import wandb
wandb.login()

#@title imports
import wandb
import torch
import torchvision
from torch import nn
from torchvision import transforms as T
from fastai.data.all import *
from fastai.vision.all import *
from fastai.callback.wandb import *
from timm.optim.rmsprop_tf import RMSpropTF
from timm.optim.lookahead import Lookahead

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f'Using device: {device}')

#@title dataset from HF
from torchvision import transforms as T
from datasets import load_dataset
dataset = load_dataset('huggan/CelebA-faces')
tfm = T.Compose([T.Resize(64), T.CenterCrop(64)])
def transforms(examples):
    examples["image"] = [tfm(image.convert("RGB")) for image in examples["image"]]
    return examples
dataset = dataset.with_transform(transforms)['train']

# Example 64px image
dataset[0]['image']

# Class for crappified image
class PILImageNoised(PILImage): pass
class TensorImageNoised(TensorImage):
    def show(self, ctx=None, **kwargs):
        super().show(ctx=ctx, **kwargs)
PILImageNoised._tensor_cls = TensorImageNoised

# Transform (TODO experiment)
class Crappify(Transform):
    def encodes(self, x:TensorImageNoised): 
        x = IntToFloatTensor()(x)
        blurred = T.GaussianBlur(3)(x) # Add some random blur
        noise_amount = torch.rand(x.shape[0], device=x.device)
        noise = torch.rand_like(x, device=x.device)
        noised = torch.lerp(blurred, noise, noise_amount.view(-1, 1, 1, 1)) * 255
        return noised

# Dataloader
dblock = DataBlock(blocks=(ImageBlock(cls=PILImageNoised),ImageBlock(cls=PILImage)),
                   get_items=lambda pth: range(len(dataset)), # Gets the indexes
                   getters=[lambda idx: np.array(dataset[idx]['image'])]*2,
                   batch_tfms=[Crappify])
# dls = dblock.dataloaders('', bs=128)
dls = dblock.dataloaders('', bs=64) # Half batch size to save mem when using extra nn for perceptual loss
dls.show_batch()

# Create a learner and pick LR
# learn = Learner(dls, model, loss_func=MSELossFlat())

loss_fn_alex = lpips.LPIPS(net='alex').to(device)
loss_fn_mse = MSELossFlat()

def combined_loss(preds, y):
  return loss_fn_alex.forward(preds, y).mean() + loss_fn_mse(preds, y)

opt_func = partial(OptimWrapper, opt=RmsLookahead)

learn = Learner(dls, model, loss_func=combined_loss, opt_func=opt_func)
learn.lr_find()

learn.model.load_state_dict(torch.load('./faces_10e_model.pt'))

#@title sample with naive method from first test
n_steps = 40 # Set steps to 1 to see raw preds on first input
x = torch.rand(64, 3, 64, 64).to(device) # Raw noise starting point

for i in range(n_steps):
    with torch.no_grad():
        pred = model(x)
    mix_factor = 1/(n_steps - i)
    x = x*(1-mix_factor) + pred*mix_factor

fig, ax = plt.subplots(1, 1, figsize=(16, 16))
ax.imshow(torchvision.utils.make_grid(x.detach().cpu(), nrow=8).permute(1, 2, 0))

#@title sample with an optimiser
n_steps = 10 # We can stop early if pred noise is close to 0
torch.random.manual_seed(1000)
x = torch.rand(64, 3, 64, 64).to(device)
x.requires_grad = True

lr_max=5e-2
div=100000.
div_final=1e5
pct_start=0.5

lr_sched_fn = combined_cos(pct_start, lr_max/div, lr_max, lr_max/div_final)
optim = OptimWrapper(opt=torch.optim.SGD([x], lr=lr_max, momentum=0.01))

#optim = OptimWrapper(opt=torch.optim.Adam([x], lr=lr_max))

for i in range(n_steps):
    pos = i/(n_steps)
    lr = lr_sched_fn(pos)
    optim.set_hyper('lr', lr)
    
    with torch.no_grad():
        net_output = (model(x) + TF.hflip(model(TF.hflip(x))))/2.0
        noise_pred = x - net_output

    # # Early stopping
    if (noise_pred.float()).var() < 0.0025 or i == n_steps-1:
        x = net_output
        print('Stopping at:', i)
        break

    # noise_pred is basically the grad, so GD on this should find a minimum!
    x.grad = noise_pred.float()
    print(i, x.grad.mean(), x.grad.var(), noise_pred.float().var(), lr) # Useful to watch the noise variance
    optim.step()
    optim.zero_grad() # Not really needed since we're setting .grad ourselves but anyway...
    

fig, ax = plt.subplots(1, 1, figsize=(16, 16))
ax.imshow(torchvision.utils.make_grid(x.detach().cpu(), nrow=8).permute(1, 2, 0))

#@title sample with an optimiser
n_steps = 50 # We can stop early if pred noise is close to 0
torch.random.manual_seed(1000)
x = torch.rand(64, 3, 64, 64).to(device)
x.requires_grad = True

lr_max=1e-2
div=100000.
div_final=1e5
pct_start=0.25

lr_sched_fn = combined_cos(pct_start, lr_max/div, lr_max, lr_max/div_final)
optim = OptimWrapper(opt=torch.optim.SGD([x], lr=lr_max, momentum=0.01))

for i in range(n_steps):
    pos = i/(n_steps)
    lr = lr_sched_fn(pos)
    optim.set_hyper('lr', lr)
    
    with torch.no_grad():
        net_output = (model(x) + TF.hflip(model(TF.hflip(x))))/2.0
        noise_pred = (x - net_output)
        
    # # Early stopping
    if (noise_pred.float()).var() < 0.0025 or i == n_steps-1:
        x = net_output
        print('Stopping at:', i)
        break

    # noise_pred is basically the grad, so GD on this should find a minimum!
    x.grad = noise_pred.float()
    if i%10==0:
        print(i, x.grad.mean(), x.grad.var(), noise_pred.float().var(), lr) # Useful to watch the noise variance
    torch.nn.utils.clip_grad_norm_(x, 250.0)
    optim.step()
    optim.zero_grad() # Not really needed since we're setting .grad ourselves but anyway...
    

fig, ax = plt.subplots(1, 1, figsize=(16, 16))
ax.imshow(torchvision.utils.make_grid(x.detach().cpu(), nrow=8).permute(1, 2, 0))