update rnncell synops calculation

neurobench/utils.py

@@ -249,12 +249,13 @@ def single_layer_MACs(inputs, layer, total=False):
         # NOTE: these activation functions are currently NOT included in NeuroBench
         # if no explicit states are passed to recurrent layers, then h and c are initialized to zero (pytorch convention)
         layer_bin = make_binary_copy(layer, all_ones=total)
-        # transpose from batches, timesteps, features to features, batches
-        # print(layer_bin.weight_ih.shape)
+        # layer_weight_ih is [4*hidden_size, input_size]
+        # inputs[0].transpose(0, -1) is [input_size, batch_size]
         out_ih = torch.matmul(
             layer_bin.weight_ih, inputs[0].transpose(0, -1)
         )  # accounts for i,f,g,o
         out_hh = torch.zeros_like(out_ih)
+        # out shape is 4*h, batch, for hidden feature dim h
 
         biases = 0
         bias_ih = 0
@@ -290,10 +291,16 @@ def single_layer_MACs(inputs, layer, total=False):
 
             out += biases  # biases are added here for computation of c and h which depend on correct computation of ifgo
             out[out != 0] = 1
-            # out is vector with i,f,g,o
-            ifgo = out.reshape(4, -1)  # each row is one of i,f,g,o
+
+            # out is vector with i,f,g,o, shape is 4*h, batch
+            hidden = out.shape[0] // 4
+            ifgo = out.reshape(4, hidden, -1)  # 4, h, B
             if in_states:
-                c_1 = ifgo[1, :] * inputs[1][1] + ifgo[0, :] * ifgo[2, :]
+                # inputs[1][1] shape is [B, h]
+                # element-wise multiply (vector products f*c + i*g)
+                c_1 = (
+                    ifgo[1, :] * inputs[1][1].transpose(0, -1) + ifgo[0, :] * ifgo[2, :]
+                )
             else:
                 c_1 = ifgo[0, :] * ifgo[2, :]
 
@@ -323,9 +330,13 @@ def single_layer_MACs(inputs, layer, total=False):
             # Wir*x, Whr*h, Wiz*x, Whz*h, Win*x, Whn*h
             macs += rzn.sum()  # multiplications of all weights and inputs/hidden states
             rzn += biases  # add biases
-            rzn = rzn.reshape(3, -1)  # each row is one of r,z,n
 
-            out_hh_n = out_hh.reshape(3, -1)[2, :] + bias_hh.reshape(3, -1)[2, :]
+            hidden = rzn.shape[0] // 3
+            rzn = rzn.reshape(3, hidden, -1)  # 3, h, B
+            out_hh = out_hh.reshape(3, hidden, -1)
+            bias_hh = bias_hh.reshape(3, hidden, -1)
+
+            out_hh_n = out_hh[2, :] + bias_hh[2, :]
             r = rzn[0, :]  # get r
             z = rzn[1, :]
 
@@ -334,23 +345,21 @@ def single_layer_MACs(inputs, layer, total=False):
 
             n_hh_term_macs = (
                 r * out_hh_n
-            )  # elementwise_multiplication to find macs ofr*(Whn*h + bhn) specifically
+            )  # elementwise_multiplication to find macs of r*(Whn*h + bhn) specifically
             n_hh_term_macs[n_hh_term_macs != 0] = 1
             macs += n_hh_term_macs.sum()
 
             # note hh part of n is already binarized, does not influence calculation of macs for n
-            n = (
-                out_hh.reshape(3, -1)[2, :]
-                + bias_ih.reshape(3, -1)[2, :]
-                + n_hh_term_macs
-            )
+            n = out_hh[2, :] + bias_ih[2, :] + n_hh_term_macs
             n[n != 0] = 1
             z_a = 1 - z
             # only do this now because affects z_a
             z[z != 0] = 1
             z_a[z_a != 0] = 1
             t_1 = z_a * n
-            t_2 = z * inputs[1]
+            t_2 = z * inputs[1].transpose(
+                0, -1
+            )  # inputs are shape [B, h], all else is [h, B]
 
             t_1[t_1 != 0] = 1
             t_2[t_2 != 0] = 1