I build WebGPU PyTorch debugger, dataset browser and NumPy viewer

Get empowered by Svetovid, whose four heads represent omniscience: the ability to see through space and time in all directions at once. Maintain unified perspective across your tensors, models, and massive datasets. Peer into the architecture of your neural networks and navigate high-dimensional data with absolute clarity.

array_inspector.py

NumPy Inspector #

Visualize any NumPy array with instant dimension awareness. Support for high-dimensional, large arrays.

import svetoviz_webgpu as sv

data = np.random.rand(4, 4, 32, 32).astype(np.float32)

sv.array(data, gap_size=10)

Learn More

pytorch_debugger.py

PyTorch Debugger #

Directly load your PyTorch module, interact, and debug activations in real-time.

import svetoviz_webgpu as sv

img = Image.open("sample_image.jpg").convert("RGB")
img_np = np.array(img).astype(np.float32) / 255.0

input_tensor = torch.from_numpy(img_np).permute(2, 0, 1).unsqueeze(0)

conv2d = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, stride=1, padding=1)

def terminal_callback(buffer, message, images, files):
    output = conv2d(input_tensor)

    buffer.send_system_message(f"Processed image. Shape: {list(output.shape)}")

sv.pytorch(module=conv2d, terminal_callback=terminal_callback)

Learn More

model_viewer.py

Model Analysis #

Deep-dive into model architecture. Explore layers down to a single parameter.

import svetoviz_webgpu as sv

image_processor = DetrImageProcessor.from_pretrained(
    "facebook/detr-resnet-50",
    revision="no_timm",
    device_map="cpu")
model = DetrForObjectDetection.from_pretrained(
    "facebook/detr-resnet-50",
    revision="no_timm",
    device_map="cpu")

sv.model(
    model=model,
    image_processor=image_processor
)

Learn More

model_debug.py

Model Debugger #

Interact with your model and investigate activations in real-time.

import svetoviz_webgpu as sv

tokenizer = GPT2Tokenizer.from_pretrained('gpt2', device_map="cpu")
model = GPT2LMHeadModel.from_pretrained('gpt2', device_map="cpu")

def terminal_callback(buffer, message, images, files):
    text = message  # "Replace me by any text you'd like."
    buffer.send_user_message(text)
    inputs = tokenizer(text, return_tensors="pt")

    outputs = model.generate(**inputs, max_new_tokens=50)

    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
    buffer.send_system_message(decoded)

sv.model(
    model=model,
    tokenizer=tokenizer,
    terminal_callback=terminal_callback
)

Learn More

dataset_handler.py

Dataset Explorer #

Efficiently stream and visualize large-scale HuggingFace datasets from local storage or remote server.

import svetoviz_webgpu as sv

hf_name = "eltorio/ROCOv2-radiology"
dataset = load_dataset(hf_name, split="train")
view = DatasetView(
    name=hf_name,
    dataset=dataset,
    image_columns=["image"],
    cell_width=400,
    cell_height=400
)
sv.save_to_disc(view=view,
                compression="png_0",
                directory=f"/Volumes/Untitled/{hf_name}")

sv.load_from_disc(directory=f"/Volumes/Untitled/{hf_name}")

Learn More

image_browser.py

Image Browser #

Browse through massive collections of images regardless of resolution.

import svetoviz_webgpu as sv

view = DirectorySpiralView(directory="/Desktop/nasaimages")

sv.save_to_disc(view=view,
                compression="jpeg_0",
                directory="/Volumes/Untitled/universe")

sv.load_from_disc(directory="/Volumes/Untitled/universe")

Learn More

Installation