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 numpy as np

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

from torchinfo import summary

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 (
    USE_FACE,
    USE_LHAND,
    USE_POSE,
    USE_RHAND,
    get_fullbody_landmarks
)
from modules_gislr.layers import (
    Identity,
    GPoolRecognitionHead,
    TransformerEncoderLayer,
    TransformerEncoder,
)
from modules_gislr.train_functions import (
    test_loop,
    val_loop,
    train_loop
)
from modules_gislr.transforms import (
    PartsBasedNormalization,
    ReplaceNan,
    SelectLandmarksAndFeature,
    ToTensor,
    matrix_interp
)
from modules_gislr.utils import (
    make_causal_mask,
    select_reluwise_activation
)

class SelectiveResize():
    def __init__(self,
                 min_tlen=8,
                 max_tlen=None):
        assert min_tlen is not None or max_tlen is not None
        self.min_tlen = min_tlen
        self.max_tlen = max_tlen

    def __call__(self,
                 data: Dict[str, Any]) -> Dict[str, Any]:

        feature = data["feature"]
        orig_shape = feature.shape
        tlen = orig_shape[1]
        resize_mode = None
        if self.min_tlen and tlen < self.min_tlen:
            resize_mode = "enlarge"
        if self.max_tlen and tlen > self.max_tlen:
            resize_mode = "shrink"
        if resize_mode is None:
            return data

        # Apply warping.
        # `[C, T, J] -> [T, C*J]`
        feature = feature.transpose([1, 0, 2]).reshape([tlen, -1])

        assert len(feature.shape) == 2, f"{feature.shape}"
        aug_tlen = self.min_tlen if resize_mode == "enlarge" else self.max_tlen

        x = np.linspace(0, tlen-1, num=aug_tlen)
        xs = np.arange(tlen)
        newfeature = matrix_interp(x, xs, feature)
        # `[T, C*J] -> [T, C, J] -> [C, T, J]`
        newfeature = newfeature.reshape([-1, orig_shape[0], orig_shape[2]])
        newfeature = newfeature.transpose([1, 0, 2])

        data["feature"] = newfeature
        return data

class TransformerEnISLR(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_layer_norm_eps=1e-5,
                 tren_norm_first=True,
                 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.
        enlayer = TransformerEncoderLayer(
                dim_model=inter_channels,
                num_heads=tren_num_heads,
                dim_ffw=tren_dim_ffw,
                dropout=tren_dropout,
                activation=activation,
                layer_norm_eps=tren_layer_norm_eps,
                norm_first=tren_norm_first,
                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,
            layer_norm_eps=tren_layer_norm_eps,
            norm_first=tren_norm_first,
            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_default = Compose([ToTensor()])

transforms_resize_32 = Compose([
    SelectiveResize(min_tlen=8, max_tlen=32),
    ToTensor()])

transforms_resize_64 = Compose([
    SelectiveResize(min_tlen=8, max_tlen=64),
    ToTensor()])

transforms_resize_128 = Compose([
    SelectiveResize(min_tlen=8, max_tlen=128),
    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)

for trans in [transforms_default, transforms_resize_32, transforms_resize_64, transforms_resize_128]:
    dataset = HDF5Dataset(hdf5_files, pre_transforms=pre_transforms, transforms=trans)
    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"
tren_num_layers = 2
tren_num_heads = 2
tren_dim_ffw = 256
tren_dropout_pe = 0.1
tren_dropout = 0.1
tren_layer_norm_eps = 1e-5
tren_norm_first = True
tren_add_bias = True
tren_add_tailnorm = True

pooling_type = "none"

model = TransformerEnISLR(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_layer_norm_eps=tren_layer_norm_eps,
                          tren_norm_first=tren_norm_first,
                          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)

pooling_type = "none"
max_tlen = 543

model = TransformerEnISLR(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_layer_norm_eps=tren_layer_norm_eps,
                          tren_norm_first=tren_norm_first,
                          tren_add_bias=tren_add_bias,
                          tren_add_tailnorm=tren_add_tailnorm)
# input_size: `[N, C, T, J]`
summary(model, input_size=(batch_size, len(use_features), max_tlen, len(use_landmarks)))

====================================================================================================
Layer (type:depth-idx)                             Output Shape              Param #
====================================================================================================
TransformerEnISLR                                  [2, 10]                   --
├─Linear: 1-1                                      [2, 543, 64]              16,704
├─ReLU: 1-2                                        [2, 543, 64]              --
├─Identity: 1-3                                    [2, 543, 64]              --
├─TransformerEncoder: 1-4                          [2, 543, 64]              --
│    └─PositionalEncoding: 2-1                     [2, 543, 64]              --
│    │    └─Dropout: 3-1                           [2, 543, 64]              --
│    └─ModuleList: 2-2                             --                        --
│    │    └─TransformerEncoderLayer: 3-2           [2, 543, 64]              49,984
│    │    └─TransformerEncoderLayer: 3-3           [2, 543, 64]              49,984
│    └─LayerNorm: 2-3                              [2, 543, 64]              128
├─GPoolRecognitionHead: 1-5                        [2, 10]                   --
│    └─Linear: 2-4                                 [2, 10]                   650
====================================================================================================
Total params: 117,450
Trainable params: 117,450
Non-trainable params: 0
Total mult-adds (M): 0.23
====================================================================================================
Input size (MB): 1.13
Forward/backward pass size (MB): 13.34
Params size (MB): 0.47
Estimated Total Size (MB): 14.94
====================================================================================================

# 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_w_norm,
    transforms=transforms_default,
    load_into_ram=load_into_ram)
val_dataset = HDF5Dataset(
    val_hdf5files,
    pre_transforms=pre_transforms_w_norm,
    transforms=transforms_default,
    load_into_ram=load_into_ram)
test_dataset = HDF5Dataset(
    test_hdf5files,
    pre_transforms=pre_transforms_w_norm,
    transforms=transforms_default,
    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)

tren_norm_first = True
tren_add_tailnorm = True
pooling_type = "none"

model_default = TransformerEnISLR(
    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_layer_norm_eps=tren_layer_norm_eps,
    tren_norm_first=tren_norm_first,
    tren_add_bias=tren_add_bias,
    tren_add_tailnorm=tren_add_tailnorm)
print(model_default)

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

# Train, validation, and evaluation.
model_default.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_default, loss_fn, optimizer, device)
    val_loss = val_loop(val_dataloader, model_default, loss_fn, device)
    val_losses.append(val_loss)

    if (epoch+1) % eval_every_n_epochs == 0:
        acc = test_loop(test_dataloader, model_default, device)
        test_accs.append(acc)
train_losses_default = np.array(train_losses)
val_losses_default = np.array(val_losses)
test_accs_default = np.array(test_accs)

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