Detect and lower depthwise conv to linalg.generic #2678
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
asaadaldien
force-pushed
the
ataei-linalg_depthwise_conv
branch
from
95aa614 to
4ce60a0
Compare
hanhanW
requested changes
Jul 28, 2020
Comment on lines
+395
to
+399
| // y[n, d1, d2, ....dn, ci * groupSize + co] = sum(k1, k2, ....kn, | ||
| // x[n, d1 * stride1 + k1, d1 * stride2 + k2, ...dn * striden + kn] | ||
| // * w[k1, k2, ...kn, ci, co]) |
There was a problem hiding this comment.
You should include dilation in the comment too since it is supported.
| }, | ||
| feature_group_count = 2 : i64, | ||
| padding = dense<0> : tensor<2x2xi64>, | ||
| rhs_dilation = dense<1> : tensor<2xi64>, |
There was a problem hiding this comment.
I dont know if we support dilation in convolution itself, but would be good to check that dilation loweres correctly (The linalg.conv lowers the dilation properly)
There was a problem hiding this comment.
There is a front end problem when constructing tf.depthwise_conv2d op the saved model doesn't have correct dilation attributes but default values instead. For conv I will add more test cases
So I will create a separate issues to track :
- Increase coverage of conv tests to include dialed convolution.
- Triage dilated depthwise convolution savedModel issue
asaadaldien
force-pushed
the
ataei-linalg_depthwise_conv
branch
2 times, most recently
from
5eba976 to
2cf9303
Compare
MaheshRavishankar
approved these changes
Aug 4, 2020
hanhanW
approved these changes
Aug 5, 2020
| @@ -26,19 +26,37 @@ class Conv2dModule(tf.Module): | |||
| tf.TensorSpec([2, 4, 5, 2], tf.float32), | |||
| tf.TensorSpec([2, 2, 2, 3], tf.float32), | |||
| ]) | |||
| def conv2d_2452x2223_valid(self, img, kernel): | |||
| def conv2d_2423x2223_valid(self, img, kernel): | |||
| return tf.nn.depthwise_conv2d( | ||
| img, kernel, [1, 1, 1, 1], "VALID", name="result") | ||
|
|
||
| @tf.function(input_signature=[ | ||
| tf.TensorSpec([2, 4, 5, 2], tf.float32), | ||
| tf.TensorSpec([2, 4, 2, 3], tf.float32), | ||
| ]) | ||
| def conv2d_2452x2223_same(self, img, kernel): | ||
| def conv2d_2423x2223_same(self, img, kernel): |
| tf.TensorSpec([2, 4, 5, 2], tf.float32), | ||
| tf.TensorSpec([2, 4, 2, 3], tf.float32), | ||
| ]) | ||
| def conv2d_2423x2223_valid_stride_2(self, img, kernel): |
| tf.TensorSpec([2, 4, 5, 2], tf.float32), | ||
| tf.TensorSpec([2, 4, 2, 3], tf.float32), | ||
| ]) | ||
| def conv2d_2423x2223_same_stride_2(self, img, kernel): |
| } | ||
| } | ||
| inputExprs.push_back(rewriter.getAffineDimExpr(ciIndex)); | ||
| // k1, k2, ...kn, ci, co |
There was a problem hiding this comment.
nit: add a blank line before the comment since they are for different exprs.
asaadaldien
force-pushed
the
ataei-linalg_depthwise_conv
branch
from
2cf9303 to
8344da9
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Lower 'mhlo.conv' with dpethwise convolution properties to 'linalg.generic' on Buffers.
Lowering depthwise convolution to linalg.generic op. The idea is to use the group convolution formulation to perform the separable depth wise convolution as the following, given an n-dimensional input x and filter w the direct convolution operation can be written as following expression that performs reduction over kernel dimensions k1..kn.