# Copyright (c) 2021 Graphcore Ltd. All rights reserved.
# Copyright 2019 Shigeki Karita
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# This file has been modified by Graphcore Ltd.
"""
This script has been adapted from some of the original EspNet found here:
[
    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
]
Main change:
    remove the ScaledPositionalEncoding class and LegacyRelPositionalEncoding class.
"""

import math
import torch


def _pre_hook(
    state_dict,
    prefix,
    local_metadata,
    strict,
    missing_keys,
    unexpected_keys,
    error_msgs,
):
    """Perform pre-hook in load_state_dict for backward compatibility.

    Note:
        We saved self.pe until v.0.5.2 but we have omitted it later.
        Therefore, we remove the item "pe" from `state_dict` for backward compatibility.

    """
    k = prefix + "pe"
    if k in state_dict:
        state_dict.pop(k)


class PositionalEncoding(torch.nn.Module):
    """Positional encoding.

    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_len (int): Maximum input length.
        reverse (bool): Whether to reverse the input position. Only for
        the class LegacyRelPositionalEncoding. We remove it in the current
        class RelPositionalEncoding.

    """

    def __init__(self, d_model, dropout_rate, max_len=50, reverse=False):  # /TODO
        """Construct an PositionalEncoding object."""
        super(PositionalEncoding, self).__init__()
        self.d_model = d_model
        self.reverse = reverse
        self.xscale = math.sqrt(self.d_model)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        self.pe = None
        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
        self._register_load_state_dict_pre_hook(_pre_hook)

    def extend_pe(self, x):
        """Reset the positional encodings."""
        if self.pe is not None:
            if self.pe.size(1) >= x.size(1):
                if self.pe.dtype != x.dtype or self.pe.device != x.device:
                    self.pe = self.pe
                return
        pe = torch.zeros(x.size(1), self.d_model)
        if self.reverse:
            position = torch.arange(x.size(1) - 1, -1, -1.0, dtype=torch.float32).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32) * -(math.log(10000.0) / self.d_model)
        )
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.pe = pe

    def forward(self, x: torch.Tensor):
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).

        """
        x = x * self.xscale + self.pe.to(x.device)[:, : x.size(1)]  # [:, :x.size(1)]
        return self.dropout(x)


class RelPositionalEncoding(torch.nn.Module):
    """Relative positional encoding module (new implementation).

    Details can be found in https://github.com/espnet/espnet/pull/2816.

    See : Appendix B in https://arxiv.org/abs/1901.02860

    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_len (int): Maximum input length.

    """

    def __init__(self, d_model, dropout_rate, max_len=5000):
        """Construct an PositionalEncoding object."""
        super(RelPositionalEncoding, self).__init__()
        self.d_model = d_model
        self.xscale = math.sqrt(self.d_model)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        self.pe = None
        self.extend_pe(torch.tensor(0.0).expand(1, max_len))

    def extend_pe(self, x):
        """Reset the positional encodings."""
        self.max_len = x.size(1)

        self.pe = torch.zeros(self.max_len, self.d_model)
        position = torch.arange(0, self.max_len, dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32) * -(math.log(10000.0) / self.d_model)
        )
        self.pe[:, 0::2] = torch.sin(position * div_term)
        self.pe[:, 1::2] = torch.cos(position * div_term)
        self.pe = self.pe.unsqueeze(0)
        self.pe = self.pe[:, 0 : 0 + x.size(1)]

    def forward(self, x: torch.Tensor):
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).

        """
        x = self.dropout(x * self.xscale)
        pos_emb = self.dropout(self.pe.to(x.device))
        return x, pos_emb