converter breaks on basic lstm

[nevakrien]

so i am on ubuntu 20 and i have this in my html file i am opening on firefox with the fillowing link https://cdn.jsdelivr.net/npm/@tensorflow/tfjs

i then run model = await tf.loadLayersModel('js_model/model.json'); which causes the error
"Initialization error: An InputLayer should be passed either abatchInputShapeor aninputShape."

the model i am runing is a stupidly simple lstm all native tensorflow. runing conversion from both h5 and keras seems to work at first but then it gives that crush.

the full model code

vocab_size = len(vectorizer)  # Number of unique words in your vocabulary

embedding_dim = 340

seq_length = 340  # Max sequence length

num_classes = 1



inputs = Input(shape=(seq_length,))

embedding = Embedding(input_dim=vocab_size, output_dim=embedding_dim, input_length=seq_length)(inputs)

cnn = Conv1D(filters=128, kernel_size=5, activation='relu', padding='same')(embedding)

cnn = BatchNormalization()(cnn)

cnn = MaxPooling1D(pool_size=4)(cnn)

cnn = Conv1D(filters=256, kernel_size=3, activation='relu', padding='same')(cnn)

cnn = BatchNormalization()(cnn)

cnn = MaxPooling1D(pool_size=2)(cnn)

lstm = LSTM(128, return_sequences=True, dropout=0.3, recurrent_dropout=0.3)(cnn)

lstm = LSTM(64, return_sequences=False, dropout=0.3, recurrent_dropout=0.3)(lstm)

dense = Dense(256, activation='relu')(lstm)

dense = Dropout(0.5)(dense)

dense = Dense(128, activation='relu')(dense)

dense = Dropout(0.5)(dense)

dense = Dense(64, activation='relu')(dense)

outputs = Dense(num_classes, activation='sigmoid')(dense)



model = Model(inputs=inputs, outputs=outputs)

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

model.summary()

[shmishra99]

Hi @nevakrien ,

This is a known issue with TFJS currently. Tfjs-converter does not support models created by Keras 3. When you try to load this model using model = await tf.loadLayersModel('js_model/model.json'), it will attempt to infer the input shape which is not present in the model.json file.

You can install Keras 2 (TensorFlow 2.15.0) to create your model and then use the tfjs-converter to convert the TensorFlow model to a TFJS model. Or alternatively, you can load your current model as a Graph Model instead of a Layer Model using const model = await tf.loadGraphModel(modelUrl);.

Let me know if it helps. Thank You!!