I'm just playing with RNN's and was having trouble getting my data into the right format for my model. I have the following dataframe:
Apple Pears Oranges ID
0 1.00 2.09 4.11 0
1 1.38 1.73 5.13 1
2 1.68 2.28 6.91 2
3 1.50 2.69 8.93 3
4 1.35 2.63 12.25 4
5 1.52 3.09 12.20 5
6 1.63 3.63 13.68 6
7 2.01 4.92 16.21 7
8 2.52 4.01 18.79 8
9 3.10 5.49 24.05 9
ID is a order/timesteps for my data.
I ran this command to try to load it into a timeseries dataset:
Dataset = keras.preprocessing.timeseries_dataset_from_array(priceHistorydf, basketHistorydf, sequence_length=10)
But when I try to train a model on this, I get the following error:
from tensorflow import keras
import numpy as np
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import *
X_train = priceHistorydf
y_train = basketHistorydf
model = Sequential()
model.add(TimeDistributed(Dense(10), input_shape=(X_train.shape[1:])))
model.add(Bidirectional(LSTM(8)))
model.add(Dense(8, activation='tanh'))
model.add(Dense(8, activation='tanh'))
model.add(Dense(y_train.shape[-1], activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer="adam")
# history = model.fit(X_train, y_train, epochs=2, batch_size=8)
history = model.fit(Dataset, epochs=2, batch_size=8)
Error:
ValueError: `TimeDistributed` Layer should be passed an `input_shape ` with at least 3 dimensions, received: [None, 4]
I'm just guessing but I realize I did not explicitly let the model know that ID is the timestep; but I'm unsure how to pass it to the model with my dataframe.
Any suggestions?
The main problem is that you are setting the input_shape argument incorrectly (i.e. X_train is the original data, and not the generated timeseries; hence X_train.shape[1:] is not correct as the input shape). Since you have used sequence_lenght=10 and each timestep has 3 features, therefore we should have input_shape=(10,3) (of course, assuming you would remove the ID column from the data first, because that's not a feature per say).
As a side note: the Dense(...) and TimeDistributed(Dense(...)) are exactly the same, because the Dense layer is applied on the last axis by default. See this answer for more information and explanation.