[Python] Why does my LSTM model raise max_seq_length <= 0 when predicting with new input of...

I'm training a multi-branch LSTM model on Kaggle using 2x T4 GPUs, with multiple input sequences of different lengths. My target is the "close" column.

1. Normalization and Data:


I normalize my data with a custom Normalizer class:

normalizer = Normalizer(start_data[cols])
data = normalizer.normalize()
values = data[cols].values

2. Model


times = [5, 15, 30, 90] # Sequence lengths

def build_lstm_model(times, lstm_units=64, dense_units=64):
inputs = []
branches = []

for t in times:
input_t = Input(shape=(t, 4))
x = LSTM(lstm_units, return_sequences=True)(input_t)
x = Dropout(0.2)(x)
x = LSTM(lstm_units)(x)
branches.append(x)
inputs.append(input_t)

merged = concatenate(branches)
dense = Dense(128, activation="relu")(merged)
dense = Dense(dense_units, activation="relu")(dense)

residual = Dense(dense_units)(merged)
added = Add()([dense, residual])

output = Dense(1)(added)

return Model(inputs=inputs, outputs=output)

3. Data Preparation


I prepare lists with the differents durations :

min_len = max(times)
x_inputs = [[] for _ in times]
y = []

for i in range(min_len, len(values)):
for idx, t in enumerate(times):
x_inputs[idx].append(values[i - t:i])
y.append(values[cols.index("close")])

x_inputs = [np.array(x) for x in x_inputs] # e.g., shape (911, 90, 4)
y = np.array(y)

4. Training (work)


strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
model = build_lstm_model(times)
model.compile(optimizer="adam", loss="mse", metrics=["mae"])

model.fit(x_inputs, y, epochs=2, batch_size=256)

5. Predictions (error)


I don't have problem when I use train data. For the last prediction, I do :

x_inputs = [values[-t:].reshape(1, t, 4) for t in times]
prediction = model.predict(x_inputs)


But this throws: InvalidArgumentError: Graph execution error: ... max_seq_length <= 0 ...

while all inputs have shape (1, t, 4).

Why is model.predict(x_inputs) throwing max_seq_length <= 0? The model works perfectly for training and predicting on full datasets. The error happens only during single sample prediction using the last t values.

Thanks for your help

I have check manually the shape, rebuild the NoteBook, retrain se model, but the error persists.

Continue reading...