sfraczek removed a comment on issue #19235:
URL:
https://github.com/apache/incubator-mxnet/issues/19235#issuecomment-700688990
I'm not experienced with Mxnet yet but I made an experiment comparing both
functions and the **old version code** seems to not work properly or I am
misunderstanding something. The new code seems to return results that look like
they make sense.
For array arr containing `[ 0.990018 -0.92912527 0.88925551 0.29950806
0.5652269 0.91779029 0.70329186 0.93344053 0.26962166 0.23452311]`
the old returns:
min: 0.0,
max: 1.0,
divergence: 0.012564257718622684
threshold: 1.0
the new returns:
min: -0.92912527
max: 0.990018
divergence: [0.]
threshold: [0.9338415]
This is my test code:
```python
@with_seed()
def test_optimal_threshold():
arr = np.array([0.990018, -0.92912527, 0.88925551, 0.29950806,
0.5652269, 0.91779029, 0.70329186, 0.93344053, 0.26962166, 0.23452311])
num_bins = 8001
num_quantized_bins=255
min_val = np.min(arr)
max_val = np.max(arr)
th = max(abs(min_val), abs(max_val))
hist, hist_edges = np.histogram(arr, bins=num_bins, range=(-th, th))
print("\narr", arr)
# print("hist", np.nonzero(hist))
min_val1, max_val1, min_divergence, opt_th =
mx.contrib.quant._get_optimal_threshold_old(mx.nd.array(hist),
num_bins=num_bins, num_quantized_bins=num_quantized_bins)
print("old", min_val1, max_val1, min_divergence, opt_th)
quantized_dtype = 'int8'
hist_data = (hist, hist_edges, min_val, max_val, max_val)
min_val2, max_val2, threshold, divergence =
mx.contrib.quant._get_optimal_threshold(hist_data, quantized_dtype,
num_quantized_bins=num_quantized_bins)
print("new", min_val2, max_val2, divergence, threshold)
```
I hope this helps?
----------------------------------------------------------------
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]
