-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
141 lines (98 loc) · 3.94 KB
/
main.py
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
import torch
import torch.nn as nn
import config
from torch.utils.data import DataLoader, Dataset
import torch.nn.functional as F
from model import BigramBaseline, BigramAttn
import argparse
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
# read text
# get set of characters
# get char to number
# get number to char
torch.manual_seed(1337)
with open(config.data, "r") as f:
text = f.read()
chars = sorted(list(set(text)))
stn = {char: i for i, char in enumerate(chars)}
nts = {i: char for i, char in enumerate(chars)}
# enc/decode a sequence
encode = lambda s: [stn[char] for char in s]
decode = lambda enc: "".join([nts[num] for num in enc])
# create X, y batches
def create_batches(data):
idx = range(len(data) - config.chunk_size)
x = torch.stack([data[i:i+config.chunk_size] for i in idx])
y = torch.stack([data[i+1:i+config.chunk_size+1] for i in idx])
return x,y
def get_data():
data = torch.tensor(encode(text), dtype=torch.float32)
train_data = data[:int(0.9*data.shape[0])]
test_data = data[int(0.9*data.shape[0]):]
X_train, y_train = create_batches(train_data)
X_test, y_test = create_batches(test_data)
if config.device == "cuda":
X_train = X_train.to(config.device)
X_test = X_test.to(config.device)
y_train = y_train.to(config.device)
y_test = y_test.to(config.device)
return X_train, X_test, y_train, y_test
class CustomDataset(Dataset):
def __init__(self, x, y):
self.x = x
self.y = y
def __len__(self):
return self.x.shape[0]
def __getitem__(self, idx):
sample_x = self.x[idx]
sample_y = self.y[idx]
return sample_x, sample_y
def evaluate(X_test, y_test, model, criterion):
total_loss = 0.0
model.eval()
nb_batches = X_test.shape[0] // config.batch_size
for i in range(nb_batches):
# batch
x = X_test[config.batch_size*i:(config.batch_size*i) + config.batch_size].long()
target = y_test[config.batch_size*i:(config.batch_size*i) + config.batch_size].long().view(-1)
preds = model(x)
loss = criterion(preds, target)
total_loss += loss.item()
return total_loss / nb_batches
def main():
run_name = f" \
simple_bigram={config.nb_blocks}_lr={config.lr}_bs={config.batch_size}_chunk_size={config.chunk_size}_#head={config.nb_heads} \
_dim_head={config.dim_head}\
"
writer = SummaryWriter("runs/" + run_name)
X_train, X_test, y_train, y_test = get_data()
nb_batches = len(X_train) // config.batch_size
criterion = torch.nn.CrossEntropyLoss()
model = BigramAttn(len(chars)).to(config.device)
print(model)
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
print("Number of batches : ", nb_batches)
print("number params : ", model.get_number_params())
print("--- Training started ---", "\n")
for epoch in range(config.epochs):
model.train()
total_loss = 0.0
for i in tqdm(range(nb_batches)):
# batch
x = X_train[config.batch_size*i:(config.batch_size*i) + config.batch_size].long()
target = y_train[config.batch_size*i:(config.batch_size*i) + config.batch_size].long().view(-1)
preds = model(x)
loss = criterion(preds, target)
with torch.no_grad():
total_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
eval_loss = evaluate(X_test, y_test, model, criterion)
print("train : ", total_loss / nb_batches, "eval : ", eval_loss)
writer.add_scalar("Train/Loss", total_loss / nb_batches, epoch)
writer.add_scalar("Eval/Loss", eval_loss, epoch)
writer.add_text(f"text_at_epoch_{epoch}", "".join(decode(model.generate(500)[0].tolist())), epoch)
if __name__ == "__main__":
main()