Caffe CIFAR-10 Build and Training

This skill provides procedural guidance for building the Caffe deep learning framework from source and training models on the CIFAR-10 dataset.

When to Use This Skill

• Building Caffe from source on Ubuntu/Debian systems

• Training CIFAR-10 or similar image classification models with Caffe

• Configuring Caffe for CPU-only execution

• Troubleshooting Caffe build and dependency issues

Critical Requirements Checklist

Before starting, identify ALL requirements from the task specification:

• Execution mode: CPU-only vs GPU (affects solver configuration)

• Iteration count: Specific number of training iterations required

• Output files: Where training logs and models should be saved

• Model checkpoints: Which iteration's model file is expected

Phase 1: Dependency Installation

System Dependencies

Install required packages before attempting to build:

apt-get update && apt-get install -y
build-essential cmake git
libprotobuf-dev libleveldb-dev libsnappy-dev
libhdf5-serial-dev protobuf-compiler
libatlas-base-dev libgflags-dev libgoogle-glog-dev liblmdb-dev
libopencv-dev libboost-all-dev
python3-dev python3-numpy python3-pip

Verification Step

Confirm critical libraries are installed:

dpkg -l | grep -E "libhdf5|libopencv|libboost"

Phase 2: Caffe Source Acquisition

Clone and Checkout

git clone https://github.com/BVLC/caffe.git cd caffe git checkout 1.0 # Note: Tag is "1.0", not "1.0.0"

Common Mistake

The release tag is 1.0 , not 1.0.0 . Verify with git tag -l if uncertain.

Phase 3: Makefile.config Configuration

Create Configuration File

cp Makefile.config.example Makefile.config

Essential Configuration Changes

Apply these modifications to Makefile.config :

CPU-Only Mode (if no GPU available):

CPU_ONLY := 1

OpenCV Version (for OpenCV 3.x or 4.x):

OPENCV_VERSION := 3

Note: OpenCV 4 may require additional compatibility patches.

HDF5 Paths (Ubuntu-specific):

INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu/hdf5/serial

Python Configuration (Python 3):

PYTHON_LIBRARIES := boost_python3 python3.8 PYTHON_INCLUDE := /usr/include/python3.8 /usr/lib/python3/dist-packages/numpy/core/include

Adjust version numbers based on installed Python version.

Configuration Verification

After editing, verify no duplicate definitions exist:

grep -n "PYTHON_INCLUDE|PYTHON_LIB|CPU_ONLY" Makefile.config

Ensure each setting appears only once in an uncommented form.

Phase 4: Building Caffe

Memory-Aware Compilation

Avoid using all CPU cores on memory-constrained systems:

For systems with limited RAM (< 8GB)

make all -j2

For systems with adequate RAM

make all -j$(nproc)

Build Failure Recovery

If the build fails or is killed (often due to memory):

Clean the build:

make clean

Rebuild with reduced parallelism:

make all -j1

Build Verification

Confirm the binary exists after build:

ls -la .build_release/tools/caffe.bin

or for CPU-only builds:

ls -la .build_release/tools/caffe

Phase 5: Dataset Preparation

Download CIFAR-10

./data/cifar10/get_cifar10.sh

Convert to LMDB Format

./examples/cifar10/create_cifar10.sh

Verification

Confirm LMDB directories exist:

ls -la examples/cifar10/cifar10_train_lmdb ls -la examples/cifar10/cifar10_test_lmdb

Phase 6: Solver Configuration

Modify Solver for Requirements

Edit examples/cifar10/cifar10_quick_solver.prototxt :

Set iteration count:

max_iter: 500 # Or as specified in task

Set execution mode:

solver_mode: CPU # Change from GPU if required

Verification

grep -E "max_iter|solver_mode" examples/cifar10/cifar10_quick_solver.prototxt

Phase 7: Training Execution

Run Training with Output Capture

./build/tools/caffe train
--solver=examples/cifar10/cifar10_quick_solver.prototxt
2>&1 | tee training_output.txt

Alternative Binary Paths

Depending on build configuration, the binary may be at:

• .build_release/tools/caffe

• build/tools/caffe

• .build_release/tools/caffe.bin

Phase 8: Verification

Required Outputs Checklist

Caffe binary exists:

test -f .build_release/tools/caffe && echo "OK" || echo "MISSING"

Model file exists (iteration-specific):

ls -la examples/cifar10/cifar10_quick_iter_*.caffemodel

Training output captured:

test -f training_output.txt && echo "OK" || echo "MISSING"

Solver configured correctly:

grep "solver_mode: CPU" examples/cifar10/cifar10_quick_solver.prototxt

Common Pitfalls

1. Premature Termination

Never stop after make clean or intermediate steps. Complete the full workflow: Dependencies -> Build -> Dataset -> Configure -> Train -> Verify

2. Missing Solver Configuration

The solver file must be modified for:

• CPU vs GPU execution mode

• Specific iteration count requirements

3. Skipping Dataset Preparation

Training will fail without LMDB data. Always run both:

• get_cifar10.sh (download)

• create_cifar10.sh (convert)

4. Build Parallelism Issues

High parallelism (-j$(nproc) ) can exhaust memory. Start with -j2 on constrained systems.

5. Duplicate Configuration Entries

Multiple edits to Makefile.config can create duplicate definitions. Always verify single definitions for each setting.

6. Wrong Git Tag

Use 1.0 not 1.0.0 for the stable release.

Decision Framework

When encountering issues:

• Build killed: Reduce parallelism, run make clean , rebuild with -j1

• Missing headers: Check HDF5 and OpenCV include paths in Makefile.config

• Python errors: Verify Python version matches configuration

• Training fails immediately: Check dataset preparation completed

• Wrong output location: Verify solver paths and output file redirection