Binary classification always predicting .5, .5 about keras-tcn HOT 5 CLOSED

@jmrichardson Hum it means that the network does not learn anything if the classes are well-balanced (50% chance of having 0, and 50% of having 1). Why don't you try another simple model? Are your features spatially correlated like an image? Because TCN is based on convolutions.

from keras-tcn.

Sorry, I assumed that we could use keras-tcn for multivariate tabular data since I saw an example in your repo for a univariate data set. I am learning more about CNNs and see that the filters convolve around image data to detect things like edges for images. Since my data set features are temporal (not IID) but not images I don't think I can use keras-tcn correct?

from keras-tcn.

@jmrichardson no you can here. If your data has shape (batch-size, timestep, input-dim), you can use a TCN. For example if you have 64 time series with 1000 time steps, you can use TCN. Your input will be (batch-size, 1000, 64). TCN uses 1-D convolutions.

from keras-tcn.

I think I have the right data shape... I have 83 time series with 31K rows. I used the following code to create a rolling window of 20 sequences/timesteps similar to your milk task example:

lookback_window = 20

# Get training data
X_train = np.array(X_train_split)
y_train = np.array(y_train_split)
x, y = [], []
for i in range(lookback_window, len(X_train)):
    x.append(X_train[i - lookback_window:i])
    y.append(y_train[i])
X_train = np.array(x)
y_train = np.array(y)

# Get validation data
X_val = np.array(X_val_split)
y_val = np.array(y_val_split)
x, y = [], []
for i in range(lookback_window, len(X_val)):
    x.append(X_val[i - lookback_window:i])
    y.append(y_val[i])
X_val = np.array(x)
y_val = np.array(y)

Here is the shape X_train shape:

print(X_train.shape)
print(y_train.shape)
(31066, 20, 83)
(31066,)

Here is the model code:

model = compiled_tcn(return_sequences=False,
                     num_feat=X_train.shape[2],
                     num_classes=2,
                     nb_filters=24,
                     kernel_size=8,
                     dilations=[2 ** i for i in range(9)],
                     nb_stacks=1,
                     max_len=X_train.shape[1],
                     use_skip_connections=True,
                     )

model.fit(X_train, y_train, validation_data=(X_val, y_val), epochs=1, batch_size=256)

from keras-tcn.

Sorry I don't have much time to look at code snippets. I don't see any errors here. Don't hesitate to post your question on a machine learning forum. I'll close this issue as I don't see any bugs here. Thanks!

from keras-tcn.