Task 5: UnpicklingError

[mculp87]

Running the baseline code for Task 5, I ran across this error

Traceback (most recent call last):
File "~/Documents/dcase_util/dcase2018_baseline/task5/task5.py", line 763, in
sys.exit(main(sys.argv))
File "~/Documents/dcase_util/dcase2018_baseline/task5/task5.py", line 716, in main
overwrite=overwrite
File "~/Documents/dcase_util/dcase2018_baseline/task5/task5.py", line 237, in do_feature_normalization
features = dcase_util.containers.FeatureRepository().load(filename=feature_filename)
File "~/anaconda3/lib/python3.6/site-packages/dcase_util/containers/data.py", line 2149, in load
super(DataRepository, self).load(filename=self.filename)
File "~/anaconda3/lib/python3.6/site-packages/dcase_util/containers/containers.py", line 502, in load
dict.update(self, Serializer.load_cpickle(filename=self.filename))
_pickle.UnpicklingError: pickle data was truncated

during the Feature Normalization.

For most of the files, this never occurs. This warning is thrown on the extracted file
output_task5/features/feature_processing_chain_3b4ae3a13a504b182ef406346a732dc6/DevNode3_ex231_16.wav

I'm running this on Ubuntu 18.04, using:

• Name : Version : Build
• anaconda-client : 1.6.14 : py36_0
• anaconda-navigator : 1.8.7 : py36_0
• anaconda-project : 0.8.2 : py36h44fb852_0
• cloudpickle : 0.5.3 : py36_0
• pickleshare : 0.7.4 : py36h63277f8_0`

My dcase_util is at the master head, version 0.2.3. The dcase2018_baseline is at commit 82fd1e5. Using pip install -r requirements.txt all package requirements are met in Task 5. The only modification I've made is to the dataset path in the task5.yaml.

[gertdekkers]

The features are normally saved in a .cpickle format. The file you're mentioning seems to be a .wav file. I'm not sure why this is the case. Are all other files in that folder also with a .wav extension? How are the dataset/feature paths defined?

[mculp87]

The dataset path is set to ~/Data/DCASE/dataset_task5/. All other parameters are kept to default. I looked in the folder, all of the files in output_task5/features/feature_processing_chain_3b4ae3a13a504b182ef406346a732dc6/ have .cpickle extensions, except for the parameters file. There are two hashes in the parameters file.

_hash: 3b4ae3a13a504b182ef406346a732dc6
method: DCASE2018_Task5_baseline
parameters:
_hash: 113564d184f826f45da520b7199f70ad

[gertdekkers]

Have you tried removing this .wav file in order to do the feature extraction again?