Add Lowering for conv2d and conv2d_transpose to ttir::ConvolutionOp #3043
Conversation
rafey1104
requested review from
acolicTT,
dgolubovicTT,
nvukobratTT and
pilkicTT
as code owners
github-actions
bot
added
the
community
|
Testing it is required. I could not test it since i lost the koyeb access |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
forge/test/mlir/test_ops_tf.py
Outdated
| ( | ||
| (1, 16, 64, 56, 56, 4, 4, 1, 1), # Flipped output/input channels | ||
| (1, 64, 64, 56, 56, 3, 3, 1, 1), | ||
| (1, 128, 128, 28, 28, 3, 3, 2, 2), # Stride > 1 |
Check failure
Code scanning / flake8
at least two spaces before inline comment Error test
forge/test/mlir/test_ops_tf.py
Outdated
| (1, 64, 64, 56, 56, 3, 3, 1, 1), | ||
| (1, 128, 128, 28, 28, 3, 3, 2, 2), # Stride > 1 | ||
| (1, 128, 128, 14, 14, 3, 3, 1, 1), | ||
| (1, 256, 256, 14, 14, 3, 3, 2, 2), # Stride > 1 |
Check failure
Code scanning / flake8
at least two spaces before inline comment Error test
forge/test/mlir/test_ops_tf.py
Outdated
| (1, 64, 64, 8, 8, 3, 3, 1, 1), | ||
| (1, 64, 64, 16, 16, 3, 3, 1, 1), | ||
| (1, 256, 256, 7, 7, 3, 3, 1, 1), | ||
| (1, 64, 256, 28, 28, 1, 1, 2, 2), # Flipped output/input channels, Stride > 1 |
Check failure
Code scanning / flake8
at least two spaces before inline comment Error test
forge/test/mlir/test_ops_tf.py
Outdated
| def __init__(self): | ||
| super().__init__() | ||
| self.conv2d_transpose = tf.keras.layers.Conv2DTranspose( | ||
| output_channels, # `filters` argument specifies output channels |
Check failure
Code scanning / flake8
at least two spaces before inline comment Error test
|
|
||
| compiled_model = forge.compile(framework_model, sample_inputs=inputs) | ||
|
|
||
| verify(inputs, framework_model, compiled_model) |
Check warning
Code scanning / flake8
no newline at end of file Warning test
Thanks for the contribution @rafey1104! Before our initial review, @marty1885, can we help @rafey1104 with his instance to test out the changes first? |
|
Sure! @rafey1104 can you send an message to |
3 participants
Add Lowering for conv2d and conv2d_transpose to ttir::ConvolutionOp
Ticket
Fix : 596
Problem description
Lower TTForge
conv2dandconv2d_transposeops to the new, unifiedttir::ConvolutionOp.What's changed
emit_mlir_ttforge_convolution_opfor loweringconv2dandconv2d_transpose.lowering_handler_mapto use the new handler for both convolution op types.AttributeMapperto remapgroups->feature_groups_countandstride->window_strides.ConvolutionLayoutAttrassuming a channel-last (NHWC) layout.ttir.nonewhen no bias is present.is_transposeattribute inttir::ConvolutionOpbased on the source op.test_conv2d_transposeto validate the new lowering path.