models/LipNetPlus.py

import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F
from torch.nn import Transformer

import math
import numpy as np

from helpers import *
from torch import Tensor
from models.PhonemeTransformer import (
    PositionalEncoding, TokenEmbedding
)


class LipNetPlus(torch.nn.Module):
    def __init__(
        self, output_classes, dropout_p=0.0, pre_gru_repeats=0,
        gru_output_size=512, embeds_size=256,
        output_vocab_size=512, dropout_t=0.1,
        src_vocab_size=4, num_encoder_layers: int = 3,
        num_decoder_layers: int = 3, nhead: int = 8,
        dim_feedforward: int = 512,
    ):
        super(LipNetPlus, self).__init__()
        assert gru_output_size % 2 == 0
        self.pre_gru_repeats = pre_gru_repeats
        self.gru_out_size = gru_output_size
        self.gru_hidden_size = gru_output_size // 2
        self.embeds_size = embeds_size

        self.output_vocab_size = output_vocab_size
        self.gru_output_size = gru_output_size
        self.dropout_t = dropout_t

        self.conv1 = nn.Conv3d(3, 32, (3, 5, 5), (1, 2, 2), (1, 2, 2))
        self.pool1 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.conv2 = nn.Conv3d(32, 64, (3, 5, 5), (1, 1, 1), (1, 2, 2))
        self.pool2 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.conv3 = nn.Conv3d(64, 96, (3, 3, 3), (1, 1, 1), (1, 1, 1))
        self.pool3 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))

        self.gru1 = nn.GRU(
            96 * 4 * 8, self.gru_hidden_size, 1, bidirectional=True
        )
        self.gru2 = nn.GRU(
            self.gru_output_size, self.gru_hidden_size, 1, bidirectional=True
        )

        self.output_classes = output_classes
        self.FC = nn.Linear(self.gru_output_size, output_classes + 1)
        self.dropout_p = dropout_p

        self.relu = nn.ReLU(inplace=True)
        self.dropout = nn.Dropout(self.dropout_p)
        self.dropout3d = nn.Dropout3d(self.dropout_p)

        self.src_tok_emb = TokenEmbedding(
            src_vocab_size, self.embeds_size
        )
        self.tgt_tok_emb = TokenEmbedding(
            output_vocab_size, self.embeds_size
        )

        self.embeds_layer = nn.Linear(
            self.gru_output_size, self.embeds_size
        )
        self.transformer = Transformer(
            d_model=self.embeds_size, nhead=nhead,
            num_encoder_layers=num_encoder_layers,
            num_decoder_layers=num_decoder_layers,
            dim_feedforward=dim_feedforward,
            dropout=dropout_t
        )
        self.positional_encoding = PositionalEncoding(
            self.embeds_size, dropout=self.dropout_t
        )
        self.generator = nn.Linear(
            self.embeds_size, self.output_vocab_size
        )
        self._init()

    def _init(self):
        init.kaiming_normal_(self.conv1.weight, nonlinearity='relu')
        init.constant_(self.conv1.bias, 0)

        init.kaiming_normal_(self.conv2.weight, nonlinearity='relu')
        init.constant_(self.conv2.bias, 0)

        init.kaiming_normal_(self.conv3.weight, nonlinearity='relu')
        init.constant_(self.conv3.bias, 0)

        init.kaiming_normal_(self.FC.weight, nonlinearity='sigmoid')
        init.constant_(self.FC.bias, 0)

        transformer_components = [
            self.transformer, self.generator,
            self.positional_encoding
        ]

        for component in transformer_components:
            for p in component.parameters():
                if p.dim() > 1:
                    nn.init.xavier_uniform_(p)

        for m in (self.gru1, self.gru2):
            stdv = math.sqrt(2 / (96 * 3 * 6 + 256))

            for i in range(0, 256 * 3, 256):
                init.uniform_(m.weight_ih_l0[i: i + 256],
                              -math.sqrt(3) * stdv, math.sqrt(3) * stdv)
                init.orthogonal_(m.weight_hh_l0[i: i + 256])
                init.constant_(m.bias_ih_l0[i: i + 256], 0)
                init.uniform_(m.weight_ih_l0_reverse[i: i + 256],
                              -math.sqrt(3) * stdv, math.sqrt(3) * stdv)
                init.orthogonal_(m.weight_hh_l0_reverse[i: i + 256])
                init.constant_(m.bias_ih_l0_reverse[i: i + 256], 0)

    def forward_gru(self, x):
        x = self.conv1(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool1(x)

        x = self.conv2(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool2(x)

        x = self.conv3(x)
        x = self.relu(x)
        x = self.dropout3d(x)
        x = self.pool3(x)

        # (B, C, T, H, W)->(T, B, C, H, W)
        x = x.permute(2, 0, 1, 3, 4).contiguous()
        # (B, C, T, H, W)->(T, B, C*H*W)
        x = x.view(x.size(0), x.size(1), -1)

        self.gru1.flatten_parameters()
        self.gru2.flatten_parameters()

        if self.pre_gru_repeats > 1:
            x = torch.repeat_interleave(
                x, dim=0, repeats=self.pre_gru_repeats
            )

        x, h = self.gru1(x)
        x = self.dropout(x)
        x, h = self.gru2(x)
        x = self.dropout(x)
        return x

    def predict_from_gru_out(self, x):
        x = self.FC(x)
        x = x.permute(1, 0, 2).contiguous()
        # assert not contains_nan_or_inf(x19)
        return x

    def forward(self, x):
        x = self.forward_gru(x)
        x = self.predict_from_gru_out(x)
        return x

    def make_src_embeds(self, x):
        x = self.embeds_layer(x)
        x = self.relu(x)
        return x

    def seq_forward(
        self, src_embeds: Tensor, trg: Tensor,
        src_mask: Tensor, tgt_mask: Tensor, src_padding_mask: Tensor,
        tgt_padding_mask: Tensor, memory_key_padding_mask: Tensor
    ):
        src_emb = self.positional_encoding(src_embeds)
        tgt_emb = self.positional_encoding(self.tgt_tok_emb(trg))

        outs = self.transformer(
            src_emb, tgt_emb, src_mask, tgt_mask, None,
            src_padding_mask, tgt_padding_mask, memory_key_padding_mask
        )
        return self.generator(outs)