# 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 script has been adapted from some of the original EspNet found here:
[
    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/decoder_layer.py
]
Main change:
    replace the BatchNorm to LayerNorm
"""

import torch
from torch import nn
from src.layers.layer_norm import LayerNorm
from src.utils.remap_tensor import remap_tensor


class DecoderLayer(nn.Module):
    """Single decoder layer module.

    Args:
        size (int): Input dimension.
        self_attn (torch.nn.Module): Self-attention module instance.
            `MultiHeadedAttention` instance can be used as the argument.
        src_attn (torch.nn.Module): Self-attention module instance.
            `MultiHeadedAttention` instance can be used as the argument.
        feed_forward (torch.nn.Module): Feed-forward module instance.
            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
            can be used as the argument.
        dropout_rate (float): Dropout rate.
        normalize_before (bool): Whether to use layer_norm before the first block.
        concat_after (bool): Whether to concat attention layer's input and output.
            if True, additional linear will be applied.
            i.e. x -> x + linear(concat(x, att(x)))
            if False, no additional linear will be applied. i.e. x -> x + att(x)


    """

    def __init__(
        self,
        size,
        self_attn,
        src_attn,
        feed_forward,
        dropout_rate,
        normalize_before=True,
        concat_after=False,
    ):
        """Construct an DecoderLayer object."""
        super(DecoderLayer, self).__init__()
        self.size = size
        self.self_attn = self_attn
        self.src_attn = src_attn
        self.feed_forward = feed_forward
        self.norm1 = LayerNorm(size)
        self.norm2 = LayerNorm(size)
        self.norm3 = LayerNorm(size)
        self.dropout = nn.Dropout(dropout_rate)
        self.normalize_before = normalize_before
        self.concat_after = concat_after
        if self.concat_after:
            self.concat_linear1 = nn.Linear(size + size, size)
            self.concat_linear2 = nn.Linear(size + size, size)

    def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None):
        """Compute decoded features.

        Args:
            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
            tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
            memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
            memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
            cache (List[torch.Tensor]): List of cached tensors.
                Each tensor shape should be (#batch, maxlen_out - 1, size).

        Returns:
            torch.Tensor: Output tensor(#batch, maxlen_out, size).
            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
            torch.Tensor: Encoded memory mask (#batch, maxlen_in).

        """
        residual = tgt
        if self.normalize_before:
            tgt = self.norm1(tgt)

        if cache is None:
            tgt_q = tgt
            tgt_q_mask = tgt_mask
        else:
            assert cache.shape == (
                tgt.shape[0],
                tgt.shape[1] - 1,
                self.size,
            ), f"{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
            tgt_q = tgt[:, -1:, :]
            residual = residual[:, -1:, :]
            tgt_q_mask = None
            if tgt_mask is not None:
                tgt_q_mask = tgt_mask[:, -1:, :]

        if self.concat_after:
            tgt_concat = torch.cat((tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)), dim=-1)
            x = residual + self.concat_linear1(tgt_concat)
        else:
            x = residual + self.dropout(self.self_attn(tgt_q, tgt, tgt, tgt_q_mask))
        if self.training:
            x = remap_tensor(x)  # fwd_grain_size=16, also can set to 8, 32
        if not self.normalize_before:
            x = self.norm1(x)

        residual = x
        if self.normalize_before:
            x = self.norm2(x)
        if self.concat_after:
            x_concat = torch.cat((x, self.src_attn(x, memory, memory, memory_mask)), dim=-1)
            x = residual + self.concat_linear2(x_concat)
        else:
            x = residual + self.dropout(self.src_attn(x, memory, memory, memory_mask))
        if not self.normalize_before:
            x = self.norm2(x)

        residual = x
        if self.normalize_before:
            x = self.norm3(x)
        x = residual + self.dropout(self.feed_forward(x))
        if not self.normalize_before:
            x = self.norm3(x)

        if cache is not None:
            x = torch.cat([cache, x], dim=1)
        if self.training:
            x = remap_tensor(x)  # fwd_grain_size=16, also can set to 8, 32
        return x, tgt_mask, memory, memory_mask