Well, I'm trying to replicate a model from this paper on Pubmed 200k Randomized controlled clinical trials dataset which you can find in this paper.
the dataset with 5 unique tokens for the output is something like this:
well, I've made some progress I tried my best to replicate the paper but in the end, I'm getting a value error:
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-55-d583261b0496> in <module>()
23
24 # 4. create an output layer
---> 25 dropout_layer = tf.keras.layers.Dropout(0.5,name='dropout_layer')(concat_model)
26 dense_layer = tf.keras.layers.Dense(200,name='dense_layer')(dropout_layer)
27 dropout_layer = tf.keras.layers.Dropout(0.5,name='dropout_layer_2')(dense_layer)
1 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/constant_op.py in convert_to_eager_tensor(value, ctx, dtype)
100 dtype = dtypes.as_dtype(dtype).as_datatype_enum
101 ctx.ensure_initialized()
--> 102 return ops.EagerTensor(value, ctx.device_name, dtype)
103
104
ValueError: Exception encountered when calling layer "dropout_layer" (type Dropout).
Attempt to convert a value (<keras.layers.merge.Concatenate object at 0x7f51a0b5f850>) with an unsupported type (<class 'keras.layers.merge.Concatenate'>) to a Tensor.
Call arguments received:
• inputs=<keras.layers.merge.Concatenate object at 0x7f51a0b5f850>
• training=False
for the code, I tried my best to comment on as many things as possible.
# 1. word model:
inputs = tf.keras.layers.Input(
shape=(1,), dtype=tf.string, name="text_input"
) # takes a list of input
text_vectorization_layer = text_vectorizer(inputs) # word level vectorizer
text_embedding_layer = embedding(text_vectorization_layer) # word level embedding
# flatten_layer = tf.keras.layers.Flatten()(text_embedding_layer)
outputs = tf.keras.layers.Dense(300, activation="relu", name="text_output")(
text_embedding_layer
) # output of 300 from fig 1. https://arxiv.org/pdf/1612.05251.pdf
token_model = tf.keras.Model(inputs, outputs) # token model
# 2. char model:
inputs = tf.keras.layers.Input(shape=(1,), dtype=tf.string, name="char_input")
char_vectorization_layer = char_vectorizer(inputs) # character level vectorization
char_embedding_layer = char_embedding(char_vectorization_layer) # char level embedding
outputs = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(25), name="bi_directional_lstm"
)(
char_embedding_layer
) # bi directional lstm as output
char_model = tf.keras.Model(inputs, outputs)
# 3. concatinating both the model
concat_model = tf.keras.layers.Concatenate([token_model, char_model])
# 4. create output layer
dropout_layer = tf.keras.layers.Dropout(0.5, name="dropout_layer")(concat_model)
dense_layer = tf.keras.layers.Dense(200, name="dense_layer")(dropout_layer)
dropout_layer = tf.keras.layers.Dropout(0.5, name="dropout_layer_2")(dense_layer)
output_layer = tf.keras.layer.Dense(5, activation="softmax", name="final_output_layer")(
dropout_layer
)
I'm not getting what will be the output from the concatenate layer that it's causing an error for the dropout layer
fixed code thanks to @Djinn :
# 1. word model:
inputs = tf.keras.layers.Input(
shape=(1,), dtype=tf.string, name="text_input"
) # takes a list of input
text_vectorization_layer = text_vectorizer(inputs) # word level vectorizer
text_embedding_layer = embedding(text_vectorization_layer) # word level embedding
flatten_layer = tf.keras.layers.Flatten()(text_embedding_layer)
outputs = tf.keras.layers.Dense(300, activation="relu", name="text_output")(
flatten_layer
) # output of 300 from fig 1. https://arxiv.org/pdf/1612.05251.pdf
token_model = tf.keras.Model(inputs, outputs) # token model
# 2. char model:
inputs = tf.keras.layers.Input(shape=(1,), dtype=tf.string, name="char_input")
char_vectorization_layer = char_vectorizer(inputs) # character level vectorization
char_embedding_layer = char_embedding(char_vectorization_layer) # char level embedding
outputs = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(25), name="bi_directional_lstm"
)(
char_embedding_layer
) # bi directional lstm as output
char_model = tf.keras.Model(inputs, outputs)
# 3. concatinating both the model
concat_model = tf.keras.layers.Concatenate()([token_model.output, char_model.output])
# 4. create output layer
dropout_layer = tf.keras.layers.Dropout(0.5, name="dropout_layer")(concat_model)
dense_layer = tf.keras.layers.Dense(200, name="dense_layer")(dropout_layer)
dropout_layer = tf.keras.layers.Dropout(0.5, name="dropout_layer_2")(dense_layer)
output_layer = tf.keras.layers.Dense(
5, activation="softmax", name="final_output_layer"
)(dropout_layer)
# 5.final model
model_2 = tf.keras.Model(
inputs=[token_model.input, char_model.input],
outputs=output_layer,
name="token_char_model",
)
