contextual&seaco ONNX export (#1481)

Shi Xian

2024-03-13 e04489ce4c0fd0095d0c79ef8f504f425e0435a8

 funasr/models/sond/encoder/fsmn_encoder.py

@@ -195,140 +195,3 @@

        return inputs, ilens, None

    def gen_tf2torch_map_dict(self):
        tensor_name_prefix_torch = self.tf2torch_tensor_name_prefix_torch
        tensor_name_prefix_tf = self.tf2torch_tensor_name_prefix_tf
        map_dict_local = {
            # torch: conv1d.weight in "out_channel in_channel kernel_size"
            # tf   : conv1d.weight in "kernel_size in_channel out_channel"
            # torch: linear.weight in "out_channel in_channel"
            # tf   :  dense.weight in "in_channel out_channel"
            # for fsmn_layers
            "{}.fsmn_layers.layeridx.ffn.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/ffn/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.fsmn_layers.layeridx.ffn.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/ffn/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.fsmn_layers.layeridx.ffn.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/ffn/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.fsmn_layers.layeridx.ffn.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/ffn/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": (2, 1, 0),
                 },
            "{}.fsmn_layers.layeridx.ffn.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/ffn/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": (2, 1, 0),
                 },
            "{}.fsmn_layers.layeridx.memory.fsmn_block.weight".format(tensor_name_prefix_torch):
                {"name": "{}/fsmn_layer_layeridx/memory/depth_conv_w".format(tensor_name_prefix_tf),
                 "squeeze": 0,
                 "transpose": (1, 2, 0),
                 },  # (1, 31, 512, 1) -> (31, 512, 1) -> (512, 1, 31)

            # for dnn_layers
            "{}.dnn_layers.layeridx.norm.bias".format(tensor_name_prefix_torch):
                {"name": "{}/dnn_layer_layeridx/LayerNorm/beta".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.dnn_layers.layeridx.norm.weight".format(tensor_name_prefix_torch):
                {"name": "{}/dnn_layer_layeridx/LayerNorm/gamma".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.dnn_layers.layeridx.w_1.bias".format(tensor_name_prefix_torch):
                {"name": "{}/dnn_layer_layeridx/conv1d/bias".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": None,
                 },
            "{}.dnn_layers.layeridx.w_1.weight".format(tensor_name_prefix_torch):
                {"name": "{}/dnn_layer_layeridx/conv1d/kernel".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": (2, 1, 0),
                 },
            "{}.dnn_layers.layeridx.w_2.weight".format(tensor_name_prefix_torch):
                {"name": "{}/dnn_layer_layeridx/conv1d_1/kernel".format(tensor_name_prefix_tf),
                 "squeeze": None,
                 "transpose": (2, 1, 0),
                 },

        }
        if self.out_units is not None:
            # add output layer
            map_dict_local.update({
                "{}.conv1d.weight".format(tensor_name_prefix_torch):
                    {"name": "{}/conv1d/kernel".format(tensor_name_prefix_tf),
                     "squeeze": None,
                     "transpose": (2, 1, 0),
                     },
                "{}.conv1d.bias".format(tensor_name_prefix_torch):
                    {"name": "{}/conv1d/bias".format(tensor_name_prefix_tf),
                     "squeeze": None,
                     "transpose": None,
                     },
            })

        return map_dict_local

    def convert_tf2torch(self,
                         var_dict_tf,
                         var_dict_torch,
                         ):

        map_dict = self.gen_tf2torch_map_dict()

        var_dict_torch_update = dict()
        for name in sorted(var_dict_torch.keys(), reverse=False):
            if name.startswith(self.tf2torch_tensor_name_prefix_torch):
                # process special (first and last) layers
                if name in map_dict:
                    name_tf = map_dict[name]["name"]
                    data_tf = var_dict_tf[name_tf]
                    if map_dict[name]["squeeze"] is not None:
                        data_tf = np.squeeze(data_tf, axis=map_dict[name]["squeeze"])
                    if map_dict[name]["transpose"] is not None:
                        data_tf = np.transpose(data_tf, map_dict[name]["transpose"])
                    data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                    assert var_dict_torch[name].size() == data_tf.size(), \
                        "{}, {}, {} != {}".format(name, name_tf,
                                                  var_dict_torch[name].size(), data_tf.size())
                    var_dict_torch_update[name] = data_tf
                    logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
                        name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
                    ))
                # process general layers
                else:
                    # self.tf2torch_tensor_name_prefix_torch may include ".", solve this case
                    names = name.replace(self.tf2torch_tensor_name_prefix_torch, "todo").split('.')
                    layeridx = int(names[2])
                    name_q = name.replace(".{}.".format(layeridx), ".layeridx.")
                    if name_q in map_dict.keys():
                        name_v = map_dict[name_q]["name"]
                        name_tf = name_v.replace("layeridx", "{}".format(layeridx))
                        data_tf = var_dict_tf[name_tf]
                        if map_dict[name_q]["squeeze"] is not None:
                            data_tf = np.squeeze(data_tf, axis=map_dict[name_q]["squeeze"])
                        if map_dict[name_q]["transpose"] is not None:
                            data_tf = np.transpose(data_tf, map_dict[name_q]["transpose"])
                        data_tf = torch.from_numpy(data_tf).type(torch.float32).to("cpu")
                        assert var_dict_torch[name].size() == data_tf.size(), \
                            "{}, {}, {} != {}".format(name, name_tf,
                                                      var_dict_torch[name].size(), data_tf.size())
                        var_dict_torch_update[name] = data_tf
                        logging.info("torch tensor: {}, {}, loading from tf tensor: {}, {}".format(
                            name, data_tf.size(), name_tf, var_dict_tf[name_tf].shape
                        ))
                    else:
                        logging.warning("{} is missed from tf checkpoint".format(name))

        return var_dict_torch_update