sxjscience opened a new issue #19595:
URL: https://github.com/apache/incubator-mxnet/issues/19595
## Description
I noticed a serialization bug when rerunning the machine translation example
in GluonNLP. This might be due to the recently upgraded serialization format.
Here's the reproducible example:
```python
import mxnet as mx
import numpy as np
mx.npx.set_np()
net = mx.gluon.nn.Embedding(input_dim=32768, output_dim=512)
net.initialize()
net.save_parameters('temp.params')
original_weight = net.collect_params()['weight'].data().asnumpy()
loaded_params = mx.npx.load('temp.params')
loaded_weight = loaded_params['weight'].asnumpy()
np.testing.assert_allclose(original_weight, loaded_weight, 1E-4, 1E-4)
```
Output:
```
---> 11 np.testing.assert_allclose(original_weight, loaded_weight, 1E-4,
1E-4)
[... skipping hidden 2 frame]
AssertionError:
Not equal to tolerance rtol=0.0001, atol=0.0001
Mismatched elements: 16735981 / 16777216 (99.8%)
Max absolute difference: 0.13994947
Max relative difference: 4780370.5
x: array([[-0.019636, 0.020329, -0.045591, ..., -0.036925, -0.007389,
0.043669],
[-0.01338 , 0.055998, 0.046903, ..., -0.006462, 0.001994,...
y: array([[-0.019636, 0.020329, -0.045591, ..., -0.036925, -0.007389,
0.043669],
[-0.01338 , 0.055998, 0.046903, ..., -0.006462, 0.001994,...
```
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
