assert not sparse_intermediate_bounds or use_sparse_conv is False

[yusiyoh]

Hello and thank you for this amazing tool!

I have also tried to use alpha-beta-CROWN with the same setup but had the same error in the title.

You can find the issue here: https://github.com/Verified-Intelligence/alpha-beta-CROWN/issues/28

With auto_LiRPA, CROWN works as expected but when I change to alpha-CROWN, I have the following error:

AssertionError                            Traceback (most recent call last)
[<ipython-input-34-0e7e4f8b53a2>](https://localhost:8080/#) in <cell line: 3>()
      2 print('Bounding method: backward (CROWN, DeepPoly)')
      3 with torch.no_grad():  # If gradients of the bounds are not needed, we can use no_grad to save memory.
----> 4   lb, ub = bounded_model.compute_bounds(x=(bounded_image,), method='alpha-CROWN')
      5 
      6 # Auxillary function to print bounds.

3 frames
[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/bound_general.py](https://localhost:8080/#) in compute_bounds(self, x, aux, C, method, IBP, forward, bound_lower, bound_upper, reuse_ibp, reuse_alpha, return_A, needed_A_dict, final_node_name, average_A, intermediate_layer_bounds, reference_bounds, intermediate_constr, alpha_idx, aux_reference_bounds, need_A_only, cutter, decision_thresh, update_mask)
   1186                 method = 'backward'
   1187             if bound_lower:
-> 1188                 ret1 = self.get_optimized_bounds(
   1189                     x=x, C=C, method=method,
   1190                     intermediate_layer_bounds=intermediate_layer_bounds,

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/optimized_bounds.py](https://localhost:8080/#) in get_optimized_bounds(self, x, aux, C, IBP, forward, method, bound_lower, bound_upper, reuse_ibp, return_A, average_A, final_node_name, intermediate_layer_bounds, reference_bounds, aux_reference_bounds, needed_A_dict, cutter, decision_thresh, epsilon_over_decision_thresh)
    455     if init_alpha:
    456         # TODO: this should set up aux_reference_bounds.
--> 457         self.init_slope(x, share_slopes=opts['use_shared_alpha'],
    458                    method=method, c=C, final_node_name=final_node_name)
    459 

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/optimized_bounds.py](https://localhost:8080/#) in init_slope(self, x, share_slopes, method, c, bound_lower, bound_upper, final_node_name, intermediate_layer_bounds, activation_opt_params, skip_bound_compute)
   1047             start_nodes.append(('_forward', 1, None))
   1048         if method in ['backward', 'forward+backward']:
-> 1049             start_nodes += self.get_alpha_crown_start_nodes(
   1050                 node, c=c, share_slopes=share_slopes,
   1051                 final_node_name=final_node_name)

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/backward_bound.py](https://localhost:8080/#) in get_alpha_crown_start_nodes(self, node, c, share_slopes, final_node_name)
    769                 # The spec dim is c only, and is shared among h, w.
    770                 output_shape = node.patch_size[nj.name][0]
--> 771             assert not sparse_intermediate_bounds or use_sparse_conv is False  # Double check our assumption holds. If this fails, then we created wrong shapes for alpha.
    772         else:
    773             # Output is linear layer, or patch converted to matrix.

My implementation is here: https://colab.research.google.com/drive/1b4PMeK0NKmeXV-mKCfeopbuoJYHa8W4S?usp=sharing Weights and data: model_data.zip

I do not get the cause of the error. I tried 'sparse_features_alpha': False, 'sparse_spec_alpha': False but got same results.

Another point is: How to visualize the perturbed image? How does BoundedTensor change the original image?

Thanks in advance and best regards

[shizhouxing]

@yusiyoh Maybe could you please try also setting sparse_intermediate_bounds to False for now?

I am not able to run the notebook for now -- it's showing normalizer not defined for me.

Another point is: How to visualize the perturbed image? How does BoundedTensor change the original image?

BoundedTensor annotates the original image with a perturbation. If it's an Linf perturbation, then there will be a lower bound and an upper bound for the perturbed image. Not sure what kind of visualization you are looking for?

[yusiyoh]

Sorry for inconvenience. This version should be okay: https://colab.research.google.com/drive/1b4PMeK0NKmeXV-mKCfeopbuoJYHa8W4S?usp=sharing

Changing sparse_intermediate_bounds to False yields another error (which I had while using alpha-beta-CROWN too):

/usr/local/lib/python3.10/dist-packages/auto_LiRPA/bound_general.py:970: UserWarning: Creating an identity matrix with size 8192x8192 for node BoundMaxPool(name="/input"). This may indicate poor performance for bound computation. If you see this message on a small network please submit a bug report.
  sparse_C = self.get_sparse_C(
/usr/local/lib/python3.10/dist-packages/auto_LiRPA/bound_general.py:970: UserWarning: Creating an identity matrix with size 4096x4096 for node BoundMaxPool(name="/input.4"). This may indicate poor performance for bound computation. If you see this message on a small network please submit a bug report.
  sparse_C = self.get_sparse_C(
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
[<ipython-input-21-b546ba971dce>](https://localhost:8080/#) in <cell line: 3>()
      2 print('Bounding method: backward (CROWN, DeepPoly)')
      3 with torch.no_grad():  # If gradients of the bounds are not needed, we can use no_grad to save memory.
----> 4   lb, ub = bounded_model.compute_bounds(x=(bounded_image,), method='alpha-CROWN')
      5 
      6 # Auxillary function to print bounds.

3 frames
[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/bound_general.py](https://localhost:8080/#) in compute_bounds(self, x, aux, C, method, IBP, forward, bound_lower, bound_upper, reuse_ibp, reuse_alpha, return_A, needed_A_dict, final_node_name, average_A, intermediate_layer_bounds, reference_bounds, intermediate_constr, alpha_idx, aux_reference_bounds, need_A_only, cutter, decision_thresh, update_mask)
   1186                 method = 'backward'
   1187             if bound_lower:
-> 1188                 ret1 = self.get_optimized_bounds(
   1189                     x=x, C=C, method=method,
   1190                     intermediate_layer_bounds=intermediate_layer_bounds,

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/optimized_bounds.py](https://localhost:8080/#) in get_optimized_bounds(self, x, aux, C, IBP, forward, method, bound_lower, bound_upper, reuse_ibp, return_A, average_A, final_node_name, intermediate_layer_bounds, reference_bounds, aux_reference_bounds, needed_A_dict, cutter, decision_thresh, epsilon_over_decision_thresh)
    455     if init_alpha:
    456         # TODO: this should set up aux_reference_bounds.
--> 457         self.init_slope(x, share_slopes=opts['use_shared_alpha'],
    458                    method=method, c=C, final_node_name=final_node_name)
    459 

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/optimized_bounds.py](https://localhost:8080/#) in init_slope(self, x, share_slopes, method, c, bound_lower, bound_upper, final_node_name, intermediate_layer_bounds, activation_opt_params, skip_bound_compute)
   1053             node.restore_optimized_params(activation_opt_params[node.name])
   1054         else:
-> 1055             node.init_opt_parameters(start_nodes)
   1056         init_intermediate_bounds[node.inputs[0].name] = (
   1057             [node.inputs[0].lower.detach(), node.inputs[0].upper.detach()])

[/usr/local/lib/python3.10/dist-packages/auto_LiRPA/operators/pooling.py](https://localhost:8080/#) in init_opt_parameters(self, start_nodes)
     48                 warnings.warn("MaxPool's optimization is not supported for forward mode")
     49                 continue
---> 50             self.alpha[ns] = torch.empty(
     51                 [1, size_s, self.input_shape[0], self.input_shape[1],
     52                 self.output_shape[-2], self.output_shape[-1],

TypeError: empty(): argument 'size' must be tuple of ints, but found element of type torch.Size at pos 2

So can I select an arbitrarily perturbed image between the lower and upper bound as a sample from the perturbation set? Or how can I visualize this perturbation set in general?

I am not sure if my question makes sense, but I am trying to have a sample from the perturbation set and forward it to the model to get the corresponding prediction.

[yusiyoh]

This might be helpful:

These are the start_nodes in init_opt_parameter() function, and all of them have non-integer sizes expect the last one (the output layer with 43 classes).