LongTensor([[Token.BOS_TOKEN.value] for _ in range(y_train.shape[0])])
|
112 |
+
decoder_input = decoder_input.to(device)
|
113 |
+
use_teacher_forcing = random.random() < teacher_forcing_ratio
|
114 |
+
if use_teacher_forcing:
|
115 |
+
for t in range(y_train.shape[1]): # Process words in all batches for timestep t
|
116 |
+
decoder_outputs, hidden = self.decoder(decoder_input, hidden, encoder_outputs)
|
117 |
+
decoder_input = y_train[:, t].unsqueeze(1)
|
118 |
+
mask_loss = maskNLLLoss(decoder_outputs, y_train[:, t], y_mask[:, t])
|
119 |
+
loss += mask_loss
|
120 |
+
else:
|
121 |
+
for t in range(y_train.shape[1]):
|
122 |
+
decoder_outputs, hidden = self.decoder(decoder_input, hidden, encoder_outputs)
|
123 |
+
decoder_input = torch.argmax(decoder_outputs, dim=1).unsqueeze(1)
|
124 |
+
mask_loss = maskNLLLoss(decoder_outputs, y_train[:, t], y_mask[:, t])
|
125 |
+
loss += mask_loss
|
126 |
+
|
127 |
+
loss.backward()
|
128 |
+
|
129 |
+
_ = nn.utils.clip_grad_norm_(self.encoder.parameters(), clip)
|
130 |
+
_ = nn.utils.clip_grad_norm_(self.decoder.parameters(), clip)
|
131 |
+
|
132 |
+
self.encoder_optimizer.step()
|
133 |
+
self.decoder_optimizer.step()
|
134 |
+
|
135 |
+
if (epoch % save_every == 0 and epoch != 0) or epoch == save_every - 1:
|
136 |
+
directory = os.path.join(save_dir, model_name, '{}-{}'.format(self.encoder_num_layers, self.decoder_num_layers, self.hidden_size))
|
137 |
+
if not os.path.exists(directory):
|
138 |
+
os.makedirs(directory)
|
139 |
+
torch.save({
|
140 |
+
'epoch': epoch + self.epoch,
|
141 |
+
'en': self.encoder.state_dict(),
|
142 |
+
'de': self.decoder.state_dict(),
|
143 |
+
'en_opt': self.encoder_optimizer.state_dict(),
|
144 |
+
'de_opt': self.decoder_optimizer.state_dict(),
|
145 |
+
'loss': loss,
|
146 |
+
'voc_dict': self.vocab.__dict__,
|
147 |
+
'embedding': self.embedding.state_dict()
|
148 |
+
}, os.path.join(directory, '{}_{}.tar'.format(epoch, 'checkpoint')))
|
149 |
+
|
150 |
+
epoch_progress.set_description(f"maskNLLLoss: {loss:.8f}")
|
151 |
+
|
152 |
+
def to(self, device):
|
153 |
+
self.encoder = self.encoder.to(device)
|
154 |
+
self.decoder = self.decoder.to(device)
|
155 |
+
self.embedding = self.embedding.to(device)
|
156 |
+
self.attn = self.attn.to(device)
|
157 |
+
|
158 |
+
def train_mode(self):
|
159 |
+
self.encoder.train()
|
160 |
+
self.decoder.train()
|
161 |
+
self.embedding.train()
|
162 |
+
self.attn.train()
|
163 |
+
|
164 |
+
def eval_mode(self):
|
165 |
+
self.encoder.eval()
|
166 |
+
self.decoder.eval()
|
167 |
+
self.embedding.eval()
|
168 |
+
self.attn.eval()
|
169 |
+
|
170 |
+
def load_checkpoint(self, checkpoint_path: str):
|
171 |
+
checkpoint = torch.load(checkpoint_path, map_location=self.device, weights_only=False)
|
172 |
+
encoder_sd = checkpoint["en"]
|
173 |
+
decoder_sd = checkpoint["de"]
|
174 |
+
embedding_sd = checkpoint["embedding"]
|
175 |
+
self.vocab.__dict__ = checkpoint["voc_dict"]
|
176 |
+
encoder_optimizer_sd = checkpoint["en_opt"]
|
177 |
+
decoder_optimizer_sd = checkpoint["de_opt"]
|
178 |
+
self.epoch = checkpoint["epoch"]
|
179 |
+
|
180 |
+
self.encoder_optimizer.load_state_dict(encoder_optimizer_sd)
|
181 |
+
self.decoder_optimizer.load_state_dict(decoder_optimizer_sd)
|
182 |
+
self.embedding.load_state_dict(embedding_sd)
|
183 |
+
self.encoder.load_state_dict(encoder_sd)
|
184 |
+
self.decoder.load_state_dict(decoder_sd)
|
185 |
+
|
186 |
+
def forward(self, input_seq: str):
|
187 |
+
input_seq = ChatDataset._ChatDataset__normalize(input_seq)
|
188 |
+
input_seq = self.vocab.sentence_indices(input_seq + ["<eos>"]).unsqueeze(0).to(self.device)
|
189 |
+
output, _ = self.searcher(input_seq, torch.tensor(input_seq.shape[1]).view(1), 10)
|
190 |
+
output = [self.vocab.index2word[i.item()] for i in output]
|
191 |
+
output = [word for word in output if word not in ("<bos>", "<eos>", "<pad>")]
|
192 |
+
return " ".join(output)
|
193 |
+
|
194 |
+
|
195 |
+
if __name__ == "__main__": # Run as module
|
196 |
+
from .chat_dataset import ChatDataset
|
197 |
+
import torch.utils.data as data
|
198 |
+
|
199 |
+
CHAT_HISTORY_PATH = "models/seq2seq/data/train/chat_history.json"
|
200 |
+
batch_size = 20
|
201 |
+
chat_dataset = ChatDataset(CHAT_HISTORY_PATH, max_message_count=10_000, batch_size=batch_size)
|
202 |
+
train_data = data.DataLoader(chat_dataset, batch_size=1, shuffle=True)
|
203 |
+
|
204 |
+
device = torch.device("cpu")
|
205 |
+
chatbot = Seq2SeqChatbot(500, chat_dataset.vocab.size, 2, 2, 0.1, device)
|
206 |
+
chatbot.load_checkpoint("models/seq2seq/checkpoint/150_checkpoint.tar")
|
207 |
+
chatbot.train_mode()
|
208 |
+
chatbot.train(3, train_data, 0.5, device, "./checkpoint/temp/", "frantics_fox", 50.0, 100)
|
models/seq2seq/searchers.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from .custom_types import Token
|
4 |
+
|
5 |
+
class GreedySearch(nn.Module):
|
6 |
+
def __init__(self, encoder, decoder, embedding, device):
|
7 |
+
super().__init__()
|
8 |
+
self.encoder = encoder
|
9 |
+
self.decoder = decoder
|
10 |
+
self.embedding = embedding
|
11 |
+
self.device = device
|
12 |
+
|
13 |
+
def forward(self, x, input_length, max_length):
|
14 |
+
encoder_outputs, hidden = self.encoder(x, input_length)
|
15 |
+
decoder_hidden = hidden[:self.decoder.num_layers]
|
16 |
+
decoder_input = torch.ones(1, 1, device=self.device, dtype=torch.long) * Token.BOS_TOKEN.value
|
17 |
+
all_tokens = torch.zeros([0], device=self.device, dtype=torch.long)
|
18 |
+
all_scores = torch.zeros([0], device=self.device)
|
19 |
+
|
20 |
+
for _ in range(max_length):
|
21 |
+
decoder_outputs, decoder_hidden = self.decoder(decoder_input, decoder_hidden, encoder_outputs)
|
22 |
+
decoder_scores, decoder_input = torch.max(decoder_outputs, dim=1)
|
23 |
+
all_tokens = torch.cat((all_tokens, decoder_input), dim=0)
|
24 |
+
all_scores = torch.cat((all_scores, decoder_scores), dim=0)
|
25 |
+
decoder_input.unsqueeze_(0)
|
26 |
+
|
27 |
+
return all_tokens, all_scores
|
models/seq2seq/vocab.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from typing import List, Dict
|
4 |
+
from .custom_types import Token
|
5 |
+
|
6 |
+
|
7 |
+
class Vocab(nn.Module):
|
8 |
+
def __init__(self, messages: List[Dict]):
|
9 |
+
super().__init__()
|
10 |
+
self.word2index: Dict[str, int] = {"<pad>": Token.PAD_TOKEN.value, "<bos>": Token.BOS_TOKEN.value, "<eos>": Token.EOS_TOKEN.value, "<unk>": Token.UNK_TOKEN.value}
|
11 |
+
self.index2word: Dict[int, str] = {Token.PAD_TOKEN.value: "<pad>", Token.BOS_TOKEN.value: "<bos>", Token.EOS_TOKEN.value: "<eos>", Token.UNK_TOKEN.value: "<unk>"}
|
12 |
+
self.word_count: Dict[str, int] = dict()
|
13 |
+
self.size = 4
|
14 |
+
|
15 |
+
for message in messages:
|
16 |
+
self.add_sentence(message["text"])
|
17 |
+
|
18 |
+
self.embedding = nn.Embedding(self.size, 300)
|
19 |
+
|
20 |
+
def add_sentence(self, sentence):
|
21 |
+
for word in sentence:
|
22 |
+
self.add_word(word)
|
23 |
+
|
24 |
+
def add_word(self, word):
|
25 |
+
if word not in self.word2index:
|
26 |
+
self.word2index[word] = self.size
|
27 |
+
self.index2word[self.size] = word
|
28 |
+
self.word_count[word] = 1
|
29 |
+
self.size += 1
|
30 |
+
else:
|
31 |
+
self.word_count[word] += 1
|
32 |
+
|
33 |
+
def sentence_indices(self, sentence: List[str]) -> torch.LongTensor:
|
34 |
+
indices = torch.LongTensor(len(sentence))
|
35 |
+
for i, word in enumerate(sentence):
|
36 |
+
indices[i] = self.word2index[word] if word in self.word2index else Token.UNK_TOKEN.value
|
37 |
+
return indices
|
38 |
+
|
39 |
+
def forward(self, indices: torch.LongTensor):
|
40 |
+
return self.embedding(indices)
|
41 |
+
|
42 |
+
def __len__(self):
|
43 |
+
return self.size
|