search

dicarlolab / archconvnets

Architecturally optimized neural networks trained with regularized backpropagation
5 stars 5 forks source link

Feature/extract to db #14

Closed ardila closed 10 years ago

ardila commented 10 years ago

@yamins81 This seems to work ok. Once you merge, I will extract the challenge set.

ardila commented 10 years ago

@yamins81 Just logging here to let you/myslef know where I'm stuck before the end of the day.

The api seems to work, but there is a serious issue: I have no idea how to get the correct indices in order to find the write ids corresponding to the features calculated. I tried self.test_dataprovider.indset as you suggested, but when I try to check the labels that theses indices correspond to against the labels that would be written to disk by the original version of this code there is a mismatch.

You should be able to get the assertion error by simply running the following command:

python extractnet.py --test-range=0-44 --train-range=0 --data-provider=general-cropped --load-query='{"experiment_data.experiment_id": "mytestrun2"}' --feature-layer=pool3_11b --data-path=/export/imgnet_storage_full/ardila/hvm_batches --dp-params='{"meta_attribute": "obj", "preproc": {"normalize": false, "dtype": "float32", "resize_to": [128, 128], "mode": "RGB", "crop":[0,0, 128, 128], "mask": null}, "batch_size": 128, "dataset_name": ["dldata.stimulus_sets.hvm", "HvMWithDiscfade"]}' --feature-path=/export/imgnet_storage_full/ardila/imgnet_features --write-db=1 --write-disk=0

yamins81 commented 10 years ago

That's an untrained model you're using. Use the reference model as mentioned in issue #11

ardila commented 10 years ago

Ok I have tested this by checking performance on HvM and it seems to match what Darren saw. How should we go about merging this?

yamins81 commented 10 years ago

@ardila this looks ok but it needs to be merged with the most recent changes in master. Especially the change to the option names in extractnet, and the fact that extractnet now can handle multiple comma-separated layer names. I'll have a look at the merge now.