search

weinman / cnn_lstm_ctc_ocr

Tensorflow-based CNN+LSTM trained with CTC-loss for OCR
GNU General Public License v3.0
498 stars 170 forks source link

TypeError #60

Closed w867066886 closed 3 years ago

w867066886 commented 5 years ago

There is a error " init() got an unexpected keyword argument 'train_distribute'" I search the function 'tf.estimator.RunConfig' and don't find the parameter "train_distribute" Coule you help me correct it? i use python2.7 and tensorflow1.5

weinman commented 5 years ago

This package requires a newer version of tensorflow. It's currently set to use tf≥1.10

w867066886 commented 5 years ago

Thansk for your answer , it is running now .

It takes a lot of the memory of gpu , but a low use of gpu-util. Im using four gtx1080ti ,but when i enter the "make test" ,it tell me " OOM when allocating tensor with shape[512,64,31,125] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc"

Should I stop trainning and to test? Look forward to your reply， thanks image

weinman commented 5 years ago

I've used the same device with no memory troubles in test. Odd.

You can reduce the batch size or run test on the CPU, since most of the computation (in test) happens there anyway. [That's why you'll have low GPU utilization.]

weinman commented 4 years ago

I think I (belatedly) understand now that perhaps you're running the test while you're also doing training? If your trainer has claimed all of the GPUs, the tester won't have room to allocate memory, because tensorflow (by default) tries to allocate all the available memory.