import torch_utils as tu

SEED = 42
tu.tools.seed_everything(SEED)

import timm

model = timm.create_model('tresnet_m', pretrained=True)
model.global_pool = tu.layers.FastGlobalConcatPool2d(flatten=True)
model.head = tu.layers.get_attention_fc(2048*2, 1) 
model.cuda()

from pytorchcv.model_provider import get_model as ptcv_get_model

model = ptcv_get_model('seresnext50_32x4d', pretrained=True)
model.features.final_pool = tu.layers.GeM() 
model.output = tu.layers.get_simple_fc(2048, 1)   
model.cuda()

# model summary
tu.models.summary(model, (3,224,224))

# 3 channels pretrained weights to 1 channel
weight_rgb = model.conv1.weight
weight_grey = weight_rgb.sum(dim=1, keepdim=True)
model.conv1 = nn.Conv2d(1, 64, kernel_size=xxx, stride=xxx, padding=xxx, bias=False)
model.conv1.weight = torch.nn.Parameter(weight_grey)

# 2D models to 3d models using ACSConv (advanced)
## using code in this repo: https://github.com/M3DV/ACSConv

optimizer_ranger = tu.Ranger(model_conv.parameters(), lr=LR)

# optimizer = torch.optim.AdamW(model_conv.parameters(), lr=LR, weight_decay=2e-4)

lr_finder = tu.LRFinder(model, optimizer, criterion, device="cuda")
lr_finder.range_test(train_loader, end_lr=10, num_iter=100)
lr_finder.plot() # to inspect the loss-learning rate graph
lr_finder.reset() # to reset the model and optimizer to their initial state

scheduler = tu.CosineAnnealingWarmUpRestarts(optimizer, T_0=T, T_mult=1, eta_max=LR, T_up=0, gamma=0.05)

# torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0, T_mult=1, eta_min=0, last_epoch=-1)

# torch.optim.lr_scheduler.OneCycleLR
# tu.OneCycleScheduler