The "lift_lexical_references" pass should not lead in an invalid model, but extract metadata instead.

[researcher-no-name]

I tried to use the "lift_lexical_references" pass on a number of models. I think that the pass removes the references arbitrarily, without checking if they are used in the model (e.g., on the initializer). I am aware that "This transformation yields a graph that does not conform to the ONNX spec.", but essentially, its purpose to "expose the data dependencies within control blocks for frameworks that use those dependencies to schedule parallel execution." is defeated that way.

How about, instead of lifting the references inside the active blocks and resulting to an invalid model, isolate the extracted information to a separate metadata file (e.g., a JSON file describing the in-scope dependencies), or inject extra attributes to related nodes? This could allow the model to be used by frameworks utilizing parallel execution, but the initial model would remain ONNX spec compliant.

As a result, while running the optimizer with this pass, I encountered upon optimization for all affected models:
line 46, in optimize\\n optimized_model_str = C.optimize(model_str, passes)

EfficientNet-Lite4 (opset=11, obtained from ONNX Hub):
Unresolved value references: efficientnet-lite4/model/blocks_0/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_0/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_0/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_0/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_1/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_1/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_1/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_1/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_1/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_1/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_1/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_1/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_1/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_10/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_10/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_10/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_10/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_10/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_10/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_10/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_10/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_10/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_11/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_11/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_11/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_11/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_11/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_11/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_12/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_12/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_12/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_12/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_12/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_12/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_13/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_13/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_13/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_13/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_13/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_13/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_14/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_14/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_14/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_14/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_14/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_14/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_15/conv2d/Conv2D_bias_fused_bn,efficintnet-lite4/model/blocks_15/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_15/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_15/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_15/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_15/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_16/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_16/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_16/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_16/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_16/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_16/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_16/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_16/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_16/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_17/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_17/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_17/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_17/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_17/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_17/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_18/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_18/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_18/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_18/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_18/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_18/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_19/Relu6_1_max__380,efficientnet-lite4/model/blocks_19/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_19/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_19/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_19/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_19/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_19/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_2/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_2/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_2/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_2/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_2/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_2/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_2/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_2/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_2/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_20/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_20/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_20/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_20/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_20/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_20/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_21/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_21/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_21/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_21/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_21/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_21/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_22/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_22/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_22/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_22/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_22/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_22/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_22/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_22/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_22/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_23/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_23/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_23/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_23/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_23/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_23/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_24/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_24/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_24/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_24/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_24/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_24/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_25/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_25/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_25/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_25/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_25/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_25/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_26/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_26/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_26/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_26/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_26/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_26/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_27/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_27/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_27/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_27/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_27/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_27/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_28/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_28/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_28/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_28/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_28/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_28/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_29/Relu6_1_min__569,efficientnet-lite4/model/blocks_29/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_29/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_29/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_29/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_29/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_29/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_3/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_3/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_3/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_3/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_3/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_3/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_4/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_4/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_4/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_4/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_4/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_4/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_5/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_5/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_5/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_5/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_5/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_5/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_6/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/blocks_6/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_6/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_6/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_6/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_6/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/blocks_6/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_6/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/blocks_6/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/blocks_7/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_7/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_7/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_7/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_7/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_7/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_8/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_8/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_8/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_8/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_8/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_8/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/blocks_9/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_9/conv2d/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_9/conv2d_1/Conv2D_bias_fused_bn,efficientnet-lite4/model/blocks_9/conv2d_1/Conv2D_weights_fused_bn,efficientnet-lite4/model/blocks_9/depthwise_conv2d/depthwise_bias_fused_bn,efficientnet-lite4/model/blocks_9/depthwise_conv2d/depthwise_weights_fused_bn,efficientnet-lite4/model/head/conv2d/Conv2D/ReadVariableOp:0,efficientnet-lite4/model/head/dense/BiasAdd/ReadVariableOp:0,efficientnet-lite4/model/head/dense/MatMul/ReadVariableOp:0,efficientnet-lite4/model/head/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp:0,efficientnet-lite4/model/head/tpu_batch_normalization/FusedBatchNormV3/ReadVariableOp_1:0,efficientnet-lite4/model/head/tpu_batch_normalization/ReadVariableOp:0,efficientnet-lite4/model/head/tpu_batch_normalization/ReadVariableOp_1:0,efficientnet-lite4/model/stem/conv2d/Conv2D_bias_fused_bn,efficientnet-lite4/model/stem/conv2d/Conv2D_weights_fused_bn

SSD (opset=12, from ONNX Hub):
Unresolved value references: backbone.additional_blocks.0.0.bias,backbone.additional_blocks.0.0.weight,backbone.additional_blocks.0.2.bias,backbone.additional_blocks.0.2.weight,backbone.additional_blocks.1.0.bias,backbone.additional_blocks.1.0.weight,backbone.additional_blocks.1.2.bias,backbone.additional_blocks.1.2.weight,backbone.additional_blocks.2.0.bias,backbone.additional_blocks.2.0.weight,backbone.additional_blocks.2.2.bias,backbone.additional_blocks.2.2.weight,backbone.additional_blocks.3.0.bias,backbone.additional_blocks.3.0.weight,backbone.additional_blocks.3.2.bias,backbone.additional_blocks.3.2.weight,backbone.additional_blocks.4.0.bias,backbone.additional_blocks.4.0.weight,backbone.additional_blocks.4.2.bias,backbone.additional_blocks.4.2.weight,backbone.conf.0.bias,backbone.conf.0.weight,backbone.conf.1.bias,backbone.conf.1.weight,backbone.conf.2.bias,backbone.conf.2.weight,backbone.conf.3.bias,backbone.conf.3.weight,backbone.conf.4.bias,backbone.conf.4.weight,backbone.conf.5.bias,backbone.conf.5.weight,backbone.loc.0.bias,backbone.loc.0.weight,backbone.loc.1.bias,backbone.loc.1.weight,backbone.loc.2.bias,backbone.loc.2.weight,backbone.loc.3.bias,backbone.loc.3.weight,backbone.loc.4.bias,backbone.loc.4.weight,backbone.loc.5.bias,backbone.loc.5.weight,backbone.model.layer1.0.weight,backbone.model.layer1.1.bias,backbone.model.layer1.1.running_mean,backbone.model.layer1.1.running_var,backbone.model.layer1.1.weight,backbone.model.layer1.4.0.bn1.bias,backbone.model.layer1.4.0.bn1.running_mean,backbone.model.layer1.4.0.bn1.running_var,backbone.model.layer1.4.0.bn1.weight,backbone.model.layer1.4.0.bn2.bias,backbone.model.layer1.4.0.bn2.running_mean,backbone.model.layer1.4.0.bn2.running_var,backbone.model.layer1.4.0.bn2.weight,backbone.model.layer1.4.0.conv1.weight,backbone.model.layer1.4.0.conv2.weight,backbone.model.layer1.4.1.bn1.bias,backbone.model.layer1.4.1.bn1.running_mean,backbone.model.layer1.4.1.bn1.running_var,backbone.model.layer1.4.1.bn1.weight,backbone.model.layer1.4.1.bn2.bias,backbone.model.layer1.4.1.bn2.running_mean,backbone.model.layer1.4.1.bn2.running_var,backbone.model.layer1.4.1.bn2.weight,backbone.model.layer1.4.1.conv1.weight,backbone.model.layer1.4.1.conv2.weight,backbone.model.layer1.4.2.bn1.bias,backbone.model.layer1.4.2.bn1.running_mean,backbone.model.layer1.4.2.bn1.running_var,backbone.model.layer1.4.2.bn1.weight,backbone.model.layer1.4.2.bn2.bias,backbone.model.layer1.4.2.bn2.running_mean,backbone.model.layer1.4.2.bn2.running_var,backbone.model.layer1.4.2.bn2.weight,backbone.model.layer1.4.2.conv1.weight,backbone.model.layer1.4.2.conv2.weight,backbone.model.layer1.5.0.bn1.bias,backbone.model.layer1.5.0.bn1.running_mean,backbone.model.layer1.5.0.bn1.running_var,backbone.model.layer1.5.0.bn1.weight,backbone.model.layer1.5.0.bn2.bias,backbone.model.layer1.5.0.bn2.running_mean,backbone.model.layer1.5.0.bn2.running_var,backbone.model.layer1.5.0.bn2.weight,backbone.model.layer1.5.0.conv1.weight,backbone.model.layer1.5.0.conv2.weight,backbone.model.layer1.5.0.downsample.0.weight,backbone.model.layer1.5.0.downsample.1.bias,backbone.model.layer1.5.0.downsample.1.running_mean,backbone.model.layer1.5.0.downsample.1.running_var,backbone.model.layer1.5.0.downsample.1.weight,backbone.model.layer1.5.1.bn1.bias,backbone.model.layer1.5.1.bn1.running_mean,backbone.model.layer1.5.1.bn1.running_var,backbone.model.layer1.5.1.bn1.weight,backbone.model.layer1.5.1.bn2.bias,backbone.model.layer1.5.1.bn2.running_mean,backbone.model.layer1.5.1.bn2.running_var,backbone.model.layer1.5.1.bn2.weight,backbone.model.layer1.5.1.conv1.weight,backbone.model.layer1.5.1.conv2.weight,backbone.model.layer1.5.2.bn1.bias,backbone.model.layer1.5.2.bn1.running_mean,backbone.model.layer1.5.2.bn1.running_var,backbone.model.layer1.5.2.bn1.weight,backbone.model.layer1.5.2.bn2.bias,backbone.model.layer1.5.2.bn2.running_mean,backbone.model.layer1.5.2.bn2.running_var,backbone.model.layer1.5.2.bn2.weight,backbone.model.layer1.5.2.conv1.weight,backbone.model.layer1.5.2.conv2.weight,backbone.model.layer1.5.3.bn1.bias,backbone.model.layer1.5.3.bn1.running_mean,backbone.model.layer1.5.3.bn1.running_var,backbone.model.layer1.5.3.bn1.weight,backbone.model.layer1.5.3.bn2.bias,backbone.model.layer1.5.3.bn2.running_mean,backbone.model.layer1.5.3.bn2.running_var,backbone.model.layer1.5.3.bn2.weight,backbone.model.layer1.5.3.conv1.weight,backbone.model.layer1.5.3.conv2.weight,backbone.model.layer2.0.0.bn1.bias,backbone.model.layer2.0.0.bn1.running_mean,backbone.model.layer2.0.0.bn1.running_var,backbone.model.layer2.0.0.bn1.weight,backbone.model.layer2.0.0.bn2.bias,backbone.model.layer2.0.0.bn2.running_mean,backbone.model.layer2.0.0.bn2.running_var,backbone.model.layer2.0.0.bn2.weight,backbone.model.layer2.0.0.conv1.weight,backbone.model.layer2.0.0.conv2.weight,backbone.model.layer2.0.0.downsample.0.weight,backbone.model.layer2.0.0.downsample.1.bias,backbone.model.layer2.0.0.downsample.1.running_mean,backbone.model.layer2.0.0.downsample.1.running_var,backbone.model.layer2.0.0.downsample.1.weight,backbone.model.layer2.0.1.bn1.bias,backbone.model.layer2.0.1.bn1.running_mean,backbone.model.layer2.0.1.bn1.running_var,backbone.model.layer2.0.1.bn1.weight,backbone.model.layer2.0.1.bn2.bias,backbone.model.layer2.0.1.bn2.running_mean,backbone.model.layer2.0.1.bn2.running_var,backbone.model.layer2.0.1.bn2.weight,backbone.model.layer2.0.1.conv1.weight,backbone.model.layer2.0.1.conv2.weight,backbone.model.layer2.0.2.bn1.bias,backbone.model.layer2.0.2.bn1.running_mean,backbone.model.layer2.0.2.bn1.running_var,backbone.model.layer2.0.2.bn1.weight,backbone.model.layer2.0.2.bn2.bias,backbone.model.layer2.0.2.bn2.running_mean,backbone.model.layer2.0.2.bn2.running_var,backbone.model.layer2.0.2.bn2.weight,backbone.model.layer2.0.2.conv1.weight,backbone.model.layer2.0.2.conv2.weight,backbone.model.layer2.0.3.bn1.bias,backbone.model.layer2.0.3.bn1.running_mean,backbone.model.layer2.0.3.bn1.running_var,backbone.model.layer2.0.3.bn1.weight,backbone.model.layer2.0.3.bn2.bias,backbone.model.layer2.0.3.bn2.running_mean,backbone.model.layer2.0.3.bn2.running_var,backbone.model.layer2.0.3.bn2.weight,backbone.model.layer2.0.3.conv1.weight,backbone.model.layer2.0.3.conv2.weight,backbone.model.layer2.0.4.bn1.bias,backbone.model.layer2.0.4.bn1.running_mean,backbone.model.layer2.0.4.bn1.running_var,backbone.model.layer2.0.4.bn1.weight,backbone.model.layer2.0.4.bn2.bias,backbone.model.layer2.0.4.bn2.running_mean,backbone.model.layer2.0.4.bn2.running_var,backbone.model.layer2.0.4.bn2.weight,backbone.model.layer2.0.4.conv1.weight,backbone.model.layer2.0.4.conv2.weight,backbone.model.layer2.0.5.bn1.bias,backbone.model.layer2.0.5.bn1.running_mean,backbone.model.layer2.0.5.bn1.running_var,backbone.model.layer2.0.5.bn1.weight,backbone.model.layer2.0.5.bn2.bias,backbone.model.layer2.0.5.bn2.running_mean,backbone.model.layer2.0.5.bn2.running_var,backbone.model.layer2.0.5.bn2.weight,backbone.model.layer2.0.5.conv1.weight,backbone.model.layer2.0.5.conv2.weight