ValueError: Graph disconnected: cannot obtain value for tensor Keras

TensorFlow General Discussion
, , ,
1

image
image1601×1376 155 KB

The picture depicts my encoder flow. Also, when i run my decoder, i keep crashing with the same error -
ValueError: Graph disconnected: cannot obtain value for tensor KerasTensor(type_spec=TensorSpec(shape=(None, 4096), dtype=tf.int32, name=‘input_word_ids’), name=‘input_word_ids’, description=“created by layer ‘input_word_ids’”) at layer “tf_longformer_model_2”. The following previous layers were accessed without issue: . I have checked the encoder flow and any disconnections near the layer input_word_ids and tf_longformer_model. However not able to do the same for my decoder_model as it keeps getting stuck with the same error. The decoder structure is - `from tensorflow.keras.layers import TimeDistributed, Lambda, Add

# Decoder

decoder_dim = 256

# Decoder
decoder_inputs = Input(shape=(None,), dtype=tf.int32, name='decoder_inputs')
coverage_input = Input(shape=(max_length,), dtype=tf.float32, name="coverage_input")
initial_h = Input(shape=(256,), name='initial_h')
initial_c = Input(shape=(256,), name='initial_c')

# Embedding layer for the decoder
decoder_embedding = Embedding(input_dim=vocab_size, output_dim=decoder_dim, name='decoder_embedding')
decoder_embedded = decoder_embedding(decoder_inputs)

# Decoder LSTM
decoder_lstm = LSTM(decoder_dim, return_sequences=True, return_state=True, name='decoder_lstm')
decoder_lstm_out, _, _ = decoder_lstm(decoder_embedded, initial_state=[state_h, state_c])

# Attention Layer
attention_layer = Attention(use_scale=True, name='attention_layer')
attention_out, attention_weights = attention_layer([decoder_lstm_out, encoder_dim_reduction], return_attention_scores=True)

# Coverage mechanism: update the coverage vector
coverage_update_layer = Add(name="coverage_update")
updated_coverage = coverage_update_layer([coverage_input, Lambda(lambda x: tf.reduce_sum(x, axis=-1))(attention_weights)])

# Reducing the dimensionality of the coverage to match the attention output
coverage_projection = TimeDistributed(Dense(decoder_dim), name="coverage_projection")(updated_coverage[:, tf.newaxis])

# Combine LSTM output, attention output, and coverage vector
decoder_concat_input = Concatenate(axis=-1, name="decoder_concat_input")([decoder_lstm_out, attention_out, coverage_projection])

# Dense layer to generate probabilities over the vocabulary
output_layer = TimeDistributed(Dense(vocab_size, activation='softmax'), name='output_layer')
decoder_outputs = output_layer(decoder_concat_input)

# Pointer-generator layer for copy mechanism
pointer_generator = Dense(1, activation='sigmoid', name="pointer_generator")(decoder_concat_input)

# Final Decoder Model
decoder_model = Model(
    inputs=[decoder_inputs, initial_h, initial_c, coverage_input],
    outputs=[decoder_outputs, updated_coverage, pointer_generator],
    name='decoder_model'
)

decoder_outputs, updated_coverage, pointer_generator = decoder_model(
    [decoder_inputs, state_h, state_c, coverage_input]
)

# Print the summary to inspect the model architecture
decoder_model.summary()
`

Please look into this and give me any suggestions or inputs I can do to rectify this. Have been stuck at this for a week now.

3

Hi @Srishruthi_Nandigam,

Sorry for the delay in response.
To resolve the Graph disconnected error in your decoder model, ensure that all input layers are correctly connected and that initial_h and initial_c are used consistently and verify that encoder_dim_reduction is properly defined and passed to the attention layer. Inspect the model summary to identify any disconnections between layers and check input matches the expected shape during training.

Thank You.