fixes act_order

Xilinx · Nov 15, 2023 · a4ea065 · a4ea065
1 parent f4c99e5
commit a4ea065
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 7 deletions.
diff --git a/src/brevitas/graph/gpfq.py b/src/brevitas/graph/gpfq.py
@@ -234,21 +234,22 @@ def single_layer_update(self):
         for t in range(weight.shape[-1]):
             for group_index in range(self.groups):
                 U[group_index] += torch.matmul(
-                    weight[group_index, :, t].unsqueeze(1),
+                    weight[group_index, :, perm[t]].unsqueeze(1),
                     self.float_input[group_index, :,
-                                     t].unsqueeze(0))  #[OC/Groups, 1] * [1, INSHAPE[1]]
-                norm = torch.linalg.norm(self.quantized_input[group_index, :, t], 2) ** 2
+                                     perm[t]].unsqueeze(0))  #[OC/Groups, 1] * [1, INSHAPE[1]]
+                norm = torch.linalg.norm(self.quantized_input[group_index, :, perm[t]], 2) ** 2
                 if norm > 0:
-                    q_arg = U[group_index].matmul(self.quantized_input[group_index, :, t]) / norm
+                    q_arg = U[group_index].matmul(
+                        self.quantized_input[group_index, :, perm[t]]) / norm
                 else:
                     q_arg = torch.zeros_like(U[group_index, :, 0])
 
-                weight[group_index, :, t] = q_arg
+                weight[group_index, :, perm[t]] = q_arg
             q = self.get_quant_weights(t, 0, permutation_list)
             for group_index in range(self.groups):
                 U[group_index] -= torch.matmul(
                     q[group_index].unsqueeze(1),
-                    self.quantized_input[group_index, :, t].unsqueeze(0))
+                    self.quantized_input[group_index, :, perm[t]].unsqueeze(0))
 
         del self.float_input
         del self.quantized_input
diff --git a/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py b/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py
@@ -207,7 +207,7 @@
     default=3,
     type=int,
     help='Exponent bit width used with float quantization for activations (default: 3)')
-add_bool_arg(parser, 'gptq', default=True, help='GPTQ (default: enabled)')
+add_bool_arg(parser, 'gptq', default=False, help='GPTQ (default: disabled)')
 add_bool_arg(parser, 'gpfq', default=False, help='GPFQ (default: disabled)')
 add_bool_arg(
     parser, 'gptq-act-order', default=False, help='GPTQ Act order heuristic (default: disabled)')