apivovarov edited a comment on issue #4142: [QNN][TFLite] Parsing QNN Add op. Adding MobilenetV2. URL: https://github.com/apache/incubator-tvm/pull/4142#issuecomment-561413217 @anijain2305 I noticed that `test_forward_qnn_mobilenet_v2_net` works only if `random.seed(0)` is set to 0. if I set seed to say 20 - the test fails. I also tried to replace deprecated `numpy.random.random_integers` with `numpy.random.randint` the test failed as well So , looks like the test only works for particular input data started with 47, 117, 67 ``` >>> np.random.seed(0) >>> data = np.random.random_integers(low=0, high=128, size=(1, 224, 224, 3)).astype('uint8') >>> data array([[[[ 47, 117, 67], [103, 9, 21], [ 36, 87, 70], ..., [ 43, 110, 118], [ 22, 109, 11], [119, 26, 48]], ... ```` ``` # seed(1) Mismatch: 33.3% Max absolute difference: 337 Max relative difference: 0.36951754 x: array([886, 540, 575]) y: array([886, 540, 912]) # seed(2) Mismatch: 100% Max absolute difference: 353 Max relative difference: 0.6537037 x: array([845, 886, 893]) y: array([886, 845, 540]) ``` The same issue exist for `test_forward_qnn_inception_v1_net` We should probably use real dog/cat image data for top3 labels comparison testing
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
