TorchCNNBuilder

---

Description in Russian is presented here.

---

TorchCNNBuilder is an open-source framework for the automatic creation of CNN architectures. This framework should first of all help researchers in the applicability of CNN models for a huge range of tasks, taking over most of the writing of the architecture code. This framework is distributed under the 3-Clause BSD license. All the functionality is written only using pytorch (no third-party dependencies).

Installation

---

The simplest way to install framework is using pip:

pip install torchcnnbuilder

Minimum technical requirements

The minimum system requirements for using the library are a Python interpreter version >3.9 and access to a computing system running Windows/Linux. The minimum hardware requirements include a processor (CPU) with 8 cores, 2GB of RAM, a graphics processor (GPU) with 8GB of VRAM, and 2GB of HDD storage.

Additional packages for examples run

Please note that when running examples from the examples folder, additional libraries are used to visualize and generate the dataset:

pip install numpy
pip install pytorch_msssim
pip install matplotlib
pip install tqdm

They are not required for the library to work, so their installation is optional.

Usage

To initialize simple model with encoder-decoder architecture call ForecasterBase class:

from torchcnnbuilder.models import ForecasterBase

model = ForecasterBase(input_size=[H, W],
                       in_time_points=C_in,
                       out_time_points=C_out,
                       n_layers=5)

Where [H, W] - size of image in pixels, C_in - number of input channels, C_out - number of out_channels.

To operate separately with encoder and decoder parts they can be called from model:

encoder = model.encoder
decoder = model.decoder

Examples

Extensive usage scenarios can be found in examples folder.

Components calls and usage in folder usage_examples.

Documentation

Check the documentation here.

Development

In order to check available local Makefile commands run in the project root:

make help

help: Show help for each of the Makefile recipes.
lint: Lint the project with flake8 lib.
doc: Build and run the doc locally.

Application Areas

TorchCNNBuilder enables CNN architectures for diverse real-world applications across multiple domains:

Environmental Monitoring

• Sea ice concentration forecasting
  Predict Arctic and Antarctic ice melt patterns to support climate research and maritime navigation safety using satellite imagery time series.

• Climate pattern recognition
  Analyze large-scale atmospheric data to identify emerging weather patterns, extreme event precursors, and long-term climate trends.

• Pollution level prediction
  Process multispectral sensor data to forecast air/water quality indices and identify pollution sources with spatial CNN architectures.

Remote Sensing

• Satellite image analysis
  Process high-resolution multispectral imagery for applications ranging from urban planning to precision agriculture using specialized encoder architectures.

• Land cover classification
  Automate large-scale terrain mapping with attention-based CNNs that handle spectral, spatial and temporal dimensions of data.

• Disaster monitoring
  Develop change detection systems that compare pre/post-event satellite imagery to assess flood, fire or earthquake damage in near-real-time.

Medical Imaging

• Automated diagnosis from X-ray/MRI scans
  Develop assistive diagnostic systems that can detect abnormalities in medical images with pixel-level precision while reducing radiologist workload.

• Tumor segmentation
  Create 3D convolutional networks for precise volumetric analysis of cancerous growths in CT/MRI scans.

• Medical time-series analysis
  Process sensor streams to predict patient deterioration through temporal features processing architectures.

Industrial Applications

• Predictive maintenance
  Monitor equipment vibration patterns and thermal signatures to forecast mechanical failures.

• Quality control in manufacturing
  Implement real-time visual inspection systems that detect defects in production lines.

Financial Forecasting

• Time-series prediction
  Build hybrid CNN-LSTM architectures that extract both spatial patterns from market heatmaps and temporal dependencies from price histories.

• Market trend analysis
  Process alternative data sources like satellite images of parking lots or social media sentiment through CNN architectures.

Sources

---

• Forecasting of Sea Ice Concentration using CNN, PDE discovery and Bayesian Networks
• Surrogate Modelling for Sea Ice Concentration using Lightweight Neural Ensemble
• Post about framework development on habr.com - in russian

Contributing

• To join the project feel free to contact us;

• Issues and Pull Requests: submit bugs found or log feature requests.

Acknowledgement

The project is supported by FASIE - Foundation for Assistance to Small Innovative Enterprises.