def prepare_solver_config(self):
use_gpu, cores = utils.get_cores()
gpu_counts = cores if use_gpu else 0
batch_size, actual_batch_size, val_batch_size = utils.get_batch_size()
step=[550,1650,2750,4400]
optimizer_params = {
"momentum": params.get_value("momentum", default=0.9),
"wd": params.get_value("wd", default=0.0005),
"learning_rate": params.get_value("learning_rate", default=0.01),
"lr_scheduler": mx.lr_scheduler.MultiFactorScheduler(step, factor=0.1),
}
self.solver_config = {
"gpu_counts": gpu_counts,
# fit_args
"optimizer": "SGD",
"optimizer_params": optimizer_params,
}
def prepare_train_config(self):
"""配置训练参数"""
# AVA-SDK 获取训练参数
"""
1) 获取所有配置 example
param_dict = params.get_all()
value1 = param_dict["key1"]
2) 获取某项value
value1 = params.get_value("key1", default=1)
"""
snapshot_interval_epochs = params.get_value(
"intervals.snapshotIntervalEpochs", default=1)
max_epochs = params.get_value("stopCondition.maxEpochs", default=3)
rand_crop = params.get_value("inputTransform.randomCrop", default=True)
rand_mirror = params.get_value("inputTransform.randomMirror",
default=True)
batch_size, actual_batch_size, val_batch_size = utils.get_batch_size()
# USING the trainning batch size as valadition batch size
val_batch_size = actual_batch_size
#crop_w, crop_h = utils.get_crop_size()
# 使用默认的一个较小的 crop_size
crop_w, crop_h = 16, 16
# 添加监控
snapshot_prefix = self.train_ins.get_snapshot_base_path() + "/snapshot"
kv_store = "device"
kv = mx.kvstore.create(kv_store)
'''
rank = int(kv.rank)
if rank > 0:
snapshot_prefix += "-%s" % rank
'''
batch_freq = 10 # 打印/上报指标的 batch 粒度
# AVA-SDK mxnet monitor callback 初始化
batch_end_cb = self.train_ins.get_monitor_callback(
"mxnet", batch_size=actual_batch_size, batch_freq=batch_freq)
epoch_end_cb = [
# mxnet default epoch callback
mx.callback.do_checkpoint(snapshot_prefix,
snapshot_interval_epochs),
self.train_ins.get_epoch_end_callback(
"mxnet",
epoch_interval=snapshot_interval_epochs,
other_files=[])
]
# 训练参数，用户可以自行配置
self.train_config = {
"input_data_shape": (CROP_CHANNELS, crop_h, crop_w),
"rand_crop": rand_crop,
"rand_mirror": rand_mirror,
"batch_size": batch_size,
"actual_batch_size": actual_batch_size,
"val_batch_size": val_batch_size,
# fit_args
"eval_metric":
mxnet_monitor.full_mxnet_metrics(), # AVA-SDK 获取mxnet metric 列表
"epoch_end_callback": epoch_end_cb,
"batch_end_callback": batch_end_cb,
"kvstore": kv,
"num_epoch": max_epochs,
}
def prepare_train_config(self):
"""配置训练参数"""
# AVA-SDK 获取训练参数
"""
1) 获取所有配置 example
param_dict = params.get_all()
value1 = param_dict["key1"]
2) 获取某项value
value1 = params.get_value("key1", default=1)
"""
snapshot_interval_epochs = params.get_value("snapshot_interval_epochs",
default=1)
max_epochs = params.get_value("max_epochs", default=3)
rand_crop = params.get_value("rand_crop", default=True)
rand_mirror = params.get_value("rand_mirror", default=True)
batch_size, actual_batch_size, val_batch_size = utils.get_batch_size()
batch_size = params.get_value("batchSize", default=8)
val_batch_size = params.get_value("valBatchSize", default=batch_size)
use_gpu, cores = utils.get_cores()
logger.info("Cores GPU=%s, count=%d", use_gpu, cores)
actual_batch_size = batch_size if not use_gpu else batch_size * cores
if use_gpu:
val_batch_size *= cores
# USING the trainning batch size as valadition batch size
val_batch_size = actual_batch_size
# crop_w, crop_h = utils.get_crop_size()
# 根据模型的输入要求选择 crop_size
crop_w = params.get_value("crop_w")
crop_h = params.get_value("crop_h")
# 添加监控
snapshot_prefix = self.train_ins.get_snapshot_base_path() + "/snapshot"
kv_store = "device"
kv = mx.kvstore.create(kv_store)
'''
rank = int(kv.rank)
if rank > 0:
snapshot_prefix += "-%s" % rank
'''
batch_freq = 10 # 打印/上报指标的 batch 粒度
batch_of_epoch = utils.ceil_by_level(
float(utils.get_sampleset_num() / actual_batch_size))
# AVA-SDK mxnet monitor callback 初始化
batch_end_cb = self.train_ins.get_monitor_callback(
"mxnet", batch_size=actual_batch_size, batch_freq=batch_freq)
epoch_end_cb = [
# mxnet default epoch callback
mx.callback.do_checkpoint(snapshot_prefix,
snapshot_interval_epochs),
self.train_ins.get_epoch_end_callback(
"mxnet",
batch_of_epoch=batch_of_epoch,
epoch_interval=snapshot_interval_epochs,
other_files=[])
]
# 训练参数，用户可以自行配置
self.train_config = {
"input_data_shape": (CROP_CHANNELS, crop_h, crop_w),
"rand_crop": rand_crop,
"rand_mirror": rand_mirror,
"batch_size": batch_size,
"actual_batch_size": actual_batch_size,
"val_batch_size": val_batch_size,
# fit_args
"eval_metric":
mxnet_monitor.full_mxnet_metrics(), # AVA-SDK 获取mxnet metric 列表
"epoch_end_callback": epoch_end_cb,
"batch_end_callback": batch_end_cb,
"kvstore": kv,
"num_epoch": max_epochs,
}