search

eembc / energyrunner

The EEMBC EnergyRunner application framework for the MLPerf Tiny benchmark.
14 stars 5 forks source link

KWS FP32 model is not FP32 #13

Closed colbybanbury closed 3 years ago

colbybanbury commented 3 years ago

The FP32 KWS model in base_models is actually a copy of the Int8 model.

Both models have the same hash (618aeb155673dae92fa6a7f26608add5) and if you open the fp32 model in Netron you can see the weights are fp32.

petertorelli commented 3 years ago

Thanks for catching this. Since both organizations mandate starting from fp32 TFLite models, I have removed all int8 to avoid confusion.

petertorelli commented 3 years ago

Re-opening because the MLC tiny model is int8. Colby, please add the real fp32 model to the tiny repo and i'll copy it here.

honson1 commented 3 years ago

@colbybanbury Touching base on this before Thursday to see when the kws fp32 model will be posted!

petertorelli commented 3 years ago

@colbybanbury Ping?

colbybanbury commented 3 years ago

The float model has been uploaded to the TinyMLPerf repo: link

No MD5 yet but I can add that.

@honson1 is this considered resolved then?

petertorelli commented 3 years ago

@colbybanbury Thanks! I just made a copy.