Convert Model to .tflite

[barnabasbartha]

Your implementation looks promising.
I'm trying converting the model to .tflite to run it on Android. Could you please help me which tensor is exactly the input and the output?
Thanks.

[stardut]

You can find the anwser in the model.py and predict.py.
Like the input,

self.inputs = tf.placeholder(tf.float32, [None, None, self.input_size])

means [batch_size, img_cols, img_rows].

The batch_size and img_cols can be any positive integer and the img_rows is 28 in my model. But you can train model and set any positive integer as you like.

About the output, you can find it in predict.py,

decode = sess.run(model.decoded, feed_dict=feed)
pre = data.decode_sparse_tensor(decode[0])

[barnabasbartha]

Thanks your reply.
I've already found those, but if I set the model.decoded[0] as the output tensor to TFLiteConverter, it says:

AttributeError: 'SparseTensor' object has no attribute 'name'

Giving only a Placeholder as an input and output with size of [1, 224, 28] and [1, 28] respectively (I'm not sure about the latter one) it gets converted to .tflite but crashes the Android app with a SIGSEGV.
So not sure how can be used such models on Android :/

[barnabasbartha]

Still no luck. Any idea?

[stardut]

Sorry, I had a busy week.
You could find the function decode_sparse_tensor in data.py , the decoded_indexes as the output and the size is [1, 8] while the input size is [1, 224, 28]. Because the word width is 28 in my model.

decoded_indexes = list()
current_i = 0
current_seq = []
for offset, i_and_index in enumerate(sparse_tensor[0]):
    i = i_and_index[0]
    if i != current_i:
         decoded_indexes.append(current_seq)
         current_i = i
         current_seq = list()
    current_seq.append(offset)
decoded_indexes.append(current_seq)