[res/tfl_recipes] Add new Net_FullyConnected_Mul

[jiwaszki]

This commit extends Net_FullyConnectedMul tflite recipes with 'no bias' case.

ONE-DCO-1.0-Signed-off-by: Jan Iwaszkiewicz [email protected]

[jiwaszki]

Related PRs:

• [circle2circle] Dredd test for FuseMulWithFullyConnectedPass #13612
• [luci/pass] Value test for FuseMulWithFullyConnectedPass #13611
• [res/tfl_recipes] Add Net_FullyConnected_Mul #13610

[seanshpark]

there is bias in the test.recipe file but you've noted as _
which I understand this as no bias. This is confusing.

if 003 is with bias, I think we need another 004 without bias.

[jiwaszki]

There is no bias here, it's a special test case requested here: #13607 (comment)

Here is the circle visual:

The third input of FC is empty:

ONE/res/TensorFlowLiteRecipes/Net_FullyConnected_Mul_003/test.recipe

Lines 36 to 46 in e3ff517

	operation {
	type: "FullyConnected"
	fullyconnected_options {
	activation: NONE
	keep_num_dims: true
	}
	input: "ifm"
	input: "fc_wgt"
	input: ""
	output: "fc_out"
	}

The B operand is the input of Mul (as in A * B):

ONE/res/TensorFlowLiteRecipes/Net_FullyConnected_Mul_003/test.recipe

Lines 47 to 55 in e3ff517

	operation {
	type: "Mul"
	mul_options {
	activation: RELU
	}
	input: "fc_out"
	input: "B"
	output: "mul_out"
	}

Is it okay or is there another notation for empty bias?

[seanshpark]

Ah, thanks for the explanation.
"B" made the confusion.
It would be nice if you renamed B as mutiplicand or scale.

Anyway, is there existing case WITH BIAS? if not, can you add one?

[jiwaszki]

#13610 -- here are already merged models, all of them have fc_bias.

It would be nice if you renamed B as mutiplicand or scale.

Will do!

[seanshpark]

LGTM