-
Notifications
You must be signed in to change notification settings - Fork 27
/
Copy pathflax_e2e_model.py
282 lines (240 loc) · 8.73 KB
/
flax_e2e_model.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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
# Copyright 2022 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Mnist example."""
import functools
from typing import Any
from absl import app
from aqt.jax.v2 import config as aqt_config
from aqt.jax.v2.flax import aqt_flax
from flax import linen as nn
from flax import struct
from flax.metrics import tensorboard
import jax
import jax.numpy as jnp
import numpy as np
import optax
import tensorflow_datasets as tfds
class CNN(nn.Module):
"""A simple CNN model."""
bn_use_stats: bool
aqt_cfg: aqt_config.DotGeneral
quant_mode: aqt_flax.QuantMode = aqt_flax.QuantMode.TRAIN
@nn.compact
def __call__(self, x):
aqt_dg = functools.partial(
aqt_flax.AqtDotGeneral,
self.aqt_cfg,
# In nn.Dense, it is RHS that has the kernel.
rhs_quant_mode=self.quant_mode,
)
use_running_avg = not self.bn_use_stats
x = nn.Conv(features=32, kernel_size=(3, 3))(x)
x = nn.BatchNorm(use_running_average=use_running_avg, dtype=x.dtype)(x)
x = nn.relu(x)
x = nn.avg_pool(x, window_shape=(2, 2), strides=(2, 2))
x = nn.Conv(features=64, kernel_size=(3, 3))(x)
x = nn.BatchNorm(use_running_average=use_running_avg, dtype=x.dtype)(x)
x = nn.relu(x)
x = nn.avg_pool(x, window_shape=(2, 2), strides=(2, 2))
x = x.reshape((x.shape[0], -1)) # flatten
x = nn.Dense(features=256, dot_general_cls=aqt_dg)(x)
x = nn.relu(x)
x = nn.Dense(features=10, dot_general_cls=aqt_dg)(x)
# Simple demonstration of how to quantize einsum.
identity = jnp.identity(10, dtype=x.dtype)
einsum = aqt_flax.AqtEinsum(self.aqt_cfg, lhs_quant_mode=self.quant_mode)
# Note for AQT developers:
# This equation is harder because jnp.einsum and einsum swap lhs and rhs.
x = einsum('bc,ab->ac', identity, x)
return x
@functools.partial(jax.jit, static_argnums=(3,))
def apply_model(state, images, labels, train):
"""Computes gradients, loss and accuracy for a single batch."""
cnn = state.cnn_train if train else state.cnn_eval
def loss_fn(model):
logits, updated_var = cnn.apply(
model,
images,
rngs={'params': jax.random.PRNGKey(0)},
mutable=True,
)
one_hot = jax.nn.one_hot(labels, 10)
loss = jnp.mean(optax.softmax_cross_entropy(logits=logits, labels=one_hot))
return loss, (logits, updated_var)
grad_fn = jax.value_and_grad(loss_fn, has_aux=True, allow_int=True)
aux, grads = grad_fn(state.model)
loss, (logits, updated_var) = aux
accuracy = jnp.mean(jnp.argmax(logits, -1) == labels)
return grads, loss, accuracy, updated_var
@jax.jit
def update_model(state, grads, updated_var):
params = state.model['params']
param_grad = grads['params']
updates, new_opt_state = state.tx.update(param_grad, state.opt_state, params)
new_params = optax.apply_updates(params, updates)
updated_var.update(params=new_params)
return state.replace(
model=updated_var,
opt_state=new_opt_state,
)
def train_epoch(state, train_ds, batch_size, 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))
epoch_loss = []
epoch_accuracy = []
for perm in perms:
batch_images = train_ds['image'][perm, ...]
batch_labels = train_ds['label'][perm, ...]
grads, loss, accuracy, updated_var = apply_model(
state, batch_images, batch_labels, train=True
)
state = update_model(state, grads, updated_var)
epoch_loss.append(loss)
epoch_accuracy.append(accuracy)
train_loss = np.mean(epoch_loss)
train_accuracy = np.mean(epoch_accuracy)
return state, train_loss, train_accuracy
def get_datasets():
"""Load MNIST train and test datasets into memory."""
print('get_datasets started')
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']) / 255.0
test_ds['image'] = jnp.float32(test_ds['image']) / 255.0
print('get_datasets DONE')
return train_ds, test_ds
class TrainState(struct.PyTreeNode):
"""Train state."""
cnn_train: Any = struct.field(pytree_node=False)
cnn_eval: Any = struct.field(pytree_node=False)
model: Any = struct.field(pytree_node=True)
tx: optax.GradientTransformation = struct.field(pytree_node=False)
opt_state: optax.OptState = struct.field(pytree_node=True)
def create_train_state(rng, aqt_cfg):
"""Creates initial `TrainState`."""
cnn_train = CNN(bn_use_stats=True, aqt_cfg=aqt_cfg)
model = cnn_train.init({'params': rng}, jnp.ones([1, 28, 28, 1]))
learning_rate = 0.1
momentum = 0.9
tx = optax.sgd(learning_rate, momentum)
cnn_eval = CNN(bn_use_stats=False, aqt_cfg=aqt_cfg)
return TrainState(
cnn_train=cnn_train,
cnn_eval=cnn_eval,
model=model,
tx=tx,
opt_state=tx.init(model['params']),
)
def train_and_evaluate(
num_epochs: int, workdir: str, aqt_cfg: aqt_config.DotGeneral
) -> TrainState:
"""Execute model training and evaluation loop."""
train_ds, test_ds = get_datasets()
rng = jax.random.key(0)
summary_writer = tensorboard.SummaryWriter(workdir)
rng, init_rng = jax.random.split(rng)
state = create_train_state(init_rng, aqt_cfg)
batch_size = 128
for epoch in range(1, num_epochs + 1):
rng, input_rng = jax.random.split(rng)
state, train_loss, train_accuracy = train_epoch(
state, train_ds, batch_size, input_rng
)
_, test_loss, test_accuracy, _ = apply_model(
state, test_ds['image'], test_ds['label'], train=False
)
print(
'epoch:% 3d, train_loss: %.30f, train_accuracy: %.30f, test_loss:'
' %.30f, test_accuracy: %.30f'
% (
epoch,
train_loss,
train_accuracy * 100,
test_loss,
test_accuracy * 100,
),
flush=True,
)
summary_writer.scalar('train_loss', train_loss, epoch)
summary_writer.scalar('train_accuracy', train_accuracy, epoch)
summary_writer.scalar('test_loss', test_loss, epoch)
summary_writer.scalar('test_accuracy', test_accuracy, epoch)
summary_writer.flush()
return state
def serving_conversion(train_state):
"""Model conversion (quantized weights freezing)."""
aqt_cfg = train_state.cnn_eval.aqt_cfg
input_shape = (1, 28, 28, 1)
cnn_freeze = CNN(
bn_use_stats=False,
aqt_cfg=aqt_cfg,
quant_mode=aqt_flax.QuantMode.CONVERT,
)
_, model_serving = cnn_freeze.apply(
train_state.model,
jnp.ones(input_shape),
rngs={'params': jax.random.PRNGKey(0)},
mutable=True,
)
cnn_serve = CNN(
bn_use_stats=False,
aqt_cfg=aqt_cfg,
quant_mode=aqt_flax.QuantMode.SERVE,
)
return cnn_serve.apply, model_serving
@jax.jit
def serve(state):
"""Take train state, freeze integer weights, and serve."""
# get sample serving data
_, test_ds = get_datasets()
sample_image, sample_label = test_ds['image'][:64], test_ds['label'][:64]
# serving
serve_fn, model_serving = serving_conversion(state)
logits = serve_fn(
model_serving, sample_image, rngs={'params': jax.random.PRNGKey(0)}
)
# compute serving loss
one_hot = jax.nn.one_hot(sample_label, 10)
loss = jnp.mean(optax.softmax_cross_entropy(logits=logits, labels=one_hot))
return loss
def serve_fn_hlo(state):
"""Example on how to inspect HLO to verify if everything was quantized."""
# get sample serving data
_, test_ds = get_datasets()
sample_image = test_ds['image'][:64]
# serving
serve_fn, model_serving = serving_conversion(state)
# The following XLA graph is only needed for debugging purpose
hlo = jax.xla_computation(serve_fn)(
model_serving,
sample_image,
rngs={'params': jax.random.PRNGKey(0)},
).as_hlo_module()
return hlo
def main(argv):
del argv
aqt_cfg = aqt_config.fully_quantized(fwd_bits=8, bwd_bits=8)
state = train_and_evaluate(
num_epochs=2, workdir='/tmp/aqt_mnist_example', aqt_cfg=aqt_cfg
)
loss = serve(state)
print('serve loss on sample ds: {}'.format(loss))
if __name__ == '__main__':
app.run(main)