diff options
| author | Louis Verhaard <louis.verhaard@arm.com> | 2021-01-07 13:35:47 +0100 |
|---|---|---|
| committer | patrik.gustavsson <patrik.gustavsson@arm.com> | 2021-01-22 13:30:25 +0000 |
| commit | ece4e65a786d61934d0c356169a3c5f11f188538 (patch) | |
| tree | d131e131aa915ad200185e7705861422878a730f | |
| parent | d4738e5d68260ed56c0b878cd5dc11c67fdbbfaa (diff) | |
| download | ethos-u-vela-ece4e65a786d61934d0c356169a3c5f11f188538.tar.gz | |
MLBEDSW-3418: Bug fixes 16-bit leaky relu
Bug fixes for 16-bit leaky relu with different quantizations for IFM/OFM:
- Overflow error occurred for alpha == 0
- The identity multiplication overwrote the result of the alpha
multiplication
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
Change-Id: I18f8d121f6e7c598b721c472b476b9285eeff543
| -rw-r--r-- | ethosu/vela/graph_optimiser.py | 15 |
1 files changed, 12 insertions, 3 deletions
diff --git a/ethosu/vela/graph_optimiser.py b/ethosu/vela/graph_optimiser.py index 3c80658..5f11178 100644 --- a/ethosu/vela/graph_optimiser.py +++ b/ethosu/vela/graph_optimiser.py @@ -918,9 +918,17 @@ def convert_lrelu_to_mul_max(op, arch): quantization = ifm.quantization.clone() quantization.min = 0 quantization.max = alpha * (quantization.quant_max - quantization.quant_min) - quantization.scale_f32 = alpha quantization.zero_point = 0 - alpha_tens = create_const_tensor(op.name + "_alpha_scalar", [], ifm.dtype, [1], np.int8, quantization=quantization) + if np.isinf(1 / np.float32(alpha)): + # Handling of alpha near zero + quantization.scale_f32 = 1 + scalar = 0 + else: + quantization.scale_f32 = alpha + scalar = 1 + alpha_tens = create_const_tensor( + op.name + "_alpha_scalar", [], ifm.dtype, [scalar], np.int8, quantization=quantization + ) mul_alpha.add_input_tensor(alpha_tens) fm_alpha = ofm.clone(op.name + "_alpha") mul_alpha.set_output_tensor(fm_alpha) @@ -944,7 +952,8 @@ def convert_lrelu_to_mul_max(op, arch): op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization ) mul_identity.add_input_tensor(identity_tens) - fm_id = ofm.clone(op.name + "_id") + # Make sure that fm_id is allocated to a different address than fm_alpha + fm_id = ofm.clone(op.name + "_id", set_unique=True) mul_identity.set_output_tensor(fm_id) mul_identity.set_ifm_ofm_shapes() DebugDatabase.add_optimised(op, mul_identity) |