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
|
import ssl
ssl._create_default_https_context = (
ssl._create_unverified_context
) # Fixed problem with downloading CIFAR10 dataset
import haiku as hk
import torch
import einx
import os
import jax
import optax
import time
import torchvision
import torchvision.transforms as transforms
import jax.numpy as jnp
from functools import partial
import einx.nn.haiku as einn
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
batch_size = 256
rng = jax.random.PRNGKey(42)
def next_rng():
global rng
rng, x = jax.random.split(rng)
return x
cifar10_path = os.path.join(os.path.dirname(__file__), "cifar10")
trainset = torchvision.datasets.CIFAR10(
root=cifar10_path, train=True, download=True, transform=transform
)
trainloader = torch.utils.data.DataLoader(
trainset, batch_size=batch_size, shuffle=True, num_workers=2
)
testset = torchvision.datasets.CIFAR10(
root=cifar10_path, train=False, download=True, transform=transform
)
testloader = torch.utils.data.DataLoader(
testset, batch_size=batch_size, shuffle=False, num_workers=2
)
class Net(hk.Module):
def __call__(self, x, training):
for c in [1024, 512, 256]:
x = einn.Linear("b [...->c]", c=c)(x)
x = einn.Norm("[b] c", decay_rate=0.99)(x, training=training)
x = jax.nn.gelu(x)
x = einn.Dropout("[...]", drop_rate=0.2)(x, training=training)
x = einn.Linear("b [...->c]", c=10)(x)
return x
net = hk.transform_with_state(lambda x, training: Net()(x, training))
inputs, labels = next(iter(trainloader))
params, state = net.init(rng=next_rng(), x=jnp.asarray(inputs), training=True) # Run on dummy batch
optimizer = optax.adam(3e-4)
opt_state = optimizer.init(params)
@partial(jax.jit, donate_argnums=(0, 1, 2))
def update_step(opt_state, params, state, images, labels, rng):
def loss_fn(params, state):
logits, new_state = net.apply(params, state, rng, images, training=True)
one_hot = jax.nn.one_hot(labels, 10)
loss = jnp.mean(optax.softmax_cross_entropy(logits=logits, labels=one_hot))
return loss, new_state
(_loss, new_state), grads = jax.value_and_grad(loss_fn, has_aux=True)(params, state)
updates, new_opt_state = optimizer.update(grads, opt_state, params)
new_params = optax.apply_updates(params, updates)
return new_opt_state, new_params, new_state
@jax.jit
def test_step(params, state, images, labels):
logits, _ = net.apply(params, state, rng, images, training=False)
accurate = jnp.argmax(logits, axis=1) == jnp.asarray(labels)
return accurate
print("Starting training")
for epoch in range(100):
t0 = time.time()
# Train
for data in trainloader:
inputs, labels = data
opt_state, params, state = update_step(
opt_state, params, state, jnp.asarray(inputs), jnp.asarray(labels), next_rng()
)
# Test
correct = 0
total = 0
for data in testloader:
images, labels = data
accurate = test_step(params, state, jnp.asarray(images), jnp.asarray(labels))
total += accurate.shape[0]
correct += jnp.sum(accurate)
print(
f"Test accuracy after {epoch + 1:5d} epochs: {float(correct) / total} "
f"({time.time() - t0:.2f}sec)"
)
|
