How to run this code

[YichiHuang]

Hi, I got a speech dataset, the length of each recording ranges from 2s~5s. How I run your code with my dataset? Can you give me some advice. Thanks.

[hrayrhar]

Hi,

First you should create the spectrograms of the recordings (you can use create_spectrograms.py for that), then make training and validation list files like this and finally, run the theano/main.py. As the length of recordings in your dataset is not constant you should either set batch_size=1 or do something to equalize the lengths in mini-batch.

[YichiHuang]

Hi, I followed your experiment procedure and run with my dataset. My dataset contains 50 classes, 400 recordings with 5 seconds for each category. First of all, I created spectrograms of each recording with size in 256x429. Then I split them in training set(1500 in total) and validation set(500 in total). Last, I run the main.py with default parameters except: --network==tc_net_mod I got the result: accuracy: 0.59 percent accuracy: 2.08 percent saving ... states/tc_net_mod.b32.bn.epoch499.test3.91200.state The accuracy is extremely low and I wonder there is something wrong with my experiment. Can you give some comments? Thanks.

[MurugeshMarvel]

getting error like this

[manbharae]

@YichiHuang can you please share your code. I trying to learn too.