0
$\begingroup$

I have the following LSTM model and I can't make inference with it:

print("Define LSTM model")

rnnmodel=Sequential()
rnnmodel.add(embedding_layer)

rnnmodel.add(LSTM(128, dropout=0.2, recurrent_dropout=0.2))
rnnmodel.add(Dense(2, activation="sigmoid"))

rnnmodel.compile(loss="binary_crossentropy",
                 optimizer="adam",
                 metrics=["accuracy"])

rnnmodel.fit(X_train, y_train,
             batch_size=256,
             epochs=1,
             validation_data=(x_val, y_val))

score, acc=rnnmodel.evaluate(test_data, test_labels, batch_size=128)
print(f"Test accuracy with RNN: {acc}")

(epoch is 1 to test) I want to make an inference with the text, let's say

text=["the product was horrible"]

I check the documentation of tf.keras.Sequential and it states I should use the predict function and the input should be "A Numpy array (or array-like), or a list of arrays (in case the model has multiple inputs)."

So what I did is:

text=["the product was horrible"]
inference_sequence=tokenizer.texts_to_sequences(text)
inference_data=pad_sequences(inference_sequence, maxlen=MAX_SEQUENCE_LENGTH)

predictions=rnnmodel.predict(inference_data)

print(predictions)

and it gives me the result [[0.63219154 0.33410403]]

However I've given only one sentence. Why it gives me two results? I checked the sigmoid documentation from here and for an confirmed it should return only one result. So what's the problem here?

I also tried other approaches to make inference like mentioned https://stackoverflow.com/questions/61443543/how-to-make-prediction-on-keras-text-classification

So I did:

text=["the product was horrible"]
rnnmodel.predict(text)

and it gives me the warning: WARNING:tensorflow:Model was constructed with shape (None, 1000) for input Tensor("embedding_input:0", shape=(None, 1000), dtype=float32), but it was called on an input with incompatible shape (None, 1). and stuck forever.

What should I do I just can't make an inference.

$\endgroup$
1
$\begingroup$

I checked the sigmoid documentation from here and for a confirmed it should return only one result. So what's the problem here?

You have used 2 Neurons at the output layer. So, each is responsible for one output.
Either, change the neuron count to 1 and y_true to the 1-D array.
Or change the activation function to Softmax.

text=["the product was horrible"]
rnnmodel.predict(text)

This will not work.
The embedding layer expects Label encoded data. So the first approach is correct.

$\endgroup$
1
  • $\begingroup$ yep you're correct. Switched to softmax working now. Thank you. $\endgroup$
    – J.Smith
    Aug 26 at 8:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.