How to extract and print a weight matrix from a ckpt file?

[Jui-Yi Tsai]

I tried to execute the attached code "word2vec_optimized.py," which is edited from the original one cloned from the github of TensorFlow.

The lines 437 to 458 in the attached code are shown as follows:

    f_voc = open('vocab.txt', 'r')

    lines = f_voc.readlines()

    rev_dict = []

    for x in lines:

    rev_dict.append(x.strip())

    f_voc.close()

    final_embeddings = np.zeros((71291, 200)) 

    restored_model = model.saver.restore(session,'model.ckpt-2265170')

    #print_tensors_in_checkpoint_file("model.ckpt-2265157", "w_out")

    fout = open('output_word2vec_test.txt', 'w')  

    for i in range(len(final_embeddings)):

    for j in range(len(final_embeddings[i])):

    if j == 0:

    fout.write(str(rev_dict[i]) + ' ' + str(final_embeddings[i][j]) + ' ')

    elif j == (len(final_embeddings[i])-1):

fout.write(str(final_embeddings[i][j]) + '\n')

    else:

fout.write(str(final_embeddings[i][j]) + ' ')

    fout.close()

    print(restored_model)

With the above lines, I tried to print the words and their corresponding vectors like the following format, which looks like an array:

apple 0.123456 -1.234567 ... 0.987654
banana 0.123456 -1.234567 ... 0.987654
cat 0.123456 -1.234567 ... 0.987654

But I can't manage to extract a matrix in a ckpt file after checking it with "inspect_checkpoint.py".....How should I do? Is there any suggestion? Thanks.

[Yuxin Wu]

You can use tf.train.NewCheckpointReader, which provides method to read the values directly.