Fix inception v3 input transform for trace & onnx #621

BowenBao · 2018-10-08T23:00:39Z

Input transform are in-place updates, which produce issues for tracing
and exporting to onnx.

The tracing issue can be reproduced by

dummy_input = torch.Tensor(torch.randn((10, 3, 299, 299))).cuda()
inception_model = models.inception_v3(pretrained=True).cuda()
inception_model = inception_model.eval()
torch.jit.trace(inception_model, dummy_input)
>>> TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error:
Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 735] (3.154785633087158 vs. 0.27894413471221924) and 9999 other locations (100.00%)
  _check_trace([example_inputs], func, executor_options, module, check_tolerance)

This fix references the section Tracer Warnings from https://pytorch.org/docs/master/jit.html.

* Input transform are in-place updates, which produce issues for tracing and exporting to onnx.

torchvision/models/inception.py

-            x[:, 1] = x[:, 1] * (0.224 / 0.5) + (0.456 - 0.5) / 0.5
-            x[:, 2] = x[:, 2] * (0.225 / 0.5) + (0.406 - 0.5) / 0.5
+            x = torch.cat((
+                torch.unsqueeze(x[:, 0], 1) * (0.229 / 0.5) + (0.485 - 0.5) / 0.5,


soumith

looks correct to me. thanks a lot for this patch!

torchvision/models/inception.py

+            x_ch0 = torch.unsqueeze(x[:, 0], 1) * (0.229 / 0.5) + (0.485 - 0.5) / 0.5
+            x_ch1 = torch.unsqueeze(x[:, 1], 1) * (0.224 / 0.5) + (0.456 - 0.5) / 0.5
+            x_ch2 = torch.unsqueeze(x[:, 2], 1) * (0.225 / 0.5) + (0.406 - 0.5) / 0.5
+            x = torch.cat((x_ch0, x_ch1, x_ch2), 1)


…ytorch#621) * [DLRMv2] Resolve benchmark name and use other constants for logging * Use mllog_constants.ADAGRAD for optimizer name

Fix inception v3 input transform for trace & onnx

d86c93e

* Input transform are in-place updates, which produce issues for tracing and exporting to onnx.

BowenBao commented Oct 8, 2018

View reviewed changes

torchvision/models/inception.py Outdated

x[:, 1] = x[:, 1] * (0.224 / 0.5) + (0.456 - 0.5) / 0.5

x[:, 2] = x[:, 2] * (0.225 / 0.5) + (0.406 - 0.5) / 0.5

x = torch.cat((

torch.unsqueeze(x[:, 0], 1) * (0.229 / 0.5) + (0.485 - 0.5) / 0.5,

This comment was marked as off-topic.

Sign in to view

nit

fe8fe70

BowenBao mentioned this pull request Oct 16, 2018

PyTorch/vision pretrained model inception v3 input transform issue for trace & onnx. pytorch/pytorch#12719

Closed

soumith approved these changes Oct 25, 2018

View reviewed changes

soumith merged commit 85369e3 into pytorch:master Oct 25, 2018

fmassa reviewed Oct 25, 2018

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix inception v3 input transform for trace & onnx #621

Fix inception v3 input transform for trace & onnx #621

BowenBao commented Oct 8, 2018

This comment was marked as off-topic.

soumith left a comment

This comment was marked as off-topic.

This comment was marked as off-topic.

Fix inception v3 input transform for trace & onnx #621

Fix inception v3 input transform for trace & onnx #621

Conversation

BowenBao commented Oct 8, 2018

This comment was marked as off-topic.

soumith left a comment

Choose a reason for hiding this comment

This comment was marked as off-topic.

This comment was marked as off-topic.