Improve number formatting in tensor print #7632

li-roy · 2018-05-17T00:44:17Z

Integer tensors won't be printed in scientific notation
Removed scaling factor for all tensors
Only leave spaces for negative signs if one of the printed elements is going to have an negative sign

Fixes #7213, fixes #6918, fixes #6865, fixes #6811, fixes #6768.

Examples

>>> torch.tensor(2341234123412341)
tensor(2341234123412341)

>>> torch.tensor([1e28, 1e-28])
tensor([1.0000e+28, 1.0000e-28])

>>> torch.tensor([1e28, -1e-28])
tensor([ 1.0000e+28, -1.0000e-28])

>>> torch.tensor([1, 2])
tensor([1, 2])

>>> torch.tensor([1, -2])
tensor([ 1, -2])

>>> torch.randn(1001)
tensor([-0.5033, -0.4142,  0.7961,  ..., -0.9400, -0.4431, -0.6192])

>>> torch.randn(20)
tensor([ 0.8574, -1.1225,  0.3135,  1.7848, -0.8342,  0.1214, -0.6393, -0.3480,
         1.4114,  0.5102, -0.8206,  1.6022,  0.6310,  1.9507, -1.5694,  1.3168,
         0.1103,  0.5242,  0.3141, -0.0692])

>>> torch.randn(5, dtype=torch.double)
tensor([-1.2961,  0.6950,  0.1357, -0.6473,  0.2605], dtype=torch.float64)

ssnl · 2018-05-17T03:10:14Z

Cool! Can you post some example outputs?

torch/_tensor_str.py

fmassa · 2018-05-22T20:45:59Z

I think the prints now are a bit verbose.
Can't we print the dtype/device/requires_grad if they are not what we would expect to be default?

Also, quite a few people wanted to have a shape information in the print. While this would break the constructor-valid contract, it could be worth considering.

li-roy · 2018-05-22T20:54:12Z

That's what it was like before, but there was some discussion on #6912 on whether it makes sense to do that, seeing as those defaults can be changed.

ezyang · 2018-05-22T20:59:57Z

Given that we are adding ellipses (which breaks Python parseability), I don't think slavishly following the "it must be a valid Python expression" makes sense. For example, we can add the shape as a comment as a newline at the very end of the expression. This is not strictly following the Python guidelines, but it follows the spirit and solves the problem.

fmassa · 2018-05-22T21:09:15Z

I personally would only print requires_grad if it's True. Same thing for device, only print it if it's not on the CPU.

Also, I think that at some point this becomes a matter of taste, and maybe adding options to the printoptions to always enable / disable certain prints might be worth considering.

alok · 2018-05-22T23:32:00Z

I don't think there should be a space before the first element of a tensor, since numpy and python lists don't do that

li-roy · 2018-05-24T00:47:09Z

@alok Numpy does when it's leaving space for a negative.

>>> np.array([1,-2])
array([ 1, -2])

li-roy · 2018-06-04T19:57:33Z

@pytorchbot retest this please

li-roy · 2018-06-05T21:16:42Z

@pytorchbot retest this please

ezyang · 2018-06-06T02:46:01Z

So, what does the new formatting look like? An expect test for a simple tensor would go a long way towards making it more obvious to reviewers.

test/test_torch.py

torch/_tensor_str.py

li-roy · 2018-06-13T19:16:40Z

@pytorchbot retest this please

ezyang · 2018-06-13T19:55:02Z

FYI, you can use assertExpected instead of hard coding the strings, and then pass --accept to automatically accept the actual string as the expected result. Saves a bunch of copy pasting.

li-roy · 2018-06-13T22:41:33Z

@pytorchbot retest this please

gchanan · 2018-06-13T22:46:10Z

FWIW I'm not a fan of assertExpected here -- these are really simple tests and it's nicer to just have them visible inline. But it's fine now that it's done.

gchanan

I have some small comments / suggestions, but I think this is good to go.

torch/_tensor_str.py

+                self.max_width = max(self.max_width, len(value_str))

+        else:
+            copy = torch.DoubleTensor(tensor.size()).copy_(tensor).view(tensor.nelement())


torch/_tensor_str.py

+                    if math.isnan(value) or math.isinf(value):
+                        self.max_width = max(self.max_width, len(value_str))
+                    else:
+                        # for finites, appending a decimal will increase the len by 1


torch/_tensor_str.py

-            else:
-                if exp_max == 0:
-                    sz = 7
+                if exp_max - exp_min > PRINT_OPTS.precision or exp_max > 8 or exp_min < -4:


torch/_tensor_str.py

+                ret = ('{:' + str(self.max_width) +
+                       '.' + str(PRINT_OPTS.precision) + 'e}').format(value)
+            else:
+                ret = ('{:.' + str(PRINT_OPTS.precision) + 'f}').format(value)


ezyang · 2018-06-14T02:23:33Z

@pytorchbot retest this please

li-roy requested review from apaszke, colesbury, ezyang, gchanan, soumith and zdevito as code owners May 17, 2018 00:44

gchanan reviewed May 18, 2018

View reviewed changes

li-roy force-pushed the printn branch from 4e9c0b6 to 7fa9a39 Compare May 31, 2018 20:16

soumith mentioned this pull request Jun 6, 2018

Print requires_grad and grad_fn in string repr of tensor #8211

Merged

li-roy force-pushed the printn branch 2 times, most recently from 4ed7b4f to 5145f95 Compare June 11, 2018 21:31

gchanan reviewed Jun 12, 2018

View reviewed changes

test/test_torch.py Outdated

This comment was marked as off-topic.

Sign in to view

gchanan reviewed Jun 12, 2018

View reviewed changes

test/test_torch.py Outdated

This comment was marked as off-topic.

Sign in to view

gchanan reviewed Jun 12, 2018

View reviewed changes

Roy Li added 3 commits June 12, 2018 18:53

Improve number formatting in tensor print

3555f07

fix bad rebase

d725581

address comments

604bb36

li-roy force-pushed the printn branch from 3862218 to 604bb36 Compare June 13, 2018 02:02

fix test

1c7bf09

Roy Li added 2 commits June 13, 2018 12:56

fix test

f6e3975

use assertExpected for tests

42236a5

gchanan approved these changes Jun 13, 2018

View reviewed changes

address comments

f91570a

address comments

10f7365

li-roy merged commit 6a85b13 into pytorch:master Jun 14, 2018

li-roy deleted the printn branch June 14, 2018 06:57

soumith mentioned this pull request Aug 6, 2018

Ugly and poorer tensor visualisation with v0.4.0 #6912

Closed

Improve number formatting in tensor print #7632

Improve number formatting in tensor print #7632

Uh oh!

Conversation

li-roy commented May 17, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ssnl commented May 17, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

fmassa commented May 22, 2018

Uh oh!

li-roy commented May 22, 2018

Uh oh!

ezyang commented May 22, 2018

Uh oh!

fmassa commented May 22, 2018

Uh oh!

alok commented May 22, 2018

Uh oh!

li-roy commented May 24, 2018

Uh oh!

li-roy commented Jun 4, 2018

Uh oh!

li-roy commented Jun 5, 2018

Uh oh!

ezyang commented Jun 6, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

li-roy commented Jun 13, 2018

Uh oh!

ezyang commented Jun 13, 2018

Uh oh!

li-roy commented Jun 13, 2018

Uh oh!

gchanan commented Jun 13, 2018

Uh oh!

gchanan left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

li-roy commented May 17, 2018 •

edited

Loading