DLPerf
commented
2 years ago Hello, I'm looking forward to your reply~
Open DLPerf opened 3 years ago
DLPerf
commented
2 years ago Hello, I'm looking forward to your reply~
Hello! I've found two types of performance issues in your program:
batch()should be called beforemap().tf.Sessionbeing defined repeatedly leads to incremental overhead.You can make your program more efficient by fixing the above two problems. Here are the tensorflow document and the Stack Overflow post to support this.
Below are detailed issues about
batch()should be called beforemap():dataset = dataset.batch(batch_size)(here) should be called beforedataset = dataset.map(parse_csv, num_parallel_calls=5)(here).dataset = dataset.batch(32)(here) should be called beforedataset = (tf.data.TextLineDataset(file_path).skip(1).map(decode_csv))(here)..batch(32)(here) should be called before.map(decode_csv, num_parallel_calls=4)(here).dataset = dataset.shuffle(1000).repeat().batch(batch_size)(here) should be called beforedataset = dataset.map(_parse_line)(here).Besides, you need to check the function called in
map()(e.g.,_parse_linecalled indataset = dataset.map(_parse_line)) whether to be affected or not to make the changed code work properly. For example, if_parse_lineneeds data with shape (x, y, z) as its input before fix, it would require data with shape (batch_size, x, y, z).Below are detailed issues about
tf.Sessionbeing defined repeatedly:with tf.Session() as sess:(here) is defined in functioneval_once(here) which is repeatedly called in a loopwhile True:(here).sess = tf.Session(master, graph=tf.get_default_graph())(here) is defined in function_run_checkpoint_once(here) which is repeatedly called in a loopwhile True:(here).with tf.Session() as sess:(here) is defined in functionrun_once(here) which is repeatedly called in a loopwhile True:(here).session = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))(here) is defined in functionrun_experiment(here) which is repeatedly called in a loopwhile paused < 360:(here).sess = tf.Session('')(here) is defined in a loopwhile True:(here).with tf.Session() as sess:(here) is defined in functionrun_once(here) which is repeatedly called in a loopwhile True:(here).with tf.Session() as sess:(here) is defined in function_eval_once(here) which is repeatedly called in a loopwhile True:(here).with tf.Session('') as sess:(here) is defined in function_add_to_tfrecord(here) which is repeatedly called in a loopfor i in range(_NUM_TRAIN_FILES):(here).sess = tf.Session()(here) is defined in functiondump_tfhub_to_hdf5(here) anddump_tfhub_to_hdf5is called in functionconvert_biggan(here) which is repeatedly called in a loopfor res in RESOLUTIONS:(here).with tf.Session() as session:(here) is defined in function_get_loss_acc(here) which is repeatedly called in a loopfor i, (weights, label) in enumerate(weight_init_list):(here).If you define
tf.Sessionout of the loop and passtf.Sessionas a parameter to the loop, your program would be much more efficient.Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.