Ultra-low power and high-performance AI processor GAP8 › Forums › GAP8 developers’ group › SDK (Software Development Kit) › tf2gap8 Regression Use Case
- Tobias DentlerParticipantOctober 2, 2019 at 1:56 pmPost count: 3
has anyone ever tried to build an example doing a regression instead of a classification task? I’m currently experimenting with adaptions of the MNIST example, but haven’t gotten a proper output.
Most likely the GTT transforms my output layers in way I haven’t fully understand yet.
The last layers of my architecture look like follows, with layer definitions coming from the mnist cnn.py file:
conv3, w3, b3 = cnn.new_conv_layer(input = conv2, num_channel_input = 16 ,filter_size=(1,2), num_filter=16, pooling=True, Relu = True)
layer_flat, num_features = cnn.flatten_layer(conv3)
layer_fc1, w_fcl, b_fcl = cnn.new_connected_layer(input=layer_flat,num_input=num_features, num_output=1, ReLu=False)
y_pred = layer_fc1
y_pred_cls = tf.argmax(y_pred, dimension = 1, name=”y_output”)
The training/testing process looks fine. However after transforming the net and flashing it on the hardware, it doesn’t perform as expected. The outputs seem to be rather arbitrary laying in the short int range instead of the expected range [0, 10].
Are there hints on how to define the last layers before the transformation if I want to perform a regression task? Happy to share more informarion if needed.
- You must be logged in to reply to this topic.