-
Notifications
You must be signed in to change notification settings - Fork 109
/
train.py
38 lines (33 loc) · 1.13 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
# -*- coding: utf-8 -*-
import torch
import numpy as np
import matplotlib.pyplot as plt
import torch.optim as optim
from tqdm import tqdm
from model import PTuneForLAMA
from load_data import train_loader
device = "cuda" if torch.cuda.is_available() else 'cpu'
template = (3,3,3)
N_EPOCHS = 5
LR = 1e-5
MAX_GRAD_NORM = 1.0
DECAY_RATE = 0.98
model = PTuneForLAMA(device, template)
model.train()
optimizer = optim.Adam(model.parameters(), lr=LR)
my_lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer=optimizer, gamma=DECAY_RATE)
loss_vals_train = []
for epoch in range(N_EPOCHS):
epoch_loss= []
pbar = tqdm(train_loader)
pbar.set_description("[Train Epoch {}]".format(epoch))
for batch_idx, batch_data in enumerate(pbar):
model.zero_grad()
loss = model(batch_data[0], batch_data[1])
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), MAX_GRAD_NORM)
epoch_loss.append(loss.item())
optimizer.step()
my_lr_scheduler.step()
loss_vals_train.append(np.mean(epoch_loss))
l1, = plt.plot(np.linspace(1, N_EPOCHS, N_EPOCHS).astype(int), loss_vals_train)