predict_convnet_10sec_model.py 文件源码

def predict_on_long_clips():
    """Load the saved model and perform inference/prediction on features obtained from inputs. 
    Splits the audio into 10second chunks and predicts on those chunks."""
    with open(FILENAMES,"r") as fh:
        filecontents=fh.read()
        filenames=filecontents.splitlines()
        random.shuffle(filenames)
        filenames=filenames[:5] #[:5] is for quickly verifying if things work
        filenames = [DATASET_LOCATION+f for f in filenames]

    session = tf.Session()
    saver = tf.train.import_meta_graph(IMPORT_META_GRAPH)
    saver.restore(session, tf.train.latest_checkpoint(IMPORT_LATEST_CHECKPOINT))
    tf.global_variables_initializer().run(session=session)

    test_x = {}
    for f in filenames:
        s, sr = librosa.load(f)
        total_chunks = s.shape[0]/max_audio_length
        waveforms = [s[max_audio_length*i:max_audio_length*(i+1)] for i in range(total_chunks)]
        test_x[f] = extract_features_from_waveforms(waveforms)

        print "FILENAME: ", f
        predictions = session.run(tf.argmax(pred, 1), feed_dict={X: test_x[f]})
        print [possible_categories[p] for p in predictions]