## 1. import 
## 2. input & hyperparameter
## 3. pretrain 
## 4. 'dataloader, model build, train, inference'
################################################


## 1. import 
import torch, os, sys
if not os.path.exists("figure"):
	os.makedirs("figure")
torch.manual_seed(0)
sys.path.append('.')
from HINT.dataloader import csv_three_feature_2_dataloader, generate_admet_dataloader_lst, csv_three_feature_2_complete_dataloader
from HINT.molecule_encode import MPNN, ADMET 
from HINT.icdcode_encode import GRAM, build_icdcode2ancestor_dict
from HINT.protocol_encode import Protocol_Embedding
from HINT.model import HINTModel
device = torch.device("cuda:0")  ## cuda:0

## 2. input & hyperparameter 
base_name = 'indication' 

train_file = 'data/' + base_name + '_train.csv'
valid_file = 'data/' + base_name + '_valid.csv'
test_file = 'data/' + base_name + '_test.csv'


mpnn_model = MPNN(mpnn_hidden_size = 50, mpnn_depth=3, device = device)


## 3. pretrain 
admet_model_path = "save_model/admet_model.ckpt"
if not os.path.exists(admet_model_path):
	admet_dataloader_lst = generate_admet_dataloader_lst(batch_size=32)
	admet_trainloader_lst = [i[0] for i in admet_dataloader_lst]
	admet_testloader_lst = [i[1] for i in admet_dataloader_lst]
	admet_model = ADMET(molecule_encoder = mpnn_model, 
						highway_num=2, 
						device = device, 
						epoch=3, 
						lr=5e-4, 
						weight_decay=0, 
						save_name = 'admet_')
	admet_model.train(admet_trainloader_lst, admet_testloader_lst)
	torch.save(admet_model, admet_model_path)
else:
	admet_model = torch.load(admet_model_path)
	admet_model = admet_model.to(device)
	admet_model.set_device(device)



## 4. dataloader, model build, train, inference
train_loader = csv_three_feature_2_dataloader(train_file, shuffle=True, batch_size=32) 
valid_loader = csv_three_feature_2_dataloader(valid_file, shuffle=False, batch_size=32) 
test_loader = csv_three_feature_2_dataloader(test_file, shuffle=False, batch_size=32) 
test_complete_loader = csv_three_feature_2_complete_dataloader(test_file, shuffle=False, batch_size = 32)

icdcode2ancestor_dict = build_icdcode2ancestor_dict()
gram_model = GRAM(embedding_dim = 50, icdcode2ancestor = icdcode2ancestor_dict, device = device)
protocol_model = Protocol_Embedding(output_dim = 50, highway_num=3, device = device)


hint_model_path = "save_model/" + base_name + ".ckpt"
if not os.path.exists(hint_model_path):
	model = HINTModel(molecule_encoder = mpnn_model, 
			 disease_encoder = gram_model, 
			 protocol_encoder = protocol_model,
			 device = device, 
			 global_embed_size = 50, 
			 highway_num_layer = 2,
			 prefix_name = base_name, 
			 gnn_hidden_size = 50,  
			 epoch = 5,
			 lr = 1e-3, 
			 weight_decay = 0, 
			)
	model.init_pretrain(admet_model)
	model.learn(train_loader, valid_loader, test_loader)
	model.bootstrap_test(test_loader)
	torch.save(model, hint_model_path)
else:
	model = torch.load(hint_model_path)
	model.bootstrap_test(test_loader)