Visualizing Saliency Map for Sentence Classification

[ndrmahmoudi]

Hi @raghakot ,

I just want to try the the module on a model for sentence classification. Actually, I would like to have plots as in this paper. Can you please let me know how to do this?

Regards, Nader

[raghakot]

It should work out of the box. The input, in this case, is not an image. I will try to whip up a notebook showing how to do this. Do you have a pretrained network on some text classification corpus?

[ndrmahmoudi]

Hi @raghakot,

Many thanks for this. I am quite new in this field. So, excuse me for asking such a time-consuming thing.

Yes I do. Do you want me to share them with you?

Regards, Nader

[raghakot]

Yep. Share them along with the dataset etc. I wanted to create examples in different domains anyways.

[ndrmahmoudi]

Alright, I sent the files to your email. Please note that I just trained the model for 10 epochs on the data available here, so it is completely simple model. Cheers, Nader

[raghakot]

I will most likely get to it this weekend. Putting in work on office days is tricky.

On Jul 12, 2017, 11:16 PM -0700, ndrmahmoudi notifications@github.com, wrote:

Alright, I sent the files to your email. Please note that I just trained the model for 10 epochs on the data available herehttps://github.com/shagunsodhani/CNN-Sentence-Classifier, so it is completely simple model. Cheers, Nader

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/raghakot/keras-vis/issues/57#issuecomment-314982168, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AO6vTEORnG-q_TiSVuE_Wjm5FHzCq9kgks5sNbYogaJpZM4OVdzm.

[FloatFlow]

Did anything come of this issue? I would be interested in seeing the notebook for this application. My situation is similar, except it is for DNA classification, and the paper I am trying to emulate is: https://arxiv.org/pdf/1608.03644.pdf.

[raghakot]

I did not get a chance to work on this yet. However, I suggest you take a look at this issue: https://github.com/raghakot/keras-vis/issues/63 which has some working code. I will try to make a clean notebook example as soon as I get a chance.

[ndrmahmoudi]

Hi all @pexmar,

I was working on it for like last 10 days and finally I have got a code adapted from here which gives saliency map for the sentences (see below).

def plot_saliency(loaded_model, pure_txt, text_class, pred_labels, text_sequence):
    text_class = 'Positive' if text_class==1 else 'Negative'
    pred_labels = 'Positive' if pred_labels==1 else 'Negative'

    input_tensors = [loaded_model.input, K.learning_phase()]
    model_input = loaded_model.layers[2].input # the input for convolution layer
    model_output = loaded_model.output[0][1]
    gradients = loaded_model.optimizer.get_gradients(model_output,model_input)
    compute_gradients = K.function(inputs=input_tensors, outputs=gradients)
    matrix = compute_gradients([text_sequence.reshape(1,30), text_class])[0][0]
    matrix = matrix[:len(pure_txt),:]

    matrix_magnify=np.zeros((matrix.shape[0]*10,matrix.shape[1]))
    for i in range(matrix.shape[0]):
        for j in range(10):
            matrix_magnify[i*10+j,:]=matrix[i,:]

    fig = plt.figure()
    ax = fig.add_subplot(111)
    plt.imshow(normalize_array(np.absolute(matrix_magnify)), interpolation='nearest', cmap=plt.cm.Blues)
    plt.yticks(np.arange(5, matrix.shape[0]*10, 10), pure_txt, weight='bold',fontsize=24)
    plt.xticks(np.arange(0, matrix.shape[1], 50), weight='bold',fontsize=24)
    plt.title('True Label: "{}" Predicted Label: "{}"'.format(text_class,pred_labels), weight='bold',fontsize=24)
    plt.colorbar()
    plt.show()

Please kindly correct me if I have done anything wrong with regard the code (I am not really expert in programming). Moreover, I have a question. When I re-run this code, it gives different results but with roughly same distribution of gradients. I mean, it shows higher gradients for important words and lower for redundant words but with other values and it changes every time I re-run the code for same sentence. Is there any problem with this?

Regards, Nader

[ssierral]

Hi,

When I re-run this code, it gives different results but with roughly same distribution of gradients. I mean, it shows higher gradients for important words and lower for redundant words but with other values and it changes every time I re-run the code for same sentence. Is there any problem with this?

This could be due to the K.learning_phase(), I see here that you are providing the text_class as the learning_phase and if you have some layers such as Dropout or BatchNormalization, it's better to set it to 0 if you don't want them to work in the testing phase. BTW your code was really helpful for an implementation I was doing using Text Multi-Input models.

Hope this helps

[ndrmahmoudi]

Thanks @ssierral, I don't use text_class as learning phase, do I?

I wanted to K.set_learning_phase(0). But it gives TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a int into a Tensor. Probably, it is not possible to change the learning phase in my model. Can you please let me know how to solve this issue?

Regards, Nader

[ssierral]

I think you set the learning phase passing a 0 to the backend function:

matrix = compute_gradients([text_sequence.reshape(1,30), 0])[0][0]

[Chiranjeev93]

hi @raghakot Can you please share the code that you shared with @ndrmahmoudi for Visualizing Saliency Map for Sentence Classification.

Thanks, Chiranjeev

[nnaliu]

Hi @raghakot, I'm also working on something similar in Pytorch right now and would love to see the code you shared with @ndrmahmoudi for mapping saliency.

[ndrmahmoudi]

Hi @nnaliu, I could fix my own problem with the function in one of my comments above. Feel free to use it and ask any question if you have.