Svetoviz

Conv3d

https://pytorch.org/docs/stable/generated/torch.nn.Conv3d.html

Sidenote: You can visualize the entire dataset using this code:

                    import svetoviz_webgpu as sv

hf_name = "g4m3r/T1w_MRI_Brain_Slices"
dataset = load_dataset(hf_name, split="train")
output_dir = f"/Volumes/Untitled/{hf_name}"
layout = DatasetLayout(
    name=hf_name,
    dataset=dataset,
    image_columns=["image"],
    cell_size=(256, 256),
)
sv.save_to_disc(
    view=layout,
    threaded=True,
    directory=output_dir,
    compression=CompressionMode.PNG_0
)
sv.load_from_disc_web(directory=output_dir)
                        

g4m3r/T1w_MRI_Brain_Slices

Spread

conv3d_demo.py

                import svetoviz_webgpu as sv

# 1. Prepare 3D volumetric data (Streaming from HF)
ds = load_dataset("g4m3r/T1w_MRI_Brain_Slices", split="train", streaming=True)

# 2. Skip first 2 noisy slices, take 30 consecutive coronal layers
slices = [np.array(ex['image'].convert("L")) for ex in ds.skip(2).take(30)]
input_volume = torch.from_numpy(np.stack(slices)).unsqueeze(0).unsqueeze(0)

# 3. Define the 3D Convolutional module
conv3d = nn.Conv3d(in_channels=1, out_channels=16, kernel_size=3, padding=1)

def terminal_callback(buffer, message, images, files):
    # 4. Process volume (1, 1, 30, H, W) through the module
    output = conv3d(input_volume)
    buffer.send_system_message(f"Voxel Grid Ready. Shape: {list(output.shape)}")

# 5. Start the interactive session
sv.pytorch_web(module=conv3d, terminal_callback=terminal_callback)
                    

Activation input and output (theme light, flat)

Output Mapping

Selected output value

Filter

3D Kernel

Kernel channel

Color map Plasma

Color map Magma

Next Section ConvTranspose3d

→