Index out of range using input dim 2,when implement crf

[forglin]

I write code in keras trying to build a QA system,And I use a lstm layer to compute a representation for question,and Evidenve lstm for analyzing evidence. At last I trying to merge them together and feed them into another lstm layer to produce the "features" to CRF.

The CRF layer for sequence labeling which indicates whether each word in the evidence is at the beginning (B), inside (I) or outside (O) of the answer. The shape of texts_evi like this [11117,110,64]which is the same to texts_q.

I think I make some mistake about dimention of array.But I can't fix it.

`q_model = Sequential() q_model.add(LSTM(64,activation='relu',return_sequences=True,name='q_lstm',input_shape=(110,64))) q_input= Input(shape=(110,64)) q = q_model(q_input)

evi_model=Sequential() evi_model.add(LSTM(64,activation='relu',name='q_lstm',return_sequences=True,input_shape=(110,64))) evi_input=Input(shape=(110,64)) encode_evi=evi_model(evi_input)

merged = keras.layers.concatenate([q, encode_evi])

lstm=LSTM(64, activation='relu',return_sequences=False)(merged) output = CRF(110, sparse_target=True)(lstm)

model = Model(inputs=[q_input,evi_input],outputs=output) model.compile(loss='categorical_crossentropy', optimizer='adam') model.fit([texts_q,texts_evi],lable, batch_size=32, epochs=10,validation_split=0.2) ` the error is:

ValueError: Index out of range using input dim 2; input has only 2 dims for 'crf_1/strided_slice' (op: 'StridedSlice') with input shapes: [?,110], [3], [3], [3] and with computed input tensors: input[3] = <1 1 1>.

[spytensor]

Did you solve it?

[Lanme]

how to solve it?

[uwaisiqbal]

This is because the crf layer expects the labels in a different shape. Normally your labels would be of shape (num_samples, max_length) but the crf layer expects them in the form (num_samples, max_length, 1) .

An easy fix is to reshape your labels as follows: labels = labels.reshape((labels.shape[0], labels.shape[1], 1))

[blog4me]

This is because the crf layer expects the labels in a different shape. Normally your labels would be of shape (num_samples, max_length) but the crf layer expects them in the form (num_samples, max_length, 1) .

An easy fix is to reshape your labels as follows: labels = labels.reshape((labels.shape[0], labels.shape[1], 1))

Hi, I'm little confused about your answer, I have the same problem just like his, I want to try your method, only don't know where to change. Can you be more specific on that issue? Thants a lot

[blog4me]

Did you solve it? I have the very same problem, If you solve it, plz help me.

[jinkaima]

I have the same problem. And solve it by y_train = np.expand_dims(y_train,2)

[HansyM]

I have similar problem, I cant solve it

[uwaisiqbal]

This is because the crf layer expects the labels in a different shape. Normally your labels would be of shape (num_samples, max_length) but the crf layer expects them in the form (num_samples, max_length, 1) . An easy fix is to reshape your labels as follows: labels = labels.reshape((labels.shape[0], labels.shape[1], 1))

Hi, I'm little confused about your answer, I have the same problem just like his, I want to try your method, only don't know where to change. Can you be more specific on that issue? Thants a lot

So you need to reshape your labels before you call model.fit You need to reshape your y_train data so that it has shape (num_samples, max_length, 1)

[DebanjanaKar]

Hi @uwaisiqbal , the solution you proposed does resolve the error mentioned in this issue. However, I am getting a new error on reshaping the labels as follows :

InvalidArgumentError: Matrix size-incompatible: In[0]: [158,1], In[1]: [4,4]
	 [[{{node loss_1/crf_2_loss/MatMul_1}}]]

Can you or anyone help me in resolving this?

[ethanyuxinzhang]

Hi @uwaisiqbal , the solution you proposed does resolve the error mentioned in this issue. However, I am getting a new error on reshaping the labels as follows :

InvalidArgumentError: Matrix size-incompatible: In[0]: [158,1], In[1]: [4,4]
	 [[{{node loss_1/crf_2_loss/MatMul_1}}]]

Can you or anyone help me in resolving this?

If you still have this problem. Please have a look at your previous layer output shape and then use Reshape layer to reshape the output. Then the problem should be solved. If NOT, try to use sparse_tartget=true. Hope this can help you.

[angyee]

ValueError: Index out of range using input dim 2; input has only 2 dims for '{{node loss/dense_loss/strided_slice_2}} = StridedSlice[Index=DT_INT32, T=DT_FLOAT, begin_mask=3, ellipsis_mask=0, end_mask=3, new_axis_mask=0, shrink_axis_mask=0](dense_target, loss/dense_loss/strided_slice_2/stack, loss/dense_loss/strided_slice_2/stack_1, loss/dense_loss/strided_slice_2/stack_2)' with input shapes: [?,?], [3], [3], [3] and with computed input tensors: input[3] = <1 1 1>

while running this repo https://github.com/suyxing/ssd_resnet model = SSD_resnet.resnet_34(300, 300, 3, classes=NUM_CLASSES)

[paktiger]

Assalam o aliekum bro @uwaisiqbal I have the same problem as mentioned above and following is my code:

input = Input(shape=(75,)) emd=Embedding(input_dim=n_words + 1,output_dim=20, input_length=75,mask_zero=False)(input) bilstm = Bidirectional(LSTM(units=75, return_sequences=True, recurrent_dropout=0.1))(emd)

input2 = Input(shape=(75,)) flat_bilstm = Flatten()(bilstm) layersConcate = concatenate([flat_bilstm,input2])

model_crf =Dense(units=75, activation="relu")(layersConcate)

crf=CRF(8, sparse_target=True) out_ = crf(model_crf) ----------------> when I pass model_crf it generates the following error:

ValueError: Index out of range using input dim 2; input has only 2 dims for 'crf_1_3/strided_slice' (op: 'StridedSlice') with input shapes: [?,8], [3], [3], [3] and with computed input tensors: input[3] = <1 1 1>.

please help me I am kind of stuck in this problem

[paktiger]

Assalam o aliekum bro @uwaisiqbal I have the same problem as mentioned above and following is my code:

input = Input(shape=(75,)) emd=Embedding(input_dim=n_words + 1,output_dim=20, input_length=75,mask_zero=False)(input) bilstm = Bidirectional(LSTM(units=75, return_sequences=True, recurrent_dropout=0.1))(emd)

input2 = Input(shape=(75,)) flat_bilstm = Flatten()(bilstm) layersConcate = concatenate([flat_bilstm,input2])

model_crf =Dense(units=75, activation="relu")(layersConcate)

crf=CRF(8, sparse_target=True) out_ = crf(model_crf) ----------------> when I pass model_crf it generates the following error:

ValueError: Index out of range using input dim 2; input has only 2 dims for 'crf_1_3/strided_slice' (op: 'StridedSlice') with input shapes: [?,8], [3], [3], [3] and with computed input tensors: input[3] = <1 1 1>.

please help me I am kind of stuck in this problem

This problem is solved since the CRF method requires three-dimensional arrays but I was passing 2D only.