diff options
author | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2022-09-13 15:22:01 +0200 |
---|---|---|
committer | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2022-09-13 14:15:20 +0000 |
commit | 7f3ccd5500458de0b56f05ed99553360c46e6b41 (patch) | |
tree | 0312fcc74a126b884a8f9f806ac2e0da265f7868 /ethosu/vela/tflite_graph_optimiser.py | |
parent | b393251346c7a5eef1496bcc92c722ba92c73fd9 (diff) | |
download | ethos-u-vela-7f3ccd5500458de0b56f05ed99553360c46e6b41.tar.gz |
MLBEDSW-6929 Fix LeakyReLU int16 regressions
Fixed LeakyReLU regressions for int16 due to scaling introduced
for handling negative alpha.
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Change-Id: I84a494fedf54bd4b47c4632645ded7d6cda445f8
Diffstat (limited to 'ethosu/vela/tflite_graph_optimiser.py')
-rw-r--r-- | ethosu/vela/tflite_graph_optimiser.py | 13 |
1 files changed, 9 insertions, 4 deletions
diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py index 052f824c..fb98e21b 100644 --- a/ethosu/vela/tflite_graph_optimiser.py +++ b/ethosu/vela/tflite_graph_optimiser.py @@ -1049,20 +1049,25 @@ def convert_lrelu_to_mul_max(op, arch): quantization.min = 0 quantization.max = alpha * (quantization.quant_max - quantization.quant_min) quantization.zero_point = 0 + alpha_dtype = mul_ifm.dtype if "alpha_scaling" in op.attrs: # The LeakyRelu was the result from convert_prelu scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"] mul_alpha.type = Op.RescaleMul mul_alpha.rescale = [alpha_scale, alpha_shift] - elif np.isinf(1 / alpha): - # Handling of alpha near zero + elif alpha == 0 or np.isinf(1 / alpha): + # Handling of alpha near or at zero quantization.scale_f32 = np.float32(1) scalar = 0 else: quantization.scale_f32 = alpha - scalar, _ = scaling.elementwise_mul_scale(ifm.quantization.scale_f32, alpha, ofm.quantization.scale_f32) + if alpha_dtype == DataType.int32: + # When the datatype is int32 we need to do the scaling with the multiplication + scalar, _ = scaling.elementwise_mul_scale(ifm.quantization.scale_f32, alpha, ofm.quantization.scale_f32) + else: + scalar = 1 alpha_tens = create_const_tensor( - op.name + "_alpha_scalar", [1, 1, 1, 1], DataType.int32, [scalar], np.int32, quantization=quantization + op.name + "_alpha_scalar", [1], alpha_dtype, [scalar], alpha_dtype.as_numpy_type(), quantization=quantization ) mul_alpha.add_input_tensor(alpha_tens) fm_alpha = ofm.clone(op.name + "_alpha", set_unique=True) |