Semantic Segmentation¶

import torch

import torch_pointcloud as tp
from torch_pointcloud.datasets import ScanNet20
from torch_pointcloud.utils.data import collate

# Load the pretrained model
model, info = tp.create_model(
    "spunet-v1m1.scannet20.pointcept",
    task="semantic-segmentation",
    pretrained=True,
    return_info=True,
)
model = model.cuda().eval()

# Get associated transform
transform = info["transform"]

# Load one room: the transform voxelizes it in place and keeps the source cloud under `origin_*`
dataset = ScanNet20(root="data", split="val", transform=transform)
sample = dataset[0]

data = collate([sample])
print(f"Data keys: {data.keys()}")

with torch.no_grad():
    x = data["x"].cuda()
    pos_grid = data["pos_grid"].cuda()
    batch = data["batch"].cuda()
    inverse = data["inverse"].cuda()
    logits = model(x, pos_grid, batch)

# Get predictions at full resolution
preds = logits[inverse].argmax(dim=-1).cpu()
print(f"Logits shape: {tuple(logits.shape)}")
print(f"Predictions shape: {tuple(preds.shape)}")

# Data keys: dict_keys(['pos', 'color', 'normal', 'scene', 'segment', 'instance', 'x', 'origin_pos', 'origin_segment', 'pos_grid', 'inverse', 'batch'])
# Logits shape: (164614, 20)
# Predictions shape: (237360,)

import torch
import torch_pointcloud.transforms as T

torch.manual_seed(0)

transform = T.Compose([
    T.Shift(keys="pos", method="min"),
    T.CopyItems(keys=["pos", "segment"], dst_keys=["origin_pos", "origin_segment"]),
    T.Voxelize(
        pos_key="pos",
        pos_reduce="grid",
        size=0.02,
        keys=["color", "segment"],
        dst_inverse_key="inverse",
    ),
])

sample = {
    "pos": torch.rand(20_000, 3) * 4.0,
    "color": torch.rand(20_000, 3) * 255,
    "segment": torch.randint(0, 20, (20_000,)),
}
sample = transform(sample)
print(f"Sample keys: {sample.keys()}")
print(f"Pos shape: {tuple(sample['pos'].shape)}")
print(f"Inverse shape: {tuple(sample['inverse'].shape)}")

# Sample keys: dict_keys(['pos', 'color', 'segment', 'origin_pos', 'origin_segment', 'inverse'])
# Pos shape: (19975, 3)
# Inverse shape: (20000,)

from torch_pointcloud.metrics import accuracy, confusion_matrix, intersection_over_union

# `preds` is at full resolution, so score it against the labels of the source cloud
cm = confusion_matrix(
    preds,
    data["origin_segment"],
    num_classes=model.num_classes,
    ignore_index=-1,
)
iou = intersection_over_union(cm, average="none")
present = cm.sum(1) > 0

print(f"accuracy {accuracy(cm):.4f}")
print(f"mIoU     {iou[present].mean():.4f}")
# accuracy 0.8856
# mIoU     0.7073

from tqdm.auto import tqdm
from torch.nn import functional as F
from torch.utils.data import DataLoader

import torch
import torch_pointcloud as tp
import torch_pointcloud.transforms as T
from torch_pointcloud.datasets import ShapeNetPart
from torch_pointcloud.utils.data import collate
from torch_pointcloud.config import DATA_DIR

device = "cuda"

# Setup the dataset and dataloader
train_dataset = ShapeNetPart(
    DATA_DIR,
    split="train",
    categories="Airplane",
    transform=T.Rescale(keys="pos"),
)
train_dataloader = DataLoader(
    train_dataset,
    batch_size=16,
    shuffle=True,
    num_workers=4,
    collate_fn=collate,
)

# Create the desired model and optimizer
model = tp.create_model(
    "pointnext-sm",
    task="semantic-segmentation",
    in_channels=0,
    num_classes=50,
).to(device)
optimizer = torch.optim.AdamW(
    model.parameters(),
    lr=1e-3,
    weight_decay=1e-4,
)

# Training loop
num_epochs = 1  # a demo pass, raise it to actually train

model.train()
for epoch in range(num_epochs):
    total_loss = 0.0
    pbar = tqdm(enumerate(train_dataloader), total=len(train_dataloader), desc=f"Epoch {epoch}")
    for i, data in pbar:
        pos = data["pos"].to(device)
        target = data["segment"].to(device)
        batch = data["batch"].to(device)

        optimizer.zero_grad()
        logits = model(None, pos, batch)
        logits = F.log_softmax(logits, dim=1)
        loss = F.nll_loss(logits, target)

        loss.backward()
        optimizer.step()
        total_loss += loss.item()
        if (i + 1) % 10 == 0:
            loss_step = loss.item()
            metrics = {"train/loss_step": f"{loss_step:.3f}"}
            pbar.set_postfix(metrics)

    loss_epoch = total_loss / len(train_dataloader)
    print(f"Loss epoch {epoch}: {loss_epoch:.3f}")


# Loading: 100%|██████████| 1958/1958 [00:00<00:00, 29210.61it/s]
# Epoch 0: 100%|██████████| 123/123 [00:02<00:00, 45.25it/s, train/loss_step=0.275]
# Loss epoch 0: 0.673
# Epoch 1: 100%|██████████| 123/123 [00:02<00:00, 45.30it/s, train/loss_step=0.220]
# Loss epoch 1: 0.244
# Epoch 2: 100%|██████████| 123/123 [00:02<00:00, 45.30it/s, train/loss_step=0.220]
# Loss epoch 2: 0.228
# Epoch 3: 100%|██████████| 123/123 [00:02<00:00, 44.70it/s, train/loss_step=0.223]
# Loss epoch 3: 0.221
# Epoch 4: 100%|██████████| 123/123 [00:02<00:00, 45.24it/s, train/loss_step=0.190]
# Loss epoch 4: 0.215
# Epoch 5: 100%|██████████| 123/123 [00:02<00:00, 45.52it/s, train/loss_step=0.231]
# Loss epoch 5: 0.211
# Epoch 6: 100%|██████████| 123/123 [00:02<00:00, 45.86it/s, train/loss_step=0.170]
# Loss epoch 6: 0.208
# Epoch 7: 100%|██████████| 123/123 [00:02<00:00, 45.59it/s, train/loss_step=0.168]
# Loss epoch 7: 0.205
# Epoch 8: 100%|██████████| 123/123 [00:02<00:00, 45.94it/s, train/loss_step=0.195]
# Loss epoch 8: 0.199
# Epoch 9: 100%|██████████| 123/123 [00:02<00:00, 44.63it/s, train/loss_step=0.161]
# Loss epoch 9: 0.198