Create DEPRECATED.md

data/cif_models/DEPRECATED.md

@@ -0,0 +1,112 @@
+# DEPRECATED
+
+## Why `data/cif_models` is deprecated
+
+The previous approach based on storing datasets as large collections of small `.cif` files and performing preprocessing directly on an HPC cluster is strongly discouraged for several reasons:
+
+### 1. Inefficient I/O on HPC systems
+
+HPC clusters are not optimized for workloads involving a large number of small files. Reading thousands (or millions) of `.cif` files leads to:
+
+* High filesystem overhead
+* Significant slowdown due to metadata access
+* Poor overall I/O performance
+
+### 2. Reduced GPU utilization
+
+Preprocessing performed on-the-fly during training:
+
+* Continuously consumes CPU resources
+* Creates a bottleneck between CPU and GPU
+* Causes **low GPU utilization**
+
+As a result, GPUs remain idle while waiting for data, which is highly inefficient.
+
+### 3. Impact on shared cluster usage
+
+Because of the inefficiencies above:
+
+* Jobs take longer than necessary
+* GPUs are occupied for extended periods
+* Other users must wait longer for access
+
+This negatively affects overall cluster fairness and efficiency.
+
+### 4. Storage quota issues
+
+During preprocessing:
+
+* Each `.cif` file often produces one or more cache files
+* Total number of files grows rapidly
+
+For large datasets:
+
+* `#cache files + #original cif files` can easily exceed SSD quotas
+* This can cause job failures or require manual cleanup
+
+---
+
+## New approach
+
+To address these issues, a new preprocessing pipeline has been introduced.
+
+### Key idea
+
+Instead of processing `.cif` files on-the-fly:
+
+* Preprocessing is done **once in advance**
+* Results are stored in a small number of **LMDB files**
+
+### Supported datasets
+
+* **QMOF** (originally `.cif`-based)
+* **ODAC23**
+
+  * IS2R LMDB format
+  * Also includes `ODAC_init.tar` with `.cif` structures (compatible with QMOF-style preprocessing)
+
+### Output format
+
+Each dataset/model combination produces:
+
+* `*_train.lmdb`
+* `*_val.lmdb`
+* `*_test.lmdb`
+
+This reduces:
+
+* File count from millions → **2–3 files total**
+* I/O overhead dramatically
+* CPU preprocessing load during training
+
+---
+
+## Benefits
+
+* Efficient sequential reads (LMDB)
+* High GPU utilization
+* Minimal CPU bottleneck
+* No filesystem overload
+* Avoids SSD quota issues
+* Faster and more stable training
+
+---
+
+## Implementation
+
+The new preprocessing pipeline is implemented in:
+
+* The **latest version of the project code** (GitLab repository)
+
+Preprocessed datasets are available in this repository under:
+
+```
+data/lmdb
+```
+
+---
+
+## Recommendation
+
+Do **not** use `data/cif_models` for training or preprocessing on HPC clusters.
+Always use the preprocessed LMDB datasets instead.