Implementation:Ggml org Ggml Mnist model init random

Template:KapsoPageMeta

Summary

mnist_model_init_random initializes an mnist_model struct with randomly generated weights and biases for either a fully connected or convolutional neural network architecture, ready for training from scratch on the MNIST handwritten digit dataset.

API Signature

mnist_model mnist_model_init_random(
    const std::string & arch,
    const std::string & backend,
    const int nbatch_logical,
    const int nbatch_physical
);

Source Location

• File: examples/mnist/mnist-common.cpp, lines 241-310
• Repository: ggml-org/ggml

Parameters

Parameter	Type	Description
arch	const std::string &	Architecture selector. Must be "mnist-fc" (fully connected) or "mnist-cnn" (convolutional).
backend	const std::string &	Backend device selector. Pass "" for the default backend, or specify explicitly (e.g., "CPU", "CUDA0").
nbatch_logical	int	Logical batch size used for gradient accumulation.
nbatch_physical	int	Physical batch size — the number of samples processed in a single forward/backward pass. Default is 500.

Return Value

Returns an mnist_model struct containing all initialized tensors, contexts, and the backend scheduler. The contents vary by architecture:

Fully Connected (mnist-fc)

Tensor	Shape	Description
fc1_weight	[784, 500]	First fully connected layer weights
fc1_bias	[500]	First fully connected layer bias
fc2_weight	[500, 10]	Second fully connected layer weights (output)
fc2_bias	[10]	Second fully connected layer bias (output)

Convolutional (mnist-cnn)

Tensor	Shape	Description
conv1_kernel	[3, 3, 1, 8]	First convolution kernel (3x3, 1 input channel, 8 output channels)
conv1_bias	[1, 1, 8]	First convolution bias
conv2_kernel	[3, 3, 8, 16]	Second convolution kernel (3x3, 8 input channels, 16 output channels)
conv2_bias	[1, 1, 16]	Second convolution bias
dense_weight	[784, 10]	Dense output layer weights
dense_bias	[10]	Dense output layer bias

Common Fields (Both Architectures)

• images — Input tensor for image data.
• buf_ggml_ctx — Two ggml contexts:
  • ctx_static — Holds the model parameters (weights and biases) that persist across batches.
  • ctx_compute — Holds temporary tensors used during forward and backward computation.
• Backend scheduler — Manages dispatching computation to the selected backend device.

Random Initialization Details

Weights are initialized using the C++ <random> standard library:

• RNG: std::mt19937 (Mersenne Twister)
• Distribution: std::normal_distribution with a small standard deviation

This provides reproducible pseudorandom initialization suitable for training from scratch.

Related Function: mnist_model_init_from_file

For loading pre-trained weights rather than initializing randomly, the companion function mnist_model_init_from_file is defined at lines 131-239 in the same source file. It deserializes model parameters from a GGUF file, enabling inference or fine-tuning without training from scratch.

Dependencies

Header	Purpose
ggml.h	Core tensor operations and context management
ggml-backend.h	Backend abstraction (CPU, CUDA, etc.)
gguf.h	GGUF file format support (used by mnist_model_init_from_file)
<random>	C++ standard library random number generation

Language

C++

Related

• Principle:Ggml_org_Ggml_Model_Architecture_Initialization
• Environment:Ggml_org_Ggml_C_Cpp_Build_Environment
• Heuristic:Ggml_org_Ggml_Gradient_Accumulation_Batch_Sizing

Source

• GGML