Correction for the is_numpy cases for mix of numpy and non numpy inputs

Author: apbose

py/torch_tensorrt/dynamo/conversion/impl/select.py

@@ -88,9 +88,16 @@ def index(
     # _LOGGER.debug(f"The index shape is {index.shape}")
     # check if the input is dynamic
     dynamic_shape = has_dynamic_shape(input.shape)
-    #is_numpy is a flag to specify if input isa numpy
-    is_numpy = False
-
+    # is_numpy is a flag to specify if all the indices are numpy or torchTensor.
+    # If any is not this flag will be set to False
+    is_numpy = True
+    _LOGGER.debug(f"Checking for the is_numpy flag")
+    for i, ind in enumerate(index):
+        if ind is None:
+            continue
+        if not (isinstance(ind, torch.Tensor) or isinstance(ind, np.ndarray)):
+            is_numpy = False
+            break
     # here we need to check if all the index are broadcastable
     # if no, then we need to broadcast
     last_index = None
@@ -101,7 +108,7 @@ def index(
             # torch.nn.parameter.Parameter=> numpy array
             # numpy array is kept as numpy
             # other cases are kept as TRTTensor
-            if (isinstance(ind, torch.Tensor) or (ind, np.ndarray)):
+            if is_numpy:
                 ind = to_numpy(ind)
                 is_numpy = True
             else:
@@ -119,8 +126,9 @@ def index(
         set_layer_name(identity_layer, target, name + "_index_identity", source_ir)
         return identity_layer.get_output(0)
     elif len(tensor_indices) == 1:
-        # This case works
-        indices_tensor = tensor_indices[0]
+        indices_tensor = get_trt_tensor(
+            ctx, tensor_indices[0], name + f"_parameter_to_fp32_tensor"
+        )
         index = adv_indx_indices[0]
         _LOGGER.debug(f"The advanced index indices is {adv_indx_indices}")
         gather_layer = ctx.net.add_gather(input, indices_tensor, index)
@@ -136,15 +144,15 @@ def index(
         rank = len(input_shape)
         adv_indx_count = len(adv_indx_indices)
         dim_tensor_list = []
+        dim_list = []
 
         for i in range(rank):
             dim = input_shape[i]
             dim_tensor = get_trt_tensor(ctx, dim, name + f"_individual_dim_{i}")
             # dim_tensor_list is a list of tensors or numpy
-            if(is_numpy):
-                dim_tensor_list.append(dim)
-            else:
-                dim_tensor_list.append(dim_tensor)
+            if is_numpy:
+                dim_list.append(dim)
+            dim_tensor_list.append(dim_tensor)
 
         # for cases like
         # t: [x_1, y_1, y_2, ..., x_m, ..., y_n] -> t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n],
@@ -164,12 +172,9 @@ def index(
         _LOGGER.debug(f"The new transpose order is {new_order}")
 
         transpose_tensor = None
-        if(is_numpy):
-            transpose_tensor = input[new_order]
-        else:
-            transpose_layer.second_transpose = tuple(new_order)
-            set_layer_name(transpose_layer, target, name + "_index_transpose", source_ir)
-            transpose_tensor = transpose_layer.get_output(0)
+        transpose_layer.second_transpose = tuple(new_order)
+        set_layer_name(transpose_layer, target, name + "_index_transpose", source_ir)
+        transpose_tensor = transpose_layer.get_output(0)
 
         # Flatten [x_1, x_2,.......x_m, y_1, y_2,.....y_n]
         # transpose_tensor_shape = ctx.net.add_shape(transpose_tensor)
@@ -182,36 +187,34 @@ def index(
         for i in range(adv_indx_count, rank):
             mult_d1 = mult_d1 * transpose_tensor_shape[i]
 
-        flatten_tensor = None
-        if(is_numpy):
-            flatten_tensor = transpose_tensor.reshape(mult_d0, mult_d1)
-        else:
-            concat_tensor_layer = ctx.net.add_concatenation(
-                [
-                    get_trt_tensor(ctx, mult_d0, name + "_d0_shape"),
-                    get_trt_tensor(ctx, mult_d1, name + "_d1_shape"),
-                ]
-            )
-            set_layer_name(concat_tensor_layer, target, name + "_index_Concat", source_ir)
-            concat_tensor = concat_tensor_layer.get_output(0)
+        concat_tensor_layer = ctx.net.add_concatenation(
+            [
+                get_trt_tensor(ctx, mult_d0, name + "_d0_shape"),
+                get_trt_tensor(ctx, mult_d1, name + "_d1_shape"),
+            ]
+        )
+        set_layer_name(concat_tensor_layer, target, name + "_index_Concat", source_ir)
+        concat_tensor = concat_tensor_layer.get_output(0)
 
-            reshape_layer = ctx.net.add_shuffle(transpose_tensor)
-            reshape_layer.set_input(1, concat_tensor)
-            flatten_tensor = reshape_layer.get_output(0)
+        reshape_layer = ctx.net.add_shuffle(transpose_tensor)
+        reshape_layer.set_input(1, concat_tensor)
+        flatten_tensor = reshape_layer.get_output(0)
 
         _LOGGER.debug(f"The flatten tensor shape is {flatten_tensor.shape}")
 
         # tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)),  ind_i is input indices[i], x_j is the
         # // j dimension of input x.
-        if(is_numpy):
-            multiplier = dim_tensor_list[adv_indx_indices[adv_indx_count - 1]]
+        if is_numpy:
+            multiplier = dim_list[adv_indx_indices[adv_indx_count - 1]]
             cum_adv_index = tensor_indices[adv_indx_count - 1]
             for i in range(adv_indx_count - 2, -1, -1):
                 adv_index = multiplier * tensor_indices[i]
                 cum_adv_index = cum_adv_index + adv_index
-                multiplier = multiplier * dim_tensor_list[adv_indx_indices[i]]
+                multiplier = multiplier * dim_list[adv_indx_indices[i]]
+            cum_adv_index = get_trt_tensor(
+                ctx, cum_adv_index, name + f"_index_sum_intermediate"
+            )
         else:
-
             multiplier = get_trt_tensor(
                 ctx,
                 dim_tensor_list[adv_indx_indices[adv_indx_count - 1]],
@@ -269,36 +272,31 @@ def index(
             == adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1
         ):
             _LOGGER.debug(f"The indices are continuous in this case")
-            if(is_numpy):
-                concat_tensor_reshape.append(-1)
-            else:
-                concat_tensor_reshape.append(
-                    get_trt_tensor(ctx, -1, name + "_dynamic_concat")
-                )
+            concat_tensor_reshape.append(
+                get_trt_tensor(ctx, -1, name + "_dynamic_concat")
+            )
             for i in range(0, rank):
                 if i not in adv_indx_indices:
                     curr_dim = dim_tensor_list[i]
                     concat_tensor_reshape.append(curr_dim)
 
             unfold_tensor = None
-            if(is_numpy):
-                unfold_tensor = gather_out.reshape(concat_tensor)
-            else:
-                concat_tensor_layer = ctx.net.add_concatenation(concat_tensor_reshape)
-                set_layer_name(
-                    concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir
-                )
-                concat_tensor = concat_tensor_layer.get_output(0)
 
-                regular_index_shuffle_layer = ctx.net.add_shuffle(gather_out)
-                regular_index_shuffle_layer.set_input(1, concat_tensor)
-                set_layer_name(
-                    regular_index_shuffle_layer,
-                    target,
-                    name + "_index_regular_index",
-                    source_ir,
-                )
-                unfold_tensor = regular_index_shuffle_layer.get_output(0)
+            concat_tensor_layer = ctx.net.add_concatenation(concat_tensor_reshape)
+            set_layer_name(
+                concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir
+            )
+            concat_tensor = concat_tensor_layer.get_output(0)
+
+            regular_index_shuffle_layer = ctx.net.add_shuffle(gather_out)
+            regular_index_shuffle_layer.set_input(1, concat_tensor)
+            set_layer_name(
+                regular_index_shuffle_layer,
+                target,
+                name + "_index_regular_index",
+                source_ir,
+            )
+            unfold_tensor = regular_index_shuffle_layer.get_output(0)
             _LOGGER.debug(f"The tensor is unfolded now")
             _LOGGER.debug(f"The unfolded tensor shape is {unfold_tensor.shape}")
 
@@ -313,17 +311,15 @@ def index(
             _LOGGER.debug(f"Transposing the indices to correct position {new_order}")
 
             transpose_tensor = None
-            if(is_numpy):
-                transpose_tensor = unfold_tensor[new_order]
-            else:
-                transpose_advanced_shuffle_layer.second_transpose = tuple(new_order)
-                set_layer_name(
-                    transpose_advanced_shuffle_layer,
-                    target,
-                    name + "_index_advanced_shuffle_transpose",
-                    source_ir,
-                )
-                transpose_tensor = transpose_advanced_shuffle_layer.get_output(0)
+
+            transpose_advanced_shuffle_layer.second_transpose = tuple(new_order)
+            set_layer_name(
+                transpose_advanced_shuffle_layer,
+                target,
+                name + "_index_advanced_shuffle_transpose",
+                source_ir,
+            )
+            transpose_tensor = transpose_advanced_shuffle_layer.get_output(0)
 
             # unfold advanced layer
             concat_final_tensor = []
@@ -338,28 +334,25 @@ def index(
                     concat_final_tensor.append(current_dim)
 
             reshape_output = []
-            if(is_numpy):
-                reshape_output = transpose_tensor.reshape(concat_final_tensor)
-            else:
-                concat_final_shape_layer = ctx.net.add_concatenation(concat_final_tensor)
-                set_layer_name(
-                    concat_final_shape_layer,
-                    target,
-                    name + "_index_continuous_concat_final_shape_layer",
-                    source_ir,
-                )
-                concat_final_tensor = concat_final_shape_layer.get_output(0)
-
-                unfold_advanced_shuffle_layer = ctx.net.add_shuffle(transpose_tensor)
-                # check this
-                unfold_advanced_shuffle_layer.set_input(1, concat_final_tensor)
-                    set_layer_name(
-                    unfold_advanced_shuffle_layer,
-                    target,
-                    name + "_unfold_advanced_index",
-                    source_ir,
-                )
-                reshape_output = unfold_advanced_shuffle_layer.get_output(0)
+            concat_final_shape_layer = ctx.net.add_concatenation(concat_final_tensor)
+            set_layer_name(
+                concat_final_shape_layer,
+                target,
+                name + "_index_continuous_concat_final_shape_layer",
+                source_ir,
+            )
+            concat_final_tensor = concat_final_shape_layer.get_output(0)
+
+            unfold_advanced_shuffle_layer = ctx.net.add_shuffle(transpose_tensor)
+            # check this
+            unfold_advanced_shuffle_layer.set_input(1, concat_final_tensor)
+            set_layer_name(
+                unfold_advanced_shuffle_layer,
+                target,
+                name + "_unfold_advanced_index",
+                source_ir,
+            )
+            reshape_output = unfold_advanced_shuffle_layer.get_output(0)
 
         else:
             _LOGGER.debug(f"The indices are not continuous in this case")
@@ -371,26 +364,24 @@ def index(
                     concat_final_tensor.append(curr_dim)
 
             reshape_output = None
-            if(is_numpy):
-                reshape_output = gather_out.reshape(concat_final_tensor)
-            else:
-                concat_final_shape_layer = ctx.net.add_concatenation(concat_final_tensor)
-                set_layer_name(
-                    concat_final_shape_layer,
-                    target,
-                    name + "_index_non_continuous_concat_final_shape_layer",
-                    source_ir,
-                )
-                concat_final_tensor = concat_final_shape_layer.get_output(0)
 
-                reshape_layer = ctx.net.add_shuffle(gather_out)
-                reshape_layer.set_input(1, concat_final_tensor)
-                set_layer_name(
-                    reshape_layer,
-                    target,
-                    name + "_index_non_continuous_shuffle_final_shape_layer",
-                    source_ir,
-                )
-                reshape_output = reshape_layer.get_output(0)
+            concat_final_shape_layer = ctx.net.add_concatenation(concat_final_tensor)
+            set_layer_name(
+                concat_final_shape_layer,
+                target,
+                name + "_index_non_continuous_concat_final_shape_layer",
+                source_ir,
+            )
+            concat_final_tensor = concat_final_shape_layer.get_output(0)
+
+            reshape_layer = ctx.net.add_shuffle(gather_out)
+            reshape_layer.set_input(1, concat_final_tensor)
+            set_layer_name(
+                reshape_layer,
+                target,
+                name + "_index_non_continuous_shuffle_final_shape_layer",
+                source_ir,
+            )
+            reshape_output = reshape_layer.get_output(0)
 
     return reshape_output