from google.colab import drive

drive.mount("/content/drive")

with h5py.File("dataset_top10/16069.hdf5", "r") as fread:
    keys = list(fread.keys())
    print(keys)
    group = fread[keys[0]]
    print(group.keys())
    feature = group["feature"][:]
    token = group["token"][:]
    print(feature.shape)
    print(token)

import copy
import json
import math
import os
import random
import sys
from functools import partial
from inspect import signature
from pathlib import Path
from typing import (
    Any,
    Dict,
    List
)

# Third party's modules
import cv2

import numpy as np

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import (
    DataLoader)

from torchvision.transforms import Compose

# Local modules
sys.path.append("modules_gislr")
from modules_gislr.dataset import (
    HDF5Dataset,
    merge_padded_batch)
from modules_gislr.defines import (
    get_fullbody_landmarks
)
from modules_gislr.layers import (
    Identity,
    GPoolRecognitionHead,
    TransformerEnISLR
)
from modules_gislr.train_functions import (
    test_loop,
    val_loop,
    train_loop
)
from modules_gislr.transforms import (
    PartsBasedNormalization,
    ReplaceNan,
    SelectLandmarksAndFeature,
    ToTensor
)

def create_norm(norm_type, dim_model, eps=1e-5, add_bias=None):
    # The argument `bias` was added at v2.1.0.
    # So, we check whether LayerNorm has this.
    sig = signature(nn.LayerNorm)
    available_bias = bool("bias" in sig.parameters)
    if norm_type == "layer":
        if available_bias:
            norm = nn.LayerNorm(dim_model, eps=eps, bias=add_bias)
        else:
            norm = nn.LayerNorm(dim_model, eps=eps)
    elif norm_type == "batch":
        norm = nn.BatchNorm1d(dim_model, eps=eps)
    return norm

def apply_norm(norm_layer, feature):
    # `[N, T, C]`
    if isinstance(norm_layer, nn.LayerNorm):
        feature = norm_layer(feature)
    elif isinstance(norm_layer, nn.BatchNorm1d):
        # `[N, T, C] -> [N, C, T]`
        feature = feature.permute([0, 2, 1])
        feature = norm_layer(feature)
        # `[N, C, T] -> [N, T, C]`
        feature = feature.permute([0, 2, 1])
    return feature

class MacaronNetEncoderLayer(nn.Module):
    def __init__(self,
                 dim_model,
                 num_heads,
                 dim_ffw,
                 dropout,
                 activation,
                 norm_type_sattn,
                 norm_type_ffw,
                 norm_eps,
                 add_bias):
        super().__init__()

        self.fc_factor = 0.5

        # =====================================================================
        # First half PFFN.
        # =====================================================================
        self.norm_ffw1 = create_norm(norm_type_ffw, dim_model, norm_eps, add_bias)
        self.ffw1 = PositionwiseFeedForward(
            dim_model=dim_model,
            dim_ffw=dim_ffw,
            dropout=dropout,
            activation=activation,
            add_bias=add_bias)
        # =====================================================================
        # MHA.
        # =====================================================================
        self.norm_sattn = create_norm(norm_type_sattn, dim_model, norm_eps, add_bias)
        self.self_attn = MultiheadAttention(
            key_dim=dim_model,
            query_dim=dim_model,
            att_dim=dim_model,
            out_dim=dim_model,
            num_heads=num_heads,
            dropout=dropout,
            add_bias=add_bias)
        # =====================================================================
        # Second half PFFN.
        # =====================================================================
        self.norm_ffw2 = create_norm(norm_type_ffw, dim_model, norm_eps, add_bias)
        self.ffw2 = PositionwiseFeedForward(
            dim_model=dim_model,
            dim_ffw=dim_ffw,
            dropout=dropout,
            activation=activation,
            add_bias=add_bias)

        self.dropout = nn.Dropout(p=dropout)

        # To store attention weights.
        self.attw = None

    def forward(self,
                feature,
                causal_mask=None,
                src_key_padding_mask=None):
        bsize, qlen = feature.shape[:2]
        if src_key_padding_mask is not None:
            san_mask = make_san_mask(src_key_padding_mask, causal_mask)
        elif causal_mask is not None:
            san_mask = causal_mask
        else:
            san_mask = None

        #################################################
        # First half PFFN.
        #################################################
        # `[N, qlen, dim_model]`
        residual = feature
        feature = apply_norm(self.norm_ffw1, feature)
        feature = self.ffw1(feature)
        feature = self.fc_factor * self.dropout(feature) + residual

        #################################################
        # MHA.
        #################################################
        # `[N, qlen, dim_model]`
        residual = feature
        feature = apply_norm(self.norm_sattn, feature)
        feature, self.attw = self.self_attn(
            key=feature,
            value=feature,
            query=feature,
            mask=san_mask)
        feature = self.dropout(feature) + residual

        #################################################
        # Second half PFFN.
        #################################################
        # `[N, qlen, dim_model]`
        residual = feature
        feature = apply_norm(self.norm_ffw2, feature)
        feature = self.ffw2(feature)
        feature = self.fc_factor * self.dropout(feature) + residual

        return feature

class TransformerEncoder(nn.Module):
    def __init__(self,
                 encoder_layer,
                 num_layers,
                 dim_model,
                 dropout_pe,
                 norm_type_tail,
                 norm_eps,
                 norm_first,
                 add_bias,
                 add_tailnorm):
        super().__init__()

        self.pos_encoder = PositionalEncoding(dim_model, dropout_pe)
        self.layers = nn.ModuleList([copy.deepcopy(encoder_layer) for _ in range(num_layers)])

        # Add LayerNorm at tail position.
        # This is applied only when norm_first is True because
        # post-normalization structure includes tail-normalization in encoder
        # layers.
        if add_tailnorm and norm_first:
            self.norm_tail = create_norm(norm_type_tail, dim_model, norm_eps, add_bias)
        else:
            self.norm_tail = Identity()

    def forward(self,
                feature,
                causal_mask,
                src_key_padding_mask):
        feature = self.pos_encoder(feature)
        for layer in self.layers:
            feature = layer(feature,
                            causal_mask,
                            src_key_padding_mask)
        feature = apply_norm(self.norm_tail, feature)
        return feature

class MacaronNetEnISLR(nn.Module):
    def __init__(self,
                 in_channels,
                 inter_channels,
                 out_channels,
                 activation="relu",
                 pooling_type="none",
                 tren_num_layers=1,
                 tren_num_heads=1,
                 tren_dim_ffw=256,
                 tren_dropout_pe=0.1,
                 tren_dropout=0.1,
                 tren_norm_type_sattn="layer",
                 tren_norm_type_ffw="layer",
                 tren_norm_type_tail="layer",
                 tren_norm_eps=1e-5,
                 tren_add_bias=True,
                 tren_add_tailnorm=True):
        super().__init__()

        # Feature extraction.
        self.linear = nn.Linear(in_channels, inter_channels)
        self.activation = select_reluwise_activation(activation)

        if pooling_type == "none":
            self.pooling = Identity()
        elif pooling_type == "average":
            self.pooling = nn.AvgPool2d(
                kernel_size=[2, 1],
                stride=[2, 1],
                padding=0)
        elif pooling_type == "max":
            self.pooling = nn.MaxPool2d(
                kernel_size=[2, 1],
                stride=[2, 1],
                padding=0)

        # Transformer-Encoder with Macaron-Net architecture.
        enlayer = MacaronNetEncoderLayer(
            dim_model=inter_channels,
            num_heads=tren_num_heads,
            dim_ffw=tren_dim_ffw,
            dropout=tren_dropout,
            activation=activation,
            norm_type_sattn=tren_norm_type_sattn,
            norm_type_ffw=tren_norm_type_ffw,
            norm_eps=tren_norm_eps,
            add_bias=tren_add_bias)
        self.tr_encoder = TransformerEncoder(
            encoder_layer=enlayer,
            num_layers=tren_num_layers,
            dim_model=inter_channels,
            dropout_pe=tren_dropout_pe,
            norm_type_tail=tren_norm_type_tail,
            norm_eps=tren_norm_eps,
            norm_first=True,  # Fixed for Macaron-Net.
            add_bias=tren_add_bias,
            add_tailnorm=tren_add_tailnorm)

        self.head = GPoolRecognitionHead(inter_channels, out_channels)

    def forward(self,
                feature,
                feature_causal_mask=None,
                feature_pad_mask=None):
        # Feature extraction.
        # `[N, C, T, J] -> [N, T, C, J] -> [N, T, C*J] -> [N, T, C']`
        N, C, T, J = feature.shape
        feature = feature.permute([0, 2, 1, 3])
        feature = feature.reshape(N, T, -1)

        feature = self.linear(feature)
        if torch.isnan(feature).any():
            raise ValueError()
        feature = self.activation(feature)
        if torch.isnan(feature).any():
            raise ValueError()

        # Apply pooling.
        feature = self.pooling(feature)
        if feature_pad_mask is not None:
            # Cast to apply pooling.
            feature_pad_mask = feature_pad_mask.to(feature.dtype)
            feature_pad_mask = self.pooling(feature_pad_mask.unsqueeze(-1)).squeeze(-1)
            # Binarization.
            # This removes averaged signals with padding.
            feature_pad_mask = feature_pad_mask > 0.5
            if feature_causal_mask is not None:
                feature_causal_mask = make_causal_mask(feature_pad_mask)

        feature = self.tr_encoder(
            feature=feature,
            causal_mask=feature_causal_mask,
            src_key_padding_mask=feature_pad_mask)
        if torch.isnan(feature).any():
            raise ValueError()

        # `[N, T, C] -> [N, C, T]`
        logit = self.head(feature.permute([0, 2, 1]), feature_pad_mask)
        if torch.isnan(feature).any():
            raise ValueError()
        return logit

# Access check.
dataset_dir = Path("dataset_top10")
files = list(dataset_dir.iterdir())
dictionary = [fin for fin in files if ".json" in fin.name][0]
hdf5_files = [fin for fin in files if ".hdf5" in fin.name]

print(dictionary)
print(hdf5_files)

# Load dictionary.
with open(dictionary, "r") as fread:
    key2token = json.load(fread)

VOCAB = len(key2token)

_, use_landmarks = get_fullbody_landmarks()
use_features = ["x", "y"]
trans_select_feature = SelectLandmarksAndFeature(landmarks=use_landmarks, features=use_features)
trans_repnan = ReplaceNan()
trans_norm = PartsBasedNormalization(align_mode="framewise", scale_mode="unique")

pre_transforms = Compose([trans_select_feature,
                          trans_repnan,
                          trans_norm])
transforms = Compose([ToTensor()])

batch_size = 2
feature_shape = (len(use_features), -1, len(use_landmarks))
token_shape = (1,)
merge_fn = partial(merge_padded_batch,
                   feature_shape=feature_shape,
                   token_shape=token_shape,
                   feature_padding_val=0.0,
                   token_padding_val=0)

dataset = HDF5Dataset(hdf5_files, pre_transforms=pre_transforms, transforms=transforms)

dataloader = DataLoader(dataset, batch_size=batch_size, collate_fn=merge_fn)
try:
    data = next(iter(dataloader))
    feature_origin = data["feature"]

    print(feature_origin.shape)
except Exception as inst:
    print(inst)

# Define model.
# in_channels: J * C (130*2=260)
#   J: use_landmarks (130)
#   C: use_channels (2)
# out_channels: 10
in_channels = len(use_landmarks) * len(use_features)
inter_channels = 64
out_channels = VOCAB
activation = "relu"
pooling_type = "none"
tren_num_layers = 2
tren_num_heads = 2
tren_dim_ffw = 256
tren_dropout_pe = 0.1
tren_dropout = 0.1
tren_norm_type_sattn = "layer"
tren_norm_type_ffw = "layer"
tren_norm_type_tail = "layer"
tren_norm_eps = 1e-5
tren_norm_first = True  # For Transformer.
tren_add_bias = True
tren_add_tailnorm = True

model = MacaronNetEnISLR(in_channels=in_channels,
                         inter_channels=inter_channels,
                         out_channels=out_channels,
                         activation=activation,
                         pooling_type=pooling_type,
                         tren_num_layers=tren_num_layers,
                         tren_num_heads=tren_num_heads,
                         tren_dim_ffw=tren_dim_ffw,
                         tren_dropout_pe=tren_dropout_pe,
                         tren_dropout=tren_dropout,
                         tren_norm_type_sattn=tren_norm_type_sattn,
                         tren_norm_type_ffw=tren_norm_type_ffw,
                         tren_norm_type_tail=tren_norm_type_tail,
                         tren_norm_eps=tren_norm_eps,
                         tren_add_bias=tren_add_bias,
                         tren_add_tailnorm=tren_add_tailnorm)
print(model)

# Sanity check.
logit = model(feature_origin)
print(logit.shape)
attw0 = model.tr_encoder.layers[0].attw.detach().cpu().numpy()
attw1 = model.tr_encoder.layers[0].attw.detach().cpu().numpy()
print(attw0.shape, attw1.shape)

# Set common parameters.
batch_size = 32
load_into_ram = True
test_pid = 16069
num_workers = os.cpu_count()
print(f"Using {num_workers} cores for data loading.")
lr = 3e-4

epochs = 50
eval_every_n_epochs = 1
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using {device} for computation.")

train_hdf5files = [fin for fin in hdf5_files if str(test_pid) not in fin.name]
val_hdf5files = [fin for fin in hdf5_files if str(test_pid) in fin.name]
test_hdf5files = [fin for fin in hdf5_files if str(test_pid) in fin.name]

_, use_landmarks = get_fullbody_landmarks()
use_features = ["x", "y"]

# Build dataloaders.
train_dataset = HDF5Dataset(train_hdf5files, pre_transforms=pre_transforms,
    transforms=transforms, load_into_ram=load_into_ram)
val_dataset = HDF5Dataset(val_hdf5files, pre_transforms=pre_transforms,
    transforms=transforms, load_into_ram=load_into_ram)
test_dataset = HDF5Dataset(test_hdf5files, pre_transforms=pre_transforms,
    transforms=transforms, load_into_ram=load_into_ram)

train_dataloader = DataLoader(train_dataset, batch_size=batch_size, collate_fn=merge_fn, num_workers=num_workers, shuffle=True)
val_dataloader = DataLoader(val_dataset, batch_size=batch_size, collate_fn=merge_fn, num_workers=num_workers, shuffle=False)
test_dataloader = DataLoader(test_dataset, batch_size=1, collate_fn=merge_fn, num_workers=num_workers, shuffle=False)

model_macaron = MacaronNetEnISLR(
    in_channels=in_channels,
    inter_channels=inter_channels,
    out_channels=out_channels,
    activation=activation,
    pooling_type=pooling_type,
    tren_num_layers=tren_num_layers,
    tren_num_heads=tren_num_heads,
    tren_dim_ffw=tren_dim_ffw,
    tren_dropout_pe=tren_dropout_pe,
    tren_dropout=tren_dropout,
    tren_norm_type_sattn=tren_norm_type_sattn,
    tren_norm_type_ffw=tren_norm_type_ffw,
    tren_norm_type_tail=tren_norm_type_tail,
    tren_norm_eps=tren_norm_eps,
    tren_add_bias=tren_add_bias,
    tren_add_tailnorm=tren_add_tailnorm)
print(model_macaron)

loss_fn = nn.CrossEntropyLoss(reduction="mean")
optimizer = torch.optim.Adam(model_macaron.parameters(), lr=lr)

# Train, validation, and evaluation.
model_macaron.to(device)

train_losses = []
val_losses = []
test_accs = []
print("Start training.")
for epoch in range(epochs):
    print("-" * 80)
    print(f"Epoch {epoch+1}")

    train_losses = train_loop(train_dataloader, model_macaron, loss_fn, optimizer, device)
    val_loss = val_loop(val_dataloader, model_macaron, loss_fn, device)
    val_losses.append(val_loss)

    if (epoch+1) % eval_every_n_epochs == 0:
        acc = test_loop(test_dataloader, model_macaron, device)
        test_accs.append(acc)
train_losses_macaron = np.array(train_losses)
val_losses_macaron = np.array(val_losses)
test_accs_macaron = np.array(test_accs)

print(f"Minimum validation loss:{val_losses_macaron.min()} at {np.argmin(val_losses_macaron)+1} epoch.")
print(f"Maximum accuracy:{test_accs_macaron.max()} at {np.argmax(test_accs_macaron)*eval_every_n_epochs+1} epoch.")