File size: 4,325 Bytes
5775f48 |
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 |
from functools import partial
import jax
from typing import Any, Callable, Sequence, Optional, NewType
from jax import lax, random, vmap, numpy as jnp
from jax.experimental.ode import odeint
import flax
from flax.training import train_state
from flax import traverse_util
from flax.core import freeze, unfreeze
from flax import linen as nn
from flax import serialization
import optax
import tensorflow_datasets as tfds
import numpy as np
from tqdm import tqdm
import os
# Define loss
@jax.jit
def cross_entropy_loss(logits, labels):
one_hot_labels = jax.nn.one_hot(labels, num_classes=10)
return -jnp.mean(jnp.sum(one_hot_labels * logits, axis=-1))
# Metric computation
@jax.jit
def compute_metrics(logits, labels):
loss = cross_entropy_loss(logits=logits, labels=labels)
accuracy = jnp.mean(jnp.argmax(logits, -1) == labels)
metrics = {
'loss': loss,
'accuracy': accuracy,
}
return metrics
def get_datasets():
"""Load MNIST train and test datasets into memory."""
ds_builder = tfds.builder('mnist')
ds_builder.download_and_prepare()
train_ds = tfds.as_numpy(ds_builder.as_dataset(split='train', batch_size=-1))
test_ds = tfds.as_numpy(ds_builder.as_dataset(split='test', batch_size=-1))
train_ds['image'] = jnp.float32(train_ds['image']) / 255.
test_ds['image'] = jnp.float32(test_ds['image']) / 255.
return train_ds, test_ds
def create_train_state(model, rng, learning_rate):
"""Creates initial 'TrainState'."""
params = model.init(rng, jnp.ones([1, 28, 28, 1]))['params']
tx = optax.adam(learning_rate)
return train_state.TrainState.create(
apply_fn=model.apply, params=params, tx=tx
)
# Training step
@jax.jit
def train_step(state, batch):
"""Train for a single step."""
def loss_fn(params):
logits = apply({'params': params}, batch['image'])
loss = cross_entropy_loss(logits=logits, labels=batch['label'])
return loss, logits
grad_fn = jax.value_and_grad(loss_fn, has_aux=True)
(_, logits), grads = grad_fn(state.params)
state = state.apply_gradients(grads=grads)
metrics = compute_metrics(logits=logits, labels=batch['label'])
return state, metrics
# Evaluation step
@jax.jit
def eval_step(params, batch):
logits = apply({'params': params}, batch['image'])
return compute_metrics(logits=logits, labels=batch['label'])
# Train function
def train_epoch(model, state, train_ds, batch_size, epoch, rng):
"""Train for a single epoch"""
train_ds_size = len(train_ds['image'])
steps_per_epoch = train_ds_size // batch_size
perms = jax.random.permutation(rng, len(train_ds['image']))
perms = perms[:steps_per_epoch * batch_size] # skip incomplete batch
perms = perms.reshape((steps_per_epoch, batch_size))
batch_metrics = []
for perm in tqdm(perms):
batch = {k: v[perm, ...] for k, v in train_ds.items()}
state, metrics = train_step(model, state, batch)
batch_metrics.append(metrics)
# compute mean of metrics across each batch in epoch.
batch_metrics_np = jax.device_get(batch_metrics)
epoch_metrics_np = {
k: np.mean([metrics[k] for metrics in batch_metrics_np])
for k in batch_metrics_np[0]
}
print('train epoch: %d, loss: %.4f, accuracy: %.2f' % (
epoch, epoch_metrics_np['loss'], epoch_metrics_np['accuracy'] * 100
))
return state
# Eval function
def eval_model(params, test_ds):
metrics = eval_step(params, test_ds)
metrics = jax.device_get(metrics)
summary = jax.tree_map(lambda x: x.item(), metrics)
return summary['loss'], summary['accuracy']
def train_and_evaluate(learning_rate, n_epoch, batch_size):
train_ds, test_ds = get_datasets()
rng = jax.random.PRNGKey(0)
rng, init_rng = jax.random.split(rng)
state = create_train_state(init_rng, learning_rate)
del init_rng # Must not be used anymore.
for epoch in tqdm(range(1, n_epoch + 1)):
rng, input_rng = jax.random.split(rng)
state = train_epoch(state, train_ds, batch_size, epoch, input_rng)
test_loss, test_accuracy = eval_model(state.params, test_ds)
print(' test epoch: %d, loss: %.2f, accuracy: %.2f' % (
epoch, test_loss, test_accuracy * 100
))
|