-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
175 lines (152 loc) · 5.7 KB
/
train.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
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
import torch
from tqdm import tqdm
import pdb
from torch import nn
import datasets
import transformers
from transformers import DataCollatorForSeq2Seq
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from arch import Transformer
def preprocess(tok, examples, max_length=100, padding="max_length"):
inputs = [ex["en"] for ex in examples["translation"]]
targets = [ex["de"] for ex in examples["translation"]]
model_inputs = tok(inputs, max_length=max_length, truncation=True, padding=padding)
labels = tok(targets, max_length=max_length, truncation=True, padding=padding)
model_inputs["labels"] = labels["input_ids"]
return model_inputs
class LabelSmoothLoss(nn.Module):
def __init__(self, pad_idx: int = 1, smooth: float = 0.1):
super().__init__()
self.smooth = smooth
self.confidence = 1 - smooth
self.pad_idx = pad_idx
self.criterion = nn.KLDivLoss(reduction="batchmean")
def forward(self, y_pred, y_true):
# Smooth
y_true *= self.confidence
y_true[y_true == 0] = self.smooth / (y_true.size(-1) - 1)
# Mask pad tokens
mask = y_true.argmax(-1) == self.pad_idx
y_pred = y_pred.clone()[mask]
y_true = y_true.clone()[mask]
return self.criterion(nn.functional.log_softmax(y_pred), y_true.to(torch.float))
def train():
tok = transformers.PreTrainedTokenizerFast.from_pretrained("bpe_tok")
batch_size = 4
max_seq_size = 1000
padding = "longest"
pad_idx = tok.vocab[tok.pad_token]
model_params = {
"src_vocab_size": len(tok.vocab),
"tgt_vocab_size": len(tok.vocab),
"embed_size": 512,
"num_layers": 6,
"heads": 8,
"device": "cuda",
"max_seq_len": max_seq_size,
"dropout": 0.1,
"src_pad_idx": pad_idx,
"tgt_pad_idx": pad_idx,
"ff_dim": 2048,
}
lr = 1e-3
betas = (0.9, 0.98)
train_steps = 100_000
eval_steps = 1_000
writer = SummaryWriter()
data = datasets.load_dataset("wmt14", "de-en").with_format("torch")
train_dataset = data["train"]
prep_train_dataset = train_dataset.map(
lambda examples: preprocess(
tok, examples, max_length=max_seq_size, padding=padding
),
batched=True,
)
prep_train_dataset = prep_train_dataset.remove_columns(
["translation", "attention_mask", "token_type_ids"]
)
val_dataset = data["validation"]
prep_val_dataset = val_dataset.map(
lambda examples: preprocess(
tok, examples, max_length=max_seq_size, padding=padding
),
batched=True,
)
prep_val_dataset = prep_val_dataset.remove_columns(
["translation", "attention_mask", "token_type_ids"]
)
print(prep_train_dataset[0])
collator = DataCollatorForSeq2Seq(
tok,
padding="max_length",
max_length=max_seq_size,
label_pad_token_id=tok.vocab[tok.pad_token],
)
train_loader = DataLoader(
prep_train_dataset, batch_size=batch_size, collate_fn=collator
)
val_loader = DataLoader(
prep_val_dataset, batch_size=batch_size, collate_fn=collator
)
model = Transformer(**model_params).to(model_params["device"])
print(model)
# criterion = nn.KLDivLoss(reduction="batchmean")
criterion = LabelSmoothLoss(pad_idx=pad_idx, smooth=0.1)
optimizer = torch.optim.Adam(model.parameters(), lr=lr, betas=betas, eps=1e-9)
def rate(step, model_size, factor, warmup):
"""
we have to default the step to 1 for LambdaLR function
to avoid zero raising to negative power.
"""
if step == 0:
step = 1
return factor * (
model_size ** (-0.5) * min(step ** (-0.5), step * warmup ** (-1.5))
)
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda step: rate(
step, model_size=model_params["embed_size"], factor=1.0, warmup=400
),
)
steps = 0
while steps < train_steps:
for inputs in tqdm(train_loader):
model.train()
y_pred = model(inputs["input_ids"], inputs["labels"])
loss = criterion(
torch.nn.functional.log_softmax(y_pred.to(torch.float64), dim=-1),
nn.functional.one_hot(
inputs["labels"].to(model_params["device"]),
model_params["tgt_vocab_size"],
).to(torch.float64),
)
writer.add_scalar("train/loss", loss.item(), steps)
writer.add_scalar("train/seq_len", y_pred.size(1), steps)
loss.backward()
optimizer.step()
optimizer.zero_grad()
lr_scheduler.step()
steps += 1
if steps % eval_steps == 0:
model.eval()
losses = []
with torch.no_grad():
for e_inputs in tqdm(val_loader):
y_pred = model(e_inputs["input_ids"], e_inputs["labels"])
loss = criterion(
torch.nn.functional.log_softmax(
y_pred.to(torch.float64), dim=-1
),
nn.functional.one_hot(
e_inputs["labels"].to(model_params["device"]),
model_params["tgt_vocab_size"],
).to(torch.float64),
)
# TODO: calc BLEU and log
losses.append(loss.item())
writer.add_scalar("val/loss", sum(losses) / len(losses), steps)
writer.flush()
if __name__ == "__main__":
train()