Tensorflow JS error with model imported from Keras (Based on the provided shape, [2,256], the tensor should have 512 values but has 43)

[Timon Lukas]

Hey guys!

I have a big problem with Tensorflow JS right now. I trained an LSTM based model with Keras and I'm now trying to import it into TFJS. To do so, I converted it using tensorflowjs_converter, and it worked fine - no errors appeared. But when I try to open it in the browser, I get this error:

Based on the provided shape, [2,256], the tensor should have 512 values but has 43

Does anyone have a clue why this happens?

The model definition:

model = Sequential()

model.add(LSTM(
    32,
    batch_input_shape=(30, 5, 3),
    return_sequences=True,
    stateful=True,
    activation='tanh',
))
model.add(Dropout(0.25))

model.add(LSTM(
    32,
    return_sequences=True,
    stateful=True,
    activation='tanh',
))
model.add(Dropout(0.25))

model.add(LSTM(
    32,
    return_sequences=False,
    stateful=True,
    activation='tanh',
))
model.add(Dropout(0.25))

model.add(Dense(3, activation='tanh', kernel_initializer='lecun_uniform'))

model.compile(loss='mse', optimizer=Adam())

This is the Keras model: https://drive.google.com/open?id=1ts8YsbRfYo9yaHYUraEcrKPgke5xkwwo

model.json: https://drive.google.com/open?id=1XXIlU6PYDpUeqc6oUxOxZAgXJY3MP9G-

Weights file: https://drive.google.com/open?id=1Nlb4_yh8lOZKiWnQyyxMI3BlTLQFhAM6

I am pretty sure that this is a problem with TFJS itself.

I am using the latest version of all involved libraries.

Thank you!

[Nikhil Thorat]

Does this error happen when you try to make a prediction? Can you paste the javascript code you're using to load and execute the model?

--
You received this message because you are subscribed to the Google Groups "TensorFlow.js Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfjs+uns...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tfjs/.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfjs/084153fd-8bc1-4267-99b8-dd836cace89a%40tensorflow.org.

[Daniel Smilkov]

Also if you can send us the stack trace, that would be great!

Daniel

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfjs/CACni9cdVwe%2BzsiAA7YierLkwU7nphU0o9-sT7gaEj2tKGfd53A%40mail.gmail.com.

[Timon Lukas]

Am Freitag, 31. August 2018 17:52:01 UTC+2 schrieb Nikhil Thorat:

Does this error happen when you try to make a prediction? Can you paste the javascript code you're using to load and execute the model?

The error happens while I'm loading the model, which makes the Promise not return a value - so I can't test predictions.

This is literally the whole code needed to create the error:

model = await tf.loadModel(`http://localhost:3000/model.json`)

(group1-shard1of1 of course also being accessible through the Web server, and loading fine)

Am Freitag, 31. August 2018 18:54:51 UTC+2 schrieb Daniel Smilkov:

Also if you can send us the stack trace, that would be great!

Yeah, no problem! I use Webpack for compiling my frontend, but I hope it's still useful to you guys.

Error: Based on the provided shape, [2,256], the tensor should have 512 values but has 43

    at assert (tf-core.esm.js?45ef:17)

    at new e (tf-core.esm.js?45ef:17)

    at Function.e.make (tf-core.esm.js?45ef:17)

    at tensor (tf-core.esm.js?45ef:17)

    at o (tf-core.esm.js?45ef:17)

    at Object.decodeWeights (tf-core.esm.js?45ef:17)

    at eval (tf-layers.esm.js?271e:17)

    at eval (tf-layers.esm.js?271e:17)

    at Object.eval [as next] (tf-layers.esm.js?271e:17)

    at o (tf-layers.esm.js?271e:17)

[Shanqing Cai]

We have seen similar errors related to mismatch between the version of keras used to generate the saved h5 file and the version of keras that tensorflowjs depends on. 

Can you let me know what version of keras you used to train and save your keras model? 

[Timon Lukas]

I am using Keras v2.2.2, which is the latest version available on conda. I'm training and saving the model on Windows with Python 3.6.6 :)

[Shanqing Cai]

Thank for the reply, Timon.

I downloaded the "model.json" and "group1-shard1of1" files you kindly provided and tested loading them using "tf.loadModel". I used @tensorflow/tfjs version 0.12.6. The loading seems to work without any error in both the browser and node.js (with tfjs-node 0.1.14). Here is what model.summary() prints after the model is done loading:

_________________________________________________________________

Layer (type)                 Output shape              Param #   

=================================================================

lstm_1 (LSTM)                [30,5,32]                 4480      

_________________________________________________________________

dropout_1 (Dropout)          [30,5,32]                 0         

_________________________________________________________________

lstm_2 (LSTM)                [30,32]                   8320      

_________________________________________________________________

dropout_2 (Dropout)          [30,32]                   0         

_________________________________________________________________

dense_1 (Dense)              [30,3]                    99        

=================================================================

Total params: 12899

Trainable params: 12899

Non-trainable params: 0

_________________________________________________________________

I wonder what is different between your setup and mine. Could you be using an out-of-date version of @tensorflow/tfjs?

Best,

Shanqing

[Timon Lukas]

Thank you for your reply and testing, Shanqing!

This is very strange - I also use @tensorflow/tfjs with version 0.12.6. I am using Webpack to build my frontend (which tfjs is a part of), could that be the problem? Have you imported tfjs directly or through building it with webpack?

[Shanqing Cai]

I haven't tried tensorflow.js with webpack before. Others - please chime in if you have experience in that regard.

Timon - Can you check in the "network" tab of your browser's devtools? The tf.loadModel call should issue two HTTP/HTTPS requests for your model, one for the model.json and one for the binary weights file (group1-shard1of1). Can you check their contents are correct, i.e., the same as the files as seen on your local file system?

--

You received this message because you are subscribed to the Google Groups "TensorFlow.js Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfjs+uns...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tfjs/.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfjs/c6903bdc-2039-4a32-97ea-10e9fc457341%40tensorflow.org.

--

---

Shanqing Cai

Software Engineer

Google

ca...@google.com

[Timon Lukas]

Hey!

So I have tried around a little more. I was able to successfully load the file in a clean environment using TFJS directly, in Node and using Webpack. I have no idea why it did not work in the specific environment, but I suspect that something was wong with sending the data. I will further investigate and come back to you, but if you don't hear anything from me, it seems to work. Thank you!