training/main.py

import torch
import torch.nn as nn
import random
from training.data.load_data import get_dataloders, save_preprocessed_data_to_output
from training.model.loss import ProxyAnchorLoss
from training.model.model import FusedFeatureModel
from training.model.feature_extractor import FeatureExtractor
from training.test_helpers import evaluate_ncm_accuracy_top3
from training.train_helpers import train_or_load_model
from training.helpers.args import parse_and_get_args


SEED_VALUE = 42

torch.manual_seed(SEED_VALUE) 
random.seed(SEED_VALUE)


FULL_TRAIN = True

RENDERS_FOLDER_NAME = "BlenderRenders7"
TEST_FOLDER_NAME = "test_data3"
RENDERS_FOLDER = f"/mnt/c/Users/KMult/Desktop/Praca_inzynierska/models/{RENDERS_FOLDER_NAME}"
TEST_FOLDER = f"/mnt/c/Users/KMult/Desktop/Praca_inzynierska/models/{TEST_FOLDER_NAME}"
VALIDATION_FOLDER = "/mnt/c/Users/KMult/Desktop/Praca_inzynierska/models/validation_data"

BATCH_PRINT_INTERVAL = 50 
NUM_EPOCHS = 50
VALIDATION_CHECK_INTERVAL = 50
VALIDATION_PATIENCE = 60
VALIDATION_MIN_DELTA = 0.001
sample_size = 8192


FILTER_HARD_DATA = False


LOAD_NETWORK = False 
LOAD_SPECIFIC_MODEL_NAME = None

# LOAD_NETWORK = True
# LOAD_SPECIFIC_MODEL_NAME = "BESTID_2_fused_feature_model.pth_full_clip1_segformer0_midas0_dpt0_gate0_batch64_traintypehardmining_bigfusionhead2_lr2e-07_margin0.8_alpha64.0_datasetsize114272_rendersBlenderRenders7_testdatatest_data3.model"

args = parse_and_get_args(LOAD_SPECIFIC_MODEL_NAME)
print(args)

BATCH_SIZE = args.batch
TRAIN_TYPE = args.train_type.lower()

MODELS_USED = {
    'clip': args.clip,
    'segformer': args.segformer,
    'midas': args.midas,
    'dpt': args.dpt,
}

VECTORS_SAVE_FOLDER = f"training/preprocessed_vectors/mega_clip_segformer_midas_dpt4.pth"
TEST_VECTORS_SAVE_FOLDER = f"training/preprocessed_vectors/test_mega_clip_segformer_midas_dpt4.pth"
feature_extractor = FeatureExtractor(use_models=MODELS_USED)

# save_preprocessed_data_to_output(root_folder=RENDERS_FOLDER, root_test_folder=TEST_FOLDER, output_file_path=VECTORS_SAVE_FOLDER, test_output_file_path=TEST_VECTORS_SAVE_FOLDER, feature_extractor=feature_extractor, processor_flags=MODELS_USED)

MODELS_USED_FOR_TRAINING = MODELS_USED.copy()
MODELS_USED_FOR_TRAINING['gate'] = args.gate

train_dataloader, validation_dataloader, test_dataloader, id_to_tag = get_dataloders(RENDERS_FOLDER, TEST_FOLDER, preprocessed_file_path=VECTORS_SAVE_FOLDER, test_preprocessed_file_path=TEST_VECTORS_SAVE_FOLDER, validation_root_dir=None, is_full_train=FULL_TRAIN, batch_size=BATCH_SIZE, sample_size=sample_size, processor_flags=MODELS_USED, filter_hard_data=FILTER_HARD_DATA)
dataset_size = len(train_dataloader.dataset)

MODEL_NAME = (
    "fused_feature_model.pth"
    + ("_full" if FULL_TRAIN else "_sampled")
    + f"_clip{args.clip}"
    + f"_segformer{args.segformer}"
    + f"_midas{args.midas}"
    + f"_dpt{args.dpt}"
    + f"_gate{args.gate}"
    + f"_batch{args.batch}"
    + f"_traintype{args.train_type}"
    + f"_bigfusionhead{args.big_fusion_head}"
    + f"_lr{args.lr}"
    + f"_margin{args.margin}"
    + f"_alpha{args.alpha}"
    + f"_datasetsize{dataset_size}"
    + f"_renders{RENDERS_FOLDER_NAME}"
    + f"_testdata{TEST_FOLDER_NAME}"
    + ".model"
)

EMBEDDING_DIM = 1024 if args.big_fusion_head >= 2 else 512
NUM_BUILDINGS = len(id_to_tag)
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = FusedFeatureModel(feature_dims=feature_extractor.feature_dims, embedding_dim=EMBEDDING_DIM, use_gate=bool(args.gate), big_fusion_head=args.big_fusion_head, use_models=MODELS_USED_FOR_TRAINING).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# criterion = nn.TripletMarginLoss(margin=0.2).to(device)
criterion = ProxyAnchorLoss(
    num_classes=NUM_BUILDINGS, 
    embedding_dim=EMBEDDING_DIM, 
    margin=args.margin, 
    alpha=args.alpha 
).to(device)

model, prototype_tensor, class_ids, total_batches_processed, min_validation_loss = train_or_load_model(
    model, 
    load_network=LOAD_NETWORK, 
    specific_model_name=LOAD_SPECIFIC_MODEL_NAME,
    optimizer=optimizer, 
    criterion=criterion, 
    train_dataloader=train_dataloader, 
    device=device, 
    num_epochs=NUM_EPOCHS, 
    batch_print_interval=BATCH_PRINT_INTERVAL, 
    val_dataloader=validation_dataloader,
    validation_interval=VALIDATION_CHECK_INTERVAL,
    model_name=MODEL_NAME,
    patience=VALIDATION_PATIENCE,
    min_delta=VALIDATION_MIN_DELTA,
    train_type=TRAIN_TYPE,
)

print("\nModel with min validation loss: ", min_validation_loss, " loaded.")

final_accuracy, accuracy_top3 = evaluate_ncm_accuracy_top3(model, test_dataloader, prototype_tensor, class_ids, id_to_tag, device)

print(f"\nFinal Test Accuracy (NCM): {final_accuracy:.4f}, Top-3 Accuracy: {accuracy_top3:.4f}")