Using the tensorgo API for TensorFlow Async Model Parallel
MIT License
Using the tensorgo API for TensorFlow Async Model Parallel
The system is designed to be simple to use, while maintaining efficiency speedup and approximate model performence(may be better). Three lines to transfer your model into a multi-gpu trainer.
from tensorgo.train.multigpu import MultiGpuTrainer
from tensorgo.train.config import TrainConfig
# [Define your own model using initial tensorflow API]
bow_model = ...
train_config = TrainConfig(dataset=training_dataset, model=bow_model, n_towers=5, commbatch=1500)
trainer = MultiGpuTrainer(train_config)
probs, labels = trainer.run([model.prob, model.label],
feed_dict={model.dropout_prob=0.2,
model.bacth_norm_on=True})