logging.basicConfig(level=logging.DEBUG)
# for single gpu
# kv = mx.kvstore.create('local')
# batch_size = 48
# devs = mx.gpu(0)
# dist_async - async sgd
# dist_sync - BSP sgd
kv = mx.kvstore.create('dist_async')
# assume each worker has two gpus
devs = [mx.gpu(i) for i in range(2)]
batch_size = 96
(train, val) = imagenet.ilsvrc12(num_parts=kv.num_workers,
part_index=kv.rank,
batch_size=batch_size,
input_shape=(3, 224, 224))
model = mx.model.FeedForward(ctx=devs,
symbol=imagenet.inception(1000),
num_epoch=20,
learning_rate=0.05,
momentum=0.9,
wd=0.00001)
model.fit(X=train,
eval_data=val,
kvstore=kv,
batch_end_callback=mx.callback.Speedometer(batch_size, 10))
# kv_type = 'local'
## distributed version, can be dist_async or dist_sync
kv_type = 'dist_async'
## batch size for one gpu
batch_size_per_gpu = 40
## number of gpus used in a worker
num_gpus = 1
## learning rate
learning_rate = 0.05
batch_size = batch_size_per_gpu * num_gpus
kv = mx.kvstore.create(kv_type)
(train, val) = imagenet.ilsvrc12(data_dir=data_dir,
num_parts=kv.num_workers,
part_index=kv.rank,
batch_size=batch_size)
logging.basicConfig(level=logging.DEBUG)
model = mx.model.FeedForward(ctx=[mx.gpu(i) for i in range(num_gpus)],
symbol=imagenet.inception(1000),
num_epoch=20,
epoch_size=1281167 / batch_size / kv.num_workers,
learning_rate=learning_rate,
momentum=0.9,
wd=0.00001)
model.fit(X=train,
eval_data=val,
kvstore=kv,
#!/usr/bin/env python
import mxnet as mx
import logging
import imagenet
logging.basicConfig(level=logging.DEBUG)
kv = mx.kvstore.create("dist_sync")
batch_size = 96
(train, val) = imagenet.ilsvrc12(
num_parts=kv.num_workers, part_index=kv.rank, batch_size=batch_size, input_shape=(3, 224, 224)
)
# assume each worker has two gpus
devs = [mx.gpu(i) for i in range(2)]
model = mx.model.FeedForward(
ctx=devs, symbol=imagenet.inception(1000), num_round=20, learning_rate=0.05, momentum=0.9, wd=0.00001
)
model.fit(X=train, eval_data=val, kvstore=kv, epoch_end_callback=mx.callback.Speedometer(batch_size, 5))
## distributed version, can be dist_async or dist_sync
kv_type = 'dist_async'
## batch size for one gpu
batch_size_per_gpu = 40
## number of gpus used in a worker
num_gpus = 1
## learning rate
learning_rate = 0.05
batch_size = batch_size_per_gpu * num_gpus
kv = mx.kvstore.create(kv_type)
(train, val) = imagenet.ilsvrc12(
data_dir = data_dir,
num_parts = kv.num_workers,
part_index = kv.rank,
batch_size = batch_size)
logging.basicConfig(level=logging.DEBUG)
model = mx.model.FeedForward(
ctx = [mx.gpu(i) for i in range(num_gpus)],
symbol = imagenet.inception(1000),
num_epoch = 20,
epoch_size = 1281167 / batch_size / kv.num_workers,
learning_rate = learning_rate,
momentum = 0.9,
wd = 0.00001)
model.fit(X = train,
