I have been trying to get the raw logits output from the output layer. I loaded the pre-trained model and loaded the input layer. The shape is fixed to [1,16,19,26]. However, the input does not always have this shape. By performing the same operations as in do_single_file_inference on the input file, the features feed in will have shape [1, num_strides, 19, 26]. I have noticed that the second shape corresponds to a variable called n_steps in Deepspeech.BiRNN, but I could not understand what does this variable mean and why is it fixed in the pre-trained model. I apologize if this question might be trivial. It would be really pleased if anyone can explain this.
–n_steps: how many timesteps to process at once by the export graph, higher
values mean more latency
(default: ‘16’)
(an integer)