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 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 matrix_interp(x, xs, ys):
    orig_shape = ys.shape

    # ========================================================================
    # Interpolation.
    # ========================================================================
    xs = xs.astype(ys.dtype)
    x = x.astype(ys.dtype)
    # Pad control points for extrapolation.
    xs = np.concatenate([[np.finfo(xs.dtype).min], xs, [np.finfo(xs.dtype).max]], axis=0)
    ys = np.concatenate([ys[:1], ys, ys[-1:]], axis=0)

    # Compute slopes, pad at the edges to flatten.
    sloops = (ys[1:] - ys[:-1]) / np.expand_dims((xs[1:] - xs[:-1]), axis=-1)
    sloops = np.pad(sloops[:-1], [(1, 1), (0, 0)])

    # Solve for intercepts.
    intercepts = ys - sloops * np.expand_dims(xs, axis=-1)

    # Search for the line parameters at each input data point.
    # Create a grid of the inputs and piece breakpoints for thresholding.
    # Rely on argmax stopping on the first true when there are duplicates,
    # which gives us an index into the parameter vectors.
    idx = np.argmax(np.expand_dims(xs, axis=-2) > np.expand_dims(x, axis=-1), axis=-1)
    sloop = sloops[idx]
    intercept = intercepts[idx]

    # Apply the linear mapping at each input data point.
    y = sloop * np.expand_dims(x, axis=-1) + intercept
    y = y.astype(ys.dtype)
    y = y.reshape([-1, orig_shape[1]])
    return y

class RandomTimeWarping():
    def __init__(self,
                 apply_ratio,
                 scale_range,
                 min_apply_size,
                 apply_post_mask=True):
        self.apply_ratio = apply_ratio
        self.scale_range = scale_range
        self.min_apply_size = min_apply_size
        self.apply_post_mask = apply_post_mask

    def __call__(self,
                 data: Dict[str, Any]) -> Dict[str, Any]:
        """Apply time warping to data.
        """
        if random.random() > self.apply_ratio:
            return data

        feature = data["feature"]
        if self.apply_post_mask:
            temp = feature.reshape([feature.shape[0], -1])
            mask = (temp == 0).all(axis=0).reshape(
                [1, feature.shape[1], feature.shape[2]])
            mask = np.bitwise_not(mask).astype(feature.dtype)
        else:
            mask = None
        orig_shape = feature.shape
        tlen = orig_shape[1]
        if tlen > self.min_apply_size:
            # `[C, T, J] -> [T, C*J]`
            feature = feature.transpose([1, 0, 2]).reshape([tlen, -1])

            assert len(feature.shape) == 2, f"{feature.shape}"
            aug_scale = np.random.random() * (self.scale_range[1] - self.scale_range[0]) + self.scale_range[0]
            aug_tlen = int((tlen-1) * aug_scale)

            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])
            if mask is not None:
                # `[1, T, J] -> [T, 1*J]`
                mask = mask.transpose([1, 0, 2]).reshape([tlen, -1])
                assert len(mask.shape) == 2, f"{mask.shape}"
                newmask = matrix_interp(x, xs, mask)
                # `[T, 1*J] -> [T, 1, J] -> [1, T, J]`
                newmask = newmask.reshape([-1, 1, orig_shape[2]])
                newmask = newmask.transpose([1, 0, 2])
                # `[1, T, J] -> [C, T, J]`
                newmask = np.repeat(newmask, orig_shape[0], axis=0)
                # Masked out interpolated position.
                newmask[newmask < 0.95] = 0.0
                newfeature *= newmask
        else:
            newfeature = feature

        data["feature"] = newfeature
        return data

    def __str__(self):
        return f"{self.__class__.__name__}:{self.__dict__}"

# 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_w_daug_twarp = Compose([
    RandomTimeWarping(
        apply_ratio=0.5,
        scale_range=(0.5, 2.0),
        min_apply_size=10),
    ToTensor()])

transforms_w_daug_tclip_twarp = Compose([
    RandomClip(
        apply_ratio=0.5,
        clip_range=(0.4, 0.6),
        min_apply_size=10),
    RandomTimeWarping(
        apply_ratio=0.5,
        scale_range=(0.5, 2.0),
        min_apply_size=10),
    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_w_daug_twarp)

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

model = TransformerEnISLR(in_channels=in_channels,
                          inter_channels=inter_channels,
                          out_channels=out_channels,
                          activation=activation,
                          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)

# 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)

tren_norm_first = True
tren_add_tailnorm = True

model_default = TransformerEnISLR(
    in_channels=in_channels,
    inter_channels=inter_channels,
    out_channels=out_channels,
    activation=activation,
    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.")