时间戳

首页 > 解决方案 > 为什么 Magenta 预处理的输出少于输入？

tensorflow - 为什么 Magenta 预处理的输出少于输入？

问题描述

我正在使用 Magenta 的Polyphony RNN来生成 MIDI 音乐。我有数据集，但是当涉及到 Magenta 的预处理时，它产生的输出比我预期的要少（它应该与输入相同）。我可以观察到有一些指标可以摆脱输入，但我真的不明白使用哪些或为什么使用它们。

我也不明白如何才能获得我没有作为输出获得的确切输入量（某些管道会产生多种变化）。

附在这篇文章之后，我将留下一张截图，说明管道处理完所有内容后的日志如何：

Polyphony RNN 管道日志：

INFO:tensorflow:Processed 3902 inputs total. Produced 819 outputs.
I0426 09:22:31.260396 140551050180416 pipeline.py:388] Processed 3902 inputs total. Produced 819 outputs.
INFO:tensorflow:DAGPipeline_PolyExtractor_eval_polyphonic_track_lengths_in_bars:
  [1,10): 9
  [10,20): 45
I0426 09:22:31.260454 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_eval_polyphonic_track_lengths_in_bars:
  [1,10): 9
  [10,20): 45
INFO:tensorflow:DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_more_than_1_program: 13317
I0426 09:22:31.260496 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_more_than_1_program: 13317
INFO:tensorflow:DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_too_long: 81
I0426 09:22:31.260534 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_too_long: 81
INFO:tensorflow:DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_too_short: 9135
I0426 09:22:31.260581 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_eval_polyphonic_tracks_discarded_too_short: 9135
INFO:tensorflow:DAGPipeline_PolyExtractor_training_polyphonic_track_lengths_in_bars:
  [1,10): 369
  [10,20): 234
  [20,30): 162
I0426 09:22:31.260628 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_training_polyphonic_track_lengths_in_bars:
  [1,10): 369
  [10,20): 234
  [20,30): 162
INFO:tensorflow:DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_more_than_1_program: 119675
I0426 09:22:31.260667 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_more_than_1_program: 119675
INFO:tensorflow:DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_too_long: 711
I0426 09:22:31.260702 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_too_long: 711
INFO:tensorflow:DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_too_short: 133902
I0426 09:22:31.260736 140551050180416 statistics.py:137] DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_too_short: 133902
INFO:tensorflow:DAGPipeline_RandomPartition_eval_poly_tracks_count: 388
I0426 09:22:31.260771 140551050180416 statistics.py:137] DAGPipeline_RandomPartition_eval_poly_tracks_count: 388
INFO:tensorflow:DAGPipeline_RandomPartition_training_poly_tracks_count: 3514
I0426 09:22:31.260805 140551050180416 statistics.py:137] DAGPipeline_RandomPartition_training_poly_tracks_count: 3514
INFO:tensorflow:DAGPipeline_TranspositionPipeline_eval_skipped_due_to_range_exceeded: 30
I0426 09:22:31.260839 140551050180416 statistics.py:137] DAGPipeline_TranspositionPipeline_eval_skipped_due_to_range_exceeded: 30
INFO:tensorflow:DAGPipeline_TranspositionPipeline_eval_transpositions_generated: 22587
I0426 09:22:31.260874 140551050180416 statistics.py:137] DAGPipeline_TranspositionPipeline_eval_transpositions_generated: 22587
INFO:tensorflow:DAGPipeline_TranspositionPipeline_training_skipped_due_to_range_exceeded: 187
I0426 09:22:31.260908 140551050180416 statistics.py:137] DAGPipeline_TranspositionPipeline_training_skipped_due_to_range_exceeded: 187
INFO:tensorflow:DAGPipeline_TranspositionPipeline_training_transpositions_generated: 255053

过滤这些歌曲的相关代码片段可以在第447 - 453行之间执行的度量中找到，称为DAGPipeline_PolyExtractor_training_polyphonic_tracks_discarded_more_than_1_program

先感谢您。

标签： tensorflowpipelinerecurrent-neural-networkmidimagenta

解决方案

推荐阅读