Pytorch在训练时冻结某些层使其不参与训练问题(更新梯度)

定义网络

# 定义一个简单的网络
class net(nn.Module):
    def __init__(self, num_class=):
        super(net, self).__init__()
        self.fc = nn.Linear(8, 4)
        self.fc = nn.Linear(4, num_class)
    
    
    def forward(self, x):
        return self.fc(self.fc1(x))

情况一：当不冻结层时

代码

model = net()

# 情况一：不冻结参数时
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=e-2)  # 传入的是所有的参数

# 训练前的模型参数
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

for epoch in range():
    x = torch.randn((, 8))
    label = torch.randint(,10,[3]).long()
    output = model(x)
    
    loss = loss_fn(output, label)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

# 训练后的模型参数
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

结果

(bbn) jyzhang@admin2-X10DAi:~/test$ python -u "/home/jyzhang/test/net.py"

model.fc1.weight Parameter containing:

tensor([[ 0.3362, -0.2676, -0.3497, -0.3009, -0.1013, -0.2316, -0.0189, 0.1430],

[-0.2486, 0.2900, -0.1818, -0.0942, 0.1445, 0.2410, -0.1407, -0.3176],

[-0.3198, 0.2039, -0.2249, 0.2819, -0.3136, -0.2794, -0.3011, -0.2270],

[ 0.3376, -0.0842, 0.2747, -0.0232, 0.0768, 0.3160, -0.1185, 0.2911]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[ 0.4277, 0.0945, 0.1768, 0.3773],

[-0.4595, -0.2447, 0.4701, 0.2873],

[ 0.3281, -0.1861, -0.2202, 0.4413],

[-0.1053, -0.1238, 0.0275, -0.0072],

[-0.4448, -0.2787, -0.0280, 0.4629],

[ 0.4063, -0.2091, 0.0706, 0.3216],

[-0.2287, -0.1352, -0.0502, 0.3434],

[-0.2946, -0.4074, 0.4926, -0.0832],

[-0.2608, 0.0165, 0.0501, -0.1673],

[ 0.2507, 0.3006, 0.0481, 0.2257]], requires_grad=True)

model.fc1.weight Parameter containing:

tensor([[ 0.3316, -0.2628, -0.3391, -0.2989, -0.0981, -0.2178, -0.0056, 0.1410],

[-0.2529, 0.2991, -0.1772, -0.0992, 0.1447, 0.2480, -0.1370, -0.3186],

[-0.3246, 0.2055, -0.2229, 0.2745, -0.3158, -0.2750, -0.2994, -0.2295],

[ 0.3366, -0.0877, 0.2693, -0.0182, 0.0807, 0.3117, -0.1184, 0.2946]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[ 0.4189, 0.0985, 0.1723, 0.3804],

[-0.4593, -0.2356, 0.4772, 0.2784],

[ 0.3269, -0.1874, -0.2173, 0.4407],

[-0.1061, -0.1248, 0.0309, -0.0062],

[-0.4322, -0.2868, -0.0319, 0.4647],

[ 0.4048, -0.2150, 0.0692, 0.3228],

[-0.2252, -0.1353, -0.0433, 0.3396],

[-0.2936, -0.4118, 0.4875, -0.0782],

[-0.2625, 0.0192, 0.0509, -0.1670],

[ 0.2474, 0.3056, 0.0418, 0.2265]], requires_grad=True)

结论

当不冻结层时，随着训练的进行，模型中的可学习参数层的参数会发生改变

情况二：采用方式一冻结fc1层时

方式一

优化器传入所有的参数

optimizer = optim.SGD(model.parameters(), lr=e-2)  # 传入的是所有的参数

将要冻结层的参数的requires_grad置为False

for name, param in model.named_parameters():
    if "fc" in name:
        param.requires_grad = False

代码

# 情况二：采用方式一冻结fc层时
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=e-2)  # 优化器传入的是所有的参数

# 训练前的模型参数
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

# 冻结fc层的参数
for name, param in model.named_parameters():
    if "fc" in name:
        param.requires_grad = False

for epoch in range():
    x = torch.randn((, 8))
    label = torch.randint(,10,[3]).long()
    output = model(x)
 
    loss = loss_fn(output, label)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

结果

(bbn) jyzhang@admin2-X10DAi:~/test$ python -u "/home/jyzhang/test/net.py"

model.fc1.weight Parameter containing:

tensor([[ 0.3163, -0.1592, -0.2360, 0.1436, 0.1158, 0.0406, -0.0627, 0.0566],

[-0.1688, 0.3519, 0.2464, -0.2693, 0.1284, 0.0544, -0.0188, 0.2404],

[ 0.0738, 0.2013, 0.0868, 0.1396, -0.2885, 0.3431, -0.1109, 0.2549],

[ 0.1222, -0.1877, 0.3511, 0.1951, 0.2147, -0.0427, -0.3374, -0.0653]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[-0.1830, -0.3147, -0.1698, 0.3235],

[-0.1347, 0.3096, 0.4895, 0.1221],

[ 0.2735, -0.2238, 0.4713, -0.0683],

[-0.3150, -0.1905, 0.3645, 0.3766],

[-0.0340, 0.3212, 0.0650, 0.1380],

[-0.2500, 0.1128, -0.3338, -0.4151],

[ 0.0446, -0.4776, -0.3655, 0.0822],

[-0.1871, -0.0602, -0.4855, -0.3604],

[-0.3296, 0.0523, -0.3424, 0.2151],

[-0.2478, 0.1424, 0.4547, -0.1969]], requires_grad=True)

model.fc1.weight Parameter containing:

tensor([[ 0.3163, -0.1592, -0.2360, 0.1436, 0.1158, 0.0406, -0.0627, 0.0566],

[-0.1688, 0.3519, 0.2464, -0.2693, 0.1284, 0.0544, -0.0188, 0.2404],

[ 0.0738, 0.2013, 0.0868, 0.1396, -0.2885, 0.3431, -0.1109, 0.2549],

[ 0.1222, -0.1877, 0.3511, 0.1951, 0.2147, -0.0427, -0.3374, -0.0653]])

model.fc2.weight Parameter containing:

tensor([[-0.1821, -0.3155, -0.1637, 0.3213],

[-0.1353, 0.3130, 0.4807, 0.1245],

[ 0.2731, -0.2206, 0.4687, -0.0718],

[-0.3138, -0.1925, 0.3561, 0.3809],

[-0.0344, 0.3152, 0.0606, 0.1332],

[-0.2501, 0.1154, -0.3267, -0.4137],

[ 0.0400, -0.4723, -0.3586, 0.0808],

[-0.1823, -0.0667, -0.4854, -0.3543],

[-0.3285, 0.0547, -0.3388, 0.2166],

[-0.2497, 0.1410, 0.4551, -0.2008]], requires_grad=True)

结论

由实验的结果可以看出：只要设置requires_grad=False虽然传入模型所有的参数，仍然只更新requires_grad=True的层的参数

情况三：采用方式二冻结fc1层时

方式二

优化器传入不冻结的fc2层的参数

optimizer = optim.SGD(model.fc.parameters(), lr=1e-2)  # 优化器只传入fc2的参数

注：不需要将要冻结层的参数的requires_grad置为False

代码

# 情况三：采用方式二冻结fc层时
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.fc.parameters(), lr=1e-2)  # 优化器只传入fc2的参数
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

for epoch in range():
    x = torch.randn((, 8))
    label = torch.randint(,3,[3]).long()
    output = model(x)
 
    loss = loss_fn(output, label)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
 
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)

结果

model.fc1.weight Parameter containing:

tensor([[ 0.2519, -0.1772, -0.2229, 0.0711, -0.1681, 0.1233, -0.3217, -0.0412],

[ 0.2032, -0.2045, 0.2723, 0.3272, 0.1034, 0.1519, -0.0587, -0.3436],

[ 0.0470, 0.2379, 0.0590, 0.2400, 0.2280, 0.2045, -0.0229, -0.3484],

[-0.3023, -0.1195, 0.1792, -0.2173, -0.0492, 0.2640, -0.3511, -0.2845]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[-0.3263, -0.2938, -0.3516, -0.4578],

[-0.4549, -0.0060, 0.4696, -0.0174],

[-0.4841, 0.2861, 0.2658, 0.4483],

[-0.3093, 0.0977, -0.2735, 0.1033],

[-0.2421, 0.4489, -0.4649, 0.0110],

[-0.3671, 0.0182, -0.1027, -0.4441],

[ 0.0205, -0.0659, 0.4183, -0.2068],

[-0.1846, 0.1741, -0.2302, -0.1745],

[-0.3423, -0.2642, 0.2796, 0.4976],

[-0.0770, -0.3766, -0.0512, -0.2105]], requires_grad=True)

model.fc1.weight Parameter containing:

tensor([[ 0.2519, -0.1772, -0.2229, 0.0711, -0.1681, 0.1233, -0.3217, -0.0412],

[ 0.2032, -0.2045, 0.2723, 0.3272, 0.1034, 0.1519, -0.0587, -0.3436],

[ 0.0470, 0.2379, 0.0590, 0.2400, 0.2280, 0.2045, -0.0229, -0.3484],

[-0.3023, -0.1195, 0.1792, -0.2173, -0.0492, 0.2640, -0.3511, -0.2845]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[-0.3253, -0.2973, -0.3707, -0.4560],

[-0.4566, 0.0015, 0.4655, -0.0166],

[-0.4796, 0.2931, 0.2592, 0.4661],

[-0.3097, 0.0966, -0.2695, 0.1002],

[-0.2433, 0.4455, -0.4587, 0.0063],

[-0.3669, 0.0171, -0.0988, -0.4452],

[ 0.0198, -0.0679, 0.4203, -0.2088],

[-0.1854, 0.1717, -0.2241, -0.1781],

[-0.3429, -0.2653, 0.2822, 0.4938],

[-0.0773, -0.3765, -0.0464, -0.2127]], requires_grad=True)

结论

当优化器只传入要更新的层的参数时，只会更新优化器传入的参数，对于没有传入的参数可以求导，但是仍然不会更新参数

方式一与方式二对比总结

在训练过程中可能需要固定一部分模型的参数，只更新另一部分参数。

有两种思路实现这个目标，一个是设置不要更新参数的网络层为false，另一个就是在定义优化器时只传入要更新的参数。

最优做法是，优化器只传入requires_grad=True的参数，这样占用的内存会更小一点，效率也会更高。

最优写法

最优写法

将不更新的参数的requires_grad设置为False，同时不将该参数传入optimizer

将不更新的参数的requires_grad设置为False

# 冻结fc层的参数
for name, param in model.named_parameters():
    if "fc" in name:
        param.requires_grad = False

不将不更新的模型参数传入optimizer

# 定义一个fliter，只传入requires_grad=True的模型参数
optimizer = optim.SGD(filter(lambda p : p.requires_grad, model.parameters()), lr=e-2)

代码

# 最优写法
loss_fn = nn.CrossEntropyLoss()

# # 训练前的模型参数
print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)
print("model.fc.weight.requires_grad:", model.fc1.weight.requires_grad)
print("model.fc.weight.requires_grad:", model.fc2.weight.requires_grad)

# 冻结fc层的参数
for name, param in model.named_parameters():
    if "fc" in name:
        param.requires_grad = False

optimizer = optim.SGD(filter(lambda p : p.requires_grad, model.parameters()), lr=e-2)  # 定义一个fliter，只传入requires_grad=True的模型参数

for epoch in range():
    x = torch.randn((, 8))
    label = torch.randint(,3,[3]).long()
    output = model(x)
 
    loss = loss_fn(output, label)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

print("model.fc.weight", model.fc1.weight)
print("model.fc.weight", model.fc2.weight)
print("model.fc.weight.requires_grad:", model.fc1.weight.requires_grad)
print("model.fc.weight.requires_grad:", model.fc2.weight.requires_grad)

结果

(bbn) jyzhang@admin2-X10DAi:~/test$ python -u "/home/jyzhang/test/net.py"

model.fc1.weight Parameter containing:

tensor([[-0.1193, 0.2354, 0.2520, 0.1187, 0.2699, -0.2301, 0.1622, -0.0478],

[-0.2862, -0.1716, 0.2865, 0.2615, -0.2205, -0.2046, -0.0983, -0.1564],

[-0.3143, -0.2248, 0.2198, 0.2338, 0.1184, -0.2033, -0.3418, 0.1434],

[ 0.3107, -0.0411, -0.3016, 0.1924, -0.1756, -0.2881, 0.0528, -0.0444]],

requires_grad=True)

model.fc2.weight Parameter containing:

tensor([[-0.2548, 0.2107, -0.1293, -0.2562],

[-0.1989, -0.2624, 0.2226, 0.4861],

[-0.1501, 0.2516, 0.4311, -0.1650],

[ 0.0334, -0.0963, -0.1731, 0.1706],

[ 0.2451, -0.2102, 0.0499, 0.0497],

[-0.1464, -0.2973, 0.3692, 0.0523],

[ 0.1192, 0.3575, -0.1911, 0.1457],

[-0.0990, 0.2059, 0.2072, -0.2013],

[-0.4397, 0.4036, -0.3402, -0.0417],

[ 0.0379, 0.0128, -0.3212, -0.0867]], requires_grad=True)

model.fc1.weight.requires_grad: True

model.fc2.weight.requires_grad: True

model.fc1.weight Parameter containing:

tensor([[-0.1193, 0.2354, 0.2520, 0.1187, 0.2699, -0.2301, 0.1622, -0.0478],

[-0.2862, -0.1716, 0.2865, 0.2615, -0.2205, -0.2046, -0.0983, -0.1564],

[-0.3143, -0.2248, 0.2198, 0.2338, 0.1184, -0.2033, -0.3418, 0.1434],

[ 0.3107, -0.0411, -0.3016, 0.1924, -0.1756, -0.2881, 0.0528, -0.0444]])

model.fc2.weight Parameter containing:

tensor([[-0.2637, 0.2073, -0.1293, -0.2422],

[-0.2027, -0.2641, 0.2152, 0.4897],

[-0.1543, 0.2504, 0.4188, -0.1576],

[ 0.0356, -0.0947, -0.1698, 0.1669],

[ 0.2474, -0.2081, 0.0536, 0.0456],

[-0.1445, -0.2962, 0.3708, 0.0500],

[ 0.1219, 0.3574, -0.1876, 0.1404],

[-0.0961, 0.2058, 0.2091, -0.2046],

[-0.4368, 0.4039, -0.3376, -0.0450],

[ 0.0398, 0.0143, -0.3181, -0.0897]], requires_grad=True)

model.fc1.weight.requires_grad: False

model.fc2.weight.requires_grad: True

结论

最优写法能够节省显存和提升速度：

节省显存：不将不更新的参数传入optimizer

提升速度：将不更新的参数的requires_grad设置为False，节省了计算这部分参数梯度的时间

Pytorch在训练时冻结某些层使其不参与训练问题(更新梯度)

目录

定义网络

情况一：当不冻结层时

情况二：采用方式一冻结fc1层时

方式一

情况三：采用方式二冻结fc1层时

方式二

方式一与方式二对比总结

最优写法