Holberton Hackathon

[sidneyriffic]

I left in a few models of my models backup folders that are a bit farther along in epochs that don't quite have as much loss if you wanted to look at them. The one I put into "weights" was the lowest for my most recent model, and is the same as the 2-24 folder.

[GiorgioRegni]

Thank you Sidney!

[sidneyriffic]

The model just started spitting out a bunch of stuff with half the loss just now, heh. A quick push for that. It's in weights/the 5-13 folder.

[GiorgioRegni]

The juicy parts :) : `units= CONFIG['lstm_hidden_size'] dropout = .1

design network

inputs = Input(shape=(48, 4)) firstconv = Conv1D(30, 3)(inputs) firstconv = LSTM(30)(firstconv) firstconv = Dropout(dropout)(firstconv) firstrnn = LSTM(30)(inputs) firstrnn = Dropout(dropout)(firstrnn) firstcat = concatenate([firstconv, firstrnn]) results = Dense(60)(firstcat) results = Dropout(dropout)(results) results = Dense(4)(results) results = keras.layers.Reshape((1, 4))(results)

innerputs = concatenate([inputs, results], axis=1) for i in range(1, 12):

innerconv = Conv1D(30, 3)(innerputs)
innerconv = LSTM(30)(innerconv)
innerconv = Dropout(dropout)(innerconv)
innerrnn = LSTM(30, input_shape=(48 + i, 4))(innerputs)
innerrnn = Dropout(dropout)(innerrnn)
innercat = concatenate([innerconv, innerrnn])
innerres = Dense(60)(innercat)
innerres = Dropout(dropout)(innerres)
innerres = Dense(4)(innerres)
innerres = keras.layers.Reshape((1, 4))(innerres)
results = concatenate([results, innerres], axis = 1)
innerputs = concatenate([innerputs, innerres], axis=1)`

[sidneyriffic]

Yeah, that is the best part. It was pretty cool to think up. But I at least still need to make sure it's doing something like what I want it to or at least think it might do! Which it seems to be, fortunately.