grasp-tool 0.1.0__py3-none-any.whl

grasp_tool-0.1.0.dist-info/METADATA

@@ -0,0 +1,511 @@
+Metadata-Version: 2.3
+Name: grasp-tool
+Version: 0.1.0
+Summary: GRASP: a graph-based representation learning toolkit for spatial transcriptomics
+Author: wzx
+Author-email: unknown@example.com
+Requires-Python: >=3.9,<3.13
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Provides-Extra: ot
+Requires-Dist: POT (>=0.9,<0.10) ; extra == "ot"
+Requires-Dist: joblib (>=1.3,<2.0)
+Requires-Dist: leidenalg (>=0.10,<0.11)
+Requires-Dist: matplotlib (>=3.7,<4.0)
+Requires-Dist: networkx (>=3.2,<4.0)
+Requires-Dist: numpy (>=1.26,<2.0)
+Requires-Dist: pandas (>=2.0,<3.0)
+Requires-Dist: python-igraph (>=0.11,<0.12)
+Requires-Dist: scikit-learn (>=1.0,<2.0)
+Requires-Dist: scipy (>=1.9,<2.0)
+Requires-Dist: seaborn (>=0.13,<0.14)
+Requires-Dist: shapely (>=2.0,<3.0)
+Requires-Dist: tqdm (>=4.0,<5.0)
+Requires-Dist: umap-learn (>=0.5,<0.6)
+Description-Content-Type: text/markdown
+
+# GRASP_pack
+
+This repository is being reorganized into an installable Python package.
+
+- Distribution name: `grasp-tool`
+- Import name: `grasp_tool`
+
+## Repo layout
+
+- `grasp_tool/`: GRASP package source (canonical code)
+- `utils_code/`, `gnn_model/`: backward-compatible wrappers (legacy paths)
+- `ELLA/`: third-party reference project (NOT part of GRASP; do not package)
+- `envs/`: conda environment exports for reproducibility (runtime baseline)
+
+## Installation (recommended)
+
+The recommended workflow is:
+
+- Use conda/mamba to create a stable Python environment (and GPU toolchain if needed)
+- Use pip to install the GRASP package from PyPI (`grasp-tool`)
+
+### 1) Create a conda environment
+
+```bash
+conda create -n grasp python=3.9 -y
+conda activate grasp
+```
+
+### 2) Install GRASP from PyPI
+
+```bash
+pip install grasp-tool
+```
+
+Smoke checks (should work without training deps):
+
+```bash
+grasp-tool --help
+grasp-tool train-moco --help
+```
+
+If you're running from this repo checkout, you can also run a small demo:
+
+```bash
+grasp-tool register \
+  --pkl_file example_pkl/simulated1_data_dict.pkl \
+  --output_pkl outputs/simulated1_registered.pkl
+```
+
+## Tiny demo (end-to-end smoke test)
+
+If you have a repo checkout, you can run a fast end-to-end smoke test on a small
+subset of `example_pkl/simulated1_data_dict.pkl`.
+
+This demo runs:
+
+- `register` (with `--nc_demo 4`)
+- create a tiny `df_registered` subset + `pairs.csv`
+- `portrait` (JS distances)
+- `partition-graphs`
+- `augment-graphs`
+- `build-train-pkl`
+- `train-moco` for 1 epoch
+
+Run:
+
+```bash
+# Default: uses conda env name "grasp"
+bash scripts/tiny_demo_example_pkl.sh
+```
+
+Override the conda env name (useful if you have multiple envs):
+
+```bash
+GRASP_CONDA_ENV=<your_env_name> bash scripts/tiny_demo_example_pkl.sh
+```
+
+Outputs are written under:
+
+- `outputs/tiny_demo_example_pkl_<timestamp>/`
+
+Notes:
+
+- The final training step requires `torch` and `torch-geometric`.
+- `example_pkl/` is excluded from PyPI release artifacts; this demo is intended for
+  repo checkouts.
+- The `scripts/` directory is not part of the installed PyPI wheel. If you installed
+  via `pip install grasp-tool`, you need to clone this repo to use this demo script.
+
+### 3) Training dependencies (NOT installed by pip)
+
+`pip install grasp-tool` intentionally does NOT pull in the training stack.
+
+If you want to run training-related commands:
+
+- `build-train-pkl` (needs `torch` + `torch-geometric`)
+- `train-moco` (needs `torch` + `torch-geometric`)
+
+Install them following the official PyTorch / PyTorch Geometric (PyG) instructions.
+We recommend installing them via conda inside the same env.
+
+## Development (maintainers)
+
+Poetry is only needed for development and release.
+
+```bash
+poetry install
+poetry run grasp-tool --help
+```
+
+## Environment (reproducibility baseline)
+
+These files are meant for *reproducing a known-good runtime environment*, not for
+publishing minimal dependencies:
+
+- `envs/grasp-conda.yml`: full export (large)
+- `envs/grasp-conda-from-history.yml`: minimal export (often needs extra installs)
+
+Typical usage:
+
+```bash
+conda env create -f envs/grasp-conda.yml
+conda activate grasp
+```
+
+Optional extras (only needed for non-core utilities):
+
+- Optimal Transport utilities (`POT` / import name `ot`): `pip install grasp-tool[ot]`
+
+## Full pipeline (from scratch)
+
+The recommended example input is:
+
+- `example_pkl/simulated1_data_dict.pkl`
+
+This PKL contains the raw inputs required by `register` (and also already includes
+`df_registered`; the steps below still show how to run the full pipeline end-to-end).
+
+### 0) Register (coordinate normalization)
+
+Input: a PKL dict containing at least:
+
+- `data_df` (DataFrame; must include `cell`, `type`, `centerX`, `centerY`, `x`, `y`)
+- `cell_mask_df` (DataFrame; columns: `cell`, `x`, `y`)
+- `nuclear_boundary` (dict: cell -> DataFrame with columns `x`, `y`)
+
+Run:
+
+```bash
+python -m grasp_tool register \
+  --pkl_file example_pkl/simulated1_data_dict.pkl \
+  --output_pkl outputs/simulated1_registered.pkl
+```
+
+Output: a PKL dict with keys:
+
+- `df_registered`
+- `nuclear_boundary_df_registered`
+- `cell_radii`
+- `cell_nuclear_stats`
+
+### 0.5) (Optional) Cell/gene visualization (cellplot)
+
+This is optional and mainly used for sanity-checking transcript spatial patterns.
+
+Notes:
+
+- This command can generate a large number of images if you do not restrict `--cells` and `--genes`.
+- The input PKL must contain the required keys for the selected `--mode`.
+  - `--mode raw-cell`: expects `cell_boundary` (and optionally `nuclear_boundary`) in the raw PKL.
+    If your raw PKL does not have `cell_boundary`, use `--mode registered-gene` instead.
+
+Registered-gene plots (recommended; uses `df_registered`):
+
+```bash
+python -m grasp_tool cellplot \
+  --mode registered-gene \
+  --pkl outputs/simulated1_registered.pkl \
+  --output_dir outputs/cellplot \
+  --dataset simulated1 \
+  --cells cell_11 \
+  --genes gene_6_2_1,gene_6_3_1 \
+  --with_nuclear 1
+```
+
+Raw cell boundary plots (uses `cell_boundary` / `nuclear_boundary` in the raw PKL):
+
+```bash
+python -m grasp_tool cellplot \
+  --mode raw-cell \
+  --pkl example_pkl/simulated1_data_dict.pkl \
+  --output_dir outputs/cellplot_raw \
+  --dataset simulated1
+```
+
+### 1) (Optional) JS distance
+
+```bash
+python -m grasp_tool portrait \
+  --pkl_file outputs/simulated1_registered.pkl \
+  --output_dir outputs/portrait \
+  --use_same_r \
+  --visualize_top_n 0 \
+  --auto_params
+```
+
+### 2) Partition + build per-cell graphs (node/adj CSV)
+
+```bash
+python -m grasp_tool partition-graphs \
+  --pkl outputs/simulated1_registered.pkl \
+  --graph_root outputs/graphs \
+  --n_sectors 20 \
+  --m_rings 10 \
+  --k_neighbor 5
+```
+
+You can restrict scope for a quick smoke test:
+
+```bash
+python -m grasp_tool partition-graphs \
+  --pkl outputs/simulated1_registered.pkl \
+  --graph_root outputs/graphs_demo \
+  --cells cell_11,cell_135 \
+  --genes gene_0_0_0,gene_0_1_0
+```
+
+### 3) Graph augmentation
+
+```bash
+python -m grasp_tool augment-graphs \
+  --graph_root outputs/graphs \
+  --dropout_ratio 0.1 \
+  --seed 2025
+```
+
+### 4) Build training PKL
+
+You need a `pairs.csv` with columns: `cell,gene`.
+
+Example (generate pairs from `df_registered`):
+
+```bash
+python -c "import pickle, pandas as pd; d=pickle.load(open('outputs/simulated1_registered.pkl','rb')); pairs=d['df_registered'][['cell','gene']].drop_duplicates(); pairs.to_csv('outputs/pairs.csv', index=False); print('wrote outputs/pairs.csv', len(pairs))"
+```
+
+Then:
+
+```bash
+python -m grasp_tool build-train-pkl \
+  --pairs_csv outputs/pairs.csv \
+  --graph_root outputs/graphs \
+  --output_pkl outputs/train.pkl \
+  --dataset simulated1
+```
+
+### 5) Train (MoCo)
+
+Note: this stage requires `torch` and `torch-geometric`, which are NOT installed by
+`pip install grasp-tool`. Install them via conda first.
+
+```bash
+python -m grasp_tool train-moco \
+  --dataset simulated1 \
+  --pkl outputs/train.pkl \
+  --js 0 \
+  --n 20 \
+  --m 10 \
+  --num_epoch 300 \
+  --batch_size 64 \
+  --cuda_device 0 \
+  --output_dir outputs/embeddings
+```
+
+If you want to use JS for positive sampling:
+
+```bash
+python -m grasp_tool train-moco \
+  --dataset simulated1 \
+  --pkl outputs/train.pkl \
+  --js 1 \
+  --js_file outputs/portrait/js_distances_*.csv
+```
+
+## Outputs
+
+This section summarizes what each stage writes to disk.
+
+### register
+
+Command:
+
+```bash
+python -m grasp_tool register --pkl_file <raw.pkl> --output_pkl <registered.pkl>
+```
+
+Output (`<registered.pkl>` is a dict with):
+
+- `df_registered` (DataFrame): normalized transcript coordinates; contains at least `cell,gene,x_c_s,y_c_s` and also keeps original columns.
+- `nuclear_boundary_df_registered` (DataFrame): normalized nucleus boundary points per cell (contains `cell,x_c_s,y_c_s` plus intermediate columns).
+- `cell_radii` (dict): per-cell radius used by downstream partitioning.
+- `cell_nuclear_stats` (DataFrame): per-cell nucleus exceed stats (`exceed_percent/exceed_count/num_nuclear_points`).
+- `meta` (dict): run metadata.
+
+### cellplot (optional)
+
+Command:
+
+```bash
+python -m grasp_tool cellplot --mode <raw-cell|registered-gene> --pkl <input.pkl> --output_dir <dir>
+```
+
+Output (`<dir>`):
+
+- Raw mode (`--mode raw-cell`): writes per-cell boundary plots under `1_<dataset>_raw_cell_plot/`.
+- Registered mode (`--mode registered-gene`): writes per-cell/per-gene scatter plots under `<dataset>/registered_gene/<cell>/`.
+
+### portrait (optional)
+
+Command:
+
+```bash
+python -m grasp_tool portrait --pkl_file <registered.pkl> --output_dir <dir>
+```
+
+Output (`<dir>`):
+
+- `js_distances_*.csv`: JS distance table used for positive sampling (when `train-moco --js 1`).
+
+### partition-graphs
+
+Command:
+
+```bash
+python -m grasp_tool partition-graphs --pkl <registered.pkl> --graph_root <graph_root>
+```
+
+Output directory layout (`<graph_root>`):
+
+- `<graph_root>/<cell>/<gene>_node_matrix.csv`
+- `<graph_root>/<cell>/<gene>_adj_matrix.csv`
+- `<graph_root>/<cell>/<gene>_dis_matrix.csv`
+
+These CSVs are the on-disk graph representation consumed by `build-train-pkl`.
+
+### augment-graphs
+
+Command:
+
+```bash
+python -m grasp_tool augment-graphs --graph_root <graph_root>
+```
+
+Output directory layout:
+
+- `<graph_root>/<cell>_aug/<gene>_node_matrix.csv`
+- `<graph_root>/<cell>_aug/<gene>_adj_matrix.csv`
+
+### build-train-pkl
+
+Command:
+
+```bash
+python -m grasp_tool build-train-pkl --pairs_csv <pairs.csv> --graph_root <graph_root> --output_pkl <train.pkl>
+```
+
+Output (`<train.pkl>` is a dict with):
+
+- `original_graphs`: list of `torch_geometric.data.Data`
+- `augmented_graphs`: list of `torch_geometric.data.Data`
+- `gene_labels`, `cell_labels`: aligned labels for each graph
+- `meta`: dataset tag + graph parameters + `pairs.csv` path
+
+### train-moco
+
+Command:
+
+```bash
+python -m grasp_tool train-moco --dataset <name> --pkl <train.pkl> --output_dir <out_root>
+```
+
+Output directory layout (`<out_root>`):
+
+- `<out_root>/<run_id>/1_training_config.json`: the full resolved args snapshot
+- `<out_root>/<run_id>/epoch{E}_lr{LR}_embedding.csv`: **main representation output**
+  - columns: `feature_1..feature_d, cell, gene`
+- checkpoints:
+  - `<out_root>/<run_id>/epoch_{E}_lr_{LR}_checkpoint.pth`
+  - (optional) `<out_root>/<run_id>/best_model_epoch_{E}_lr_{LR}.pth`
+- best summary (only when clustering is enabled via `--num_clusters`):
+  - `<out_root>/<run_id>/best_metrics_lr{LR}.json`
+  - `<out_root>/<run_id>/best_{vis_method}_{cluster_method}_lr{LR}.png`
+- evaluation / visualization (from `grasp_tool/gnn/plot_refined.py`):
+  - `<out_root>/<run_id>/epoch{E}_lr{LR}_metrics*.txt`
+  - `<out_root>/<run_id>/epoch{E}_lr{LR}_clusters*.csv`
+  - `<out_root>/<run_id>/epoch{E}_lr{LR}_visualization*.png`
+- `<out_root>/<run_id>/ALL_COMPLETED.txt`: written after all learning rates finish
+
+## Key parameters (quick reference)
+
+In general, you can always inspect the full list via:
+
+```bash
+python -m grasp_tool --help
+python -m grasp_tool <command> --help
+```
+
+### register
+
+- `--pkl_file`: input raw data dict PKL
+- `--output_pkl`: output registered PKL (will contain `df_registered`)
+- `--nc_demo`: process only first N cells (smoke test)
+- `--chunk_size`: multiprocessing chunk size (speed/memory tradeoff)
+- `--clip_to_cell`: `1` to clip nucleus to cell boundary; `0` to keep outside points
+- `--remove_outliers`: `1` to drop nucleus points exceeding boundary
+- `--epsilon`: numerical stability
+
+### cellplot
+
+- `--mode`: `raw-cell` or `registered-gene`
+- `--pkl` / `--pkl_file`: input PKL path
+- `--output_dir`: output directory root
+- `--dataset`: dataset tag used in output paths (optional)
+- `--cells`: restrict to a comma-separated subset of cells (recommended)
+- `--genes`: restrict to a comma-separated subset of genes (registered-gene only; recommended)
+- `--with_nuclear`: `1` to plot nucleus boundary if present, `0` to disable (registered-gene only)
+
+### portrait
+
+This command is a pass-through wrapper. Common knobs:
+
+- `--auto_params`: auto-select `r_min/r_max/bin_size`
+- `--use_same_r`: enforce the same `r` within each gene
+- `--max_count`, `--transcript_window`: reduce compute for large datasets
+- `--output_dir`: control where `js_distances_*.csv` is written
+
+### partition-graphs
+
+- `--pkl`: registered PKL (must contain `df_registered`)
+- `--graph_root`: output root directory
+- `--n_sectors`, `--m_rings`: partition resolution
+- `--k_neighbor`: kNN graph connectivity
+- `--cells`, `--genes`: restrict scope (smoke test)
+- `--epsilon`: boundary classification tolerance
+
+### augment-graphs
+
+- `--graph_root`: directory created by `partition-graphs`
+- `--dropout_ratio`: node dropout probability
+- `--seed`: make augmentation deterministic
+- `--angle_min`, `--angle_max`: rotation angle range (degrees)
+
+### build-train-pkl
+
+- `--pairs_csv`: CSV with columns `cell,gene`
+- `--graph_root`: directory created by `partition-graphs` (and augmented by `augment-graphs`)
+- `--output_pkl`: training PKL consumed by `train-moco`
+- `--dataset`: dataset tag stored in metadata
+- `--processes`: multiprocessing workers
+
+### train-moco
+
+This command runs the packaged training entrypoint (`grasp_tool.cli.train_moco`).
+
+- `--pkl`: training PKL built by `build-train-pkl`
+- `--output_dir`: output root directory
+- `--lrs`: learning rate list (e.g. `--lrs 0.001` or `--lrs 0.001 0.002`)
+- `--use_gradient_clipping`: `1` (default) to clip gradients, `0` to disable
+- `--gradient_clip_norm`: max norm for gradient clipping
+- `--js` + `--js_file`: use JS distance for positive sampling
+- `--n`, `--m`: must match partition settings
+- `--seed`: reproducibility
+- `--num_epoch`, `--batch_size`: training schedule
+- `--cuda_device`: GPU index
+- `--num_clusters`: affects clustering evaluation (for very small datasets, set it <= num graphs)
+
+## Reproducibility tips
+
+- Always record: `n_sectors/m_rings/k_neighbor/dropout_ratio/seed` and the exact `pairs.csv`.
+- Prefer writing all outputs under `outputs/` (or a dedicated run directory).
+- For large runs, use tmux/screen; training can be slow due to evaluation + visualization.
+

grasp_tool-0.1.0.dist-info/RECORD

@@ -0,0 +1,22 @@
+grasp_tool/__init__.py,sha256=axzacuCBBe4OBgW8dimPR2QXQESPlG1kNBlA0sOOzFc,472
+grasp_tool/__main__.py,sha256=1huT110kHaY7zMU9u3_uu6CizrIXLPp3T2Mz1rc9vl8,100
+grasp_tool/cli/__init__.py,sha256=WH-ValcU5Su13EAryqlacwK_RXaZSA5rmkBK1rcR5N8,47
+grasp_tool/cli/main.py,sha256=k3jnp9Tx29-MXJPSUgg9yaUMSFr2K_tUZZEiyU_Bv9I,24896
+grasp_tool/cli/train_moco.py,sha256=m6cNzIaEdvljl1yhV1b-_jZ9SCwms_S3Wi36GJSuM3A,26216
+grasp_tool/gnn/__init__.py,sha256=t0od8HkPC3xPytTs-us8I6Lswu87dKUl4xJOY9GXZB4,77
+grasp_tool/gnn/embedding.py,sha256=t_r0O8ztHeUfxdETY4c3DjvsgzyA4qPJJ4H3_99NKu4,5862
+grasp_tool/gnn/gat_moco_final.py,sha256=bBpKlgE-zDtp9o-Z6N-HcLH9FFaumZVrZrIyud2m4QY,37514
+grasp_tool/gnn/graphloader.py,sha256=wOdsUtd9yIRznjSNlj2OSXWweVliOACYrw2zoRajuZ0,69195
+grasp_tool/gnn/plot_refined.py,sha256=JmG70jWu6mxs93zGcNhLPX9XOhVgZoSEkD3qDtdgzB8,54744
+grasp_tool/preprocessing/__init__.py,sha256=kEW0xCQogpjSTbacCWf2dJ1DVQ3riJbwGiFBm1G6TxU,66
+grasp_tool/preprocessing/augumentation.py,sha256=yo0G6jBBThzZcxKw1wKgKLquLg0qgbX0zD_riAR5Ie0,2376
+grasp_tool/preprocessing/cellplot.py,sha256=KYrlueC_XqvJ9Ko6veoFsDvFVLSWNATCW0tTJgnjjOs,18684
+grasp_tool/preprocessing/filter.py,sha256=4A9iMxKEtHgedfnMD-RC2laRyii9zbx8LMZggNGw6TQ,5527
+grasp_tool/preprocessing/network.py,sha256=1u5Iy8mnFqMDH6guaJm2fGzRwPlaJeV5jLFNerEbA68,2883
+grasp_tool/preprocessing/partition.py,sha256=SLDZzK3Xmv22gsIZOc5yGF7rXvDb5uFQMQOV8yCWb3M,23558
+grasp_tool/preprocessing/portrait.py,sha256=W6HBddomkFFGMtKisfff1cMlWWTiIsg56hLpq7ei58A,77877
+grasp_tool/preprocessing/register.py,sha256=1BLbUF7VrJnntoCZAk82kFAnf5e0mu7ZnmC4hzSptlA,39694
+grasp_tool-0.1.0.dist-info/METADATA,sha256=8OdNEdu6qHQKsvf9Zz_2l_60L-OXgy-DkEuuR_NKg6E,15058
+grasp_tool-0.1.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+grasp_tool-0.1.0.dist-info/entry_points.txt,sha256=mo5VfxgHDhsAt6CHIQsfFbLtEn9xl3eN-_LzCBAYxWQ,55
+grasp_tool-0.1.0.dist-info/RECORD,,

grasp_tool-0.1.0.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: poetry-core 2.1.3
+Root-Is-Purelib: true
+Tag: py3-none-any

grasp_tool-0.1.0.dist-info/entry_points.txt

@@ -0,0 +1,3 @@
+[console_scripts]
+grasp-tool=grasp_tool.cli.main:main
+