我正在使用本文中的代码在 Keras 中学习机器翻译。这篇文章的代码在 GPU 和 CPU 上运行良好。
现在我想利用 Google Colab TPU。代码没有按原样进行 TPU 化,我需要朝 TF 方向移动。
根据 TPU 的Fashion MNIST 笔记本,我在 Tensorflow 中使用 Keras 层,而不是相反。在进入 TPU 部分之前,我正在执行此转换以查看它是否仍可在 GPU 上运行。这意味着主要改变这个功能,从:
from keras.models import Sequential
from keras.layers import LSTM
from keras.layers import Dense
from keras.layers import Embedding
from keras.layers import RepeatVector
from keras.layers import TimeDistributed
# define NMT model
def define_model(src_vocab, tar_vocab, src_timesteps, tar_timesteps, n_units):
model = Sequential()
model.add(Embedding(src_vocab, n_units, input_length=src_timesteps, mask_zero=True))
model.add(LSTM(n_units))
model.add(RepeatVector(tar_timesteps))
model.add(LSTM(n_units, return_sequences=True))
model.add(TimeDistributed(Dense(tar_vocab, activation='softmax')))
return model
至:
import tensorflow as tf
# define NMT model
def define_model(src_vocab, tar_vocab, src_timesteps, tar_timesteps, n_units):
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Embedding(src_vocab, n_units, input_length=src_timesteps, mask_zero=True))
model.add(tf.keras.layers.LSTM(n_units))
model.add(tf.keras.layers.RepeatVector(tar_timesteps))
model.add(tf.keras.layers.LSTM(n_units, return_sequences=True))
model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(tar_vocab, activation='softmax')))
return model
然后我做
model = define_model(swh_vocab_size, eng_vocab_size, swh_length, eng_length, 256)
model.compile(optimizer='adam', loss='categorical_crossentropy')
model.fit(trainX, trainY, epochs=1, batch_size=64, validation_data=(testX, testY), callbacks=[checkpoint], verbose=2)
但是,当我跑步时,这会导致投诉:
lib\site-packages\tensorflow\python\ops\gradients_impl.py:112: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
然后在 GPU 内部适配期间,它在 BLAS 负载上失败,如下所示:
InternalError: Blas GEMM launch failed : a.shape=(64, 256), b.shape=(256, 256), m=64, n=256, k=256
[[{{node lstm/while/MatMul}} = MatMul[T=DT_FLOAT, _class=["loc:@training/Adam/gradients/lstm/while/strided_slice_grad/StridedSliceGrad"], transpose_a=false, transpose_b=false, _device="/job:localhost/replica:0/task:0/device:GPU:0"](lstm/while/TensorArrayReadV3, lstm/while/strided_slice)]]
[[{{node loss/time_distributed_loss/broadcast_weights/assert_broadcastable/AssertGuard/Assert/Switch/_175}} = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_2728_...ert/Switch", tensor_type=DT_BOOL, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]
这是在转换为 TPU 模型之前。在进行最终的 TPU 转换之前,我只是想确保事情仍然在 CPU 和 GPU 上运行。他们没有。关于为什么我不能走这么远的任何想法?