[Frontend][TFLite] L2_POOL_2D operator #5452
Conversation
| in_data = tf.placeholder( | ||
| dtype=tf.float32, name="input", shape=input_shape) | ||
| out = tf.sqrt(tf.nn.avg_pool( | ||
| tf.square(in_data), ksize=ksize, strides=strides, | ||
| padding=padding, data_format=data_format)) | ||
| out = with_fused_activation_function(out, fused_func_name) | ||
|
|
||
| compare_tflite_with_tvm(x, 'input', [in_data], [out]) |
There was a problem hiding this comment.
Very quick drive-by review :
Don't we need a qnn test here ? I haven't managed to read up L2_POOL_2D in the tflite documentation yet.
There was a problem hiding this comment.
TFLite doesn't have support for quantized L2_POOL_2D yet.
There was a problem hiding this comment.
Thanks , possibly worth a belts and braces check to ensure we catch this in the frontend to ensure we think through qnn support rather than getting an unknown failure.
python/tvm/relay/frontend/tflite.py
Outdated
| @@ -1770,6 +1775,12 @@ def convert_pool2d(self, op, pool_type): | |||
| assert self.has_same_qnn_params(input_tensor, output_tensor), \ | |||
| "qnn.op.max_pool2d requires input and output qnn params to be same" | |||
| out = _op.nn.max_pool2d(in_expr, **params) | |||
| elif pool_type == "l2": | |||
| # l2_pool_2d is equivalent to square_root(avg_pool(square(in_data))) | |||
There was a problem hiding this comment.
Add a protection check to throw error if you receive quantized inputs.
There was a problem hiding this comment.
TFLite does not support quantized L2_POOL_2D. That's why the check is not added. Now I have added a comment for the same.
There was a problem hiding this comment.
Putting in the check has the following benefits:
- Better user experience as the rest of the stack doesn't get basically undefined behaviour because of random input..
- A signal to us as a community in the form of a bug report that this is appearing.
Ramana
There was a problem hiding this comment.
ok, make sense. I have added the check.
|
@u99127, @siju-samuel, @FrozenGene, Thanks. The review comments are addressed. Could you please help in merging the PR. |
|
Thanks @maheshambule @siju-samuel @u99127 merged now |
* TFLITE fill and splitv ops * l2_pool_2d op changes in comment * TFLite l2_pool_2d op added test case in main * TFLite L2_POOL_2D added check for quantized input
* TFLITE fill and splitv ops * l2_pool_2d op changes in comment * TFLite l2_pool_2d op added test case in main * TFLite L2_POOL_2D added check for quantized input
* TFLITE fill and splitv ops * l2_pool_2d op changes in comment * TFLite l2_pool_2d op added test case in main * TFLite L2_POOL_2D added check for quantized input
Adds support for L2_POOL_2D operator in TFLITE frontened.
Ref:
https://www.tensorflow.org/lite/guide/ops_compatibility
@u99127, @siju-samuel, @FrozenGene Please review.