Paddings values for pad are loaded 1D instead of 2D from SavedModel

[JornEngelbart]

TensorFlow.js version

0.8.0

Browser version

Chrome 65.0.3325.181

Describe the problem or feature request

When loading a Tensorflow SavedModel including a pad using the loadFrozenModel function and run a simple execution it will give an error Uncaught (in promise) Error: Failed to compile fragment shader.. After some investigating it seems like the paddings value are injected as an 1D array where it should be 2D. E.g. paddings [[0, 0], [2, 0]] are injected as [0, 0, 2, 0]. This results in the map function at https://github.com/tensorflow/tfjs-core/blob/75b2596a46db8560a6d27b62397dacf366a9347e/src/kernels/webgl/pad_gpu.ts#L34 to output a list of NaNs. As far as I can see the paddings are checked when building your own model using tfjs here https://github.com/tensorflow/tfjs-core/blob/23970ba5876d54c704b8e40ffe07f976e725c45a/src/ops/array_ops.ts#L779, so I suspect it might have to do with loading the model from a (Tensorflow) SavedModel. The weightManifest in the model shows that the paddings are loaded with (correct) shape [2, 2].
Unfortunately I'm not familiar enough with the code base to PR it myself.

Code to reproduce the bug / link to feature request

See above

[dsmilkov]

@pyu10055 can you take a look at this? @JornEngelbart is it possible to share your model and weight files (post conversion is ok) to help us reproduce the problem?

[JornEngelbart]

@pyu10055 @dsmilkov I made a quick mock version.
The model was generated with:

import tensorflow as tf

input = tf.placeholder(tf.float32, [5, 1])

layer = tf.pad(input, [[1, 2], [3, 4]])

with tf.Session() as sess:
    tf.global_variables_initializer()
    print(sess.run(layer, feed_dict={input: [[1],[2],[3],[4],[5]]}))
    tf.saved_model.simple_save(sess, "saved_model/", inputs={"in": input}, outputs={"layer": layer})

The model can be executed with something like

const model = await loadFrozenModel(MODEL_URL, WEIGHTS_URL);
var start = model.execute({
    "Placeholder": tf.tensor([[1],[2],[3],[4],[5]], [5,1], "float32")
});
var out = await start.data()

The model and weight files are attached.
model.zip

[pyu10055]

fix added to the converter repo, and release on v0.1.2