version = "0.4.2"

[

This repo implements FiP proposed in the ICML 2024 paper "A Fixed-Point Approach for Causal Generative Modeling".

FiP is a transformer-based approach to learn Structural Causal Models (SCMs) from observational data. To do so, FiP uses an equivalent formulation of SCMs that does not require Directed Acyclic Graphs (DAGs), viewed as fixed-point problems on the causally ordered variables. To infer topological orders (TOs), we propose to amortize the learning of a TO inference method on synthetically generated datasets by sequentially predicting the leaves of graphs seen during training.

We use [Poetry](https://python-poetry.org/) to manage the project dependencies, they are specified in [pyproject](pyproject.toml) file. To install poetry, run:

To install the environment, run `poetry install` in the directory of fip project.

To reproduce the results obtained in the [paper](https://arxiv.org/pdf/2404.06969), you need to generate the data. A more detailed explanation on how to generate the data can be found in [README.md](src/fip/data_generation/README.md).

To generate the [AVICI](https://arxiv.org/abs/2205.12934) synthetic data, run the following command:

This executes the [avici_data.py](src/fip/dataset_generation/avici_data.py) file to generate various datasets from the dataset distributions presented in [AVICI](https://arxiv.org/abs/2205.12934). The generated data will be saved in the `src/fip/data`.

Similarly, to generate the [CSuite](https://arxiv.org/abs/2202.02195) and the [Causal NF](https://arxiv.org/abs/2306.05415) synthetic data, run the following commands:

```

In the [launchers](src/fip/launchers) directory, we provide scripts to run the experiments reported in the paper. A more detailed explanation on how to use these files can be found in [README.md](src/fip/launchers/README.md).

To train the TO inference method on AVICI data, run the following command:

The model as well as the config file will be saved in `src/fip/outputs`.

To train FiP when the DAG is known, run the following command:

The model as well as the config file will be saved in `src/fip/outputs`. If you want to train FiP, when only the TO is known, replace `--ground_truth_case graph` with `--ground_truth_case perm`. These commands assume that the datasets have been generated and saved in `src/fip/data`.

To train FiP end-to-end, run the following command:

The model as well as the config file will be saved in `src/fip/outputs`. This command assumes that a TO inference model has been trained and saved in a directory located at `src/fip/outputs/<name_of_the_directory_containing_the_amortized_model>`. This command also assumes that the datasets have been generated and saved in `src/fip/data`.

[tool.poetry]

name = "fip"

version = "0.1.0"

description = "A Fixed-Point Approach for Causal Generative Modeling"

readme = "README.md"

authors = ["Meyer Scetbon", "Joel Jennings", "Agrin Hilmkil", "Cheng Zhang", "Chao Ma"]

packages = [

{include = "fip", from = "src"}

]

license = "MIT"

[tool.poetry.dependencies]

python = "~3.10"

causica = "0.4.1"

[build-system]

requires = ["poetry-core>=1.0.0"]

build-backend = "poetry.core.masonry.api"

seed_everything: 234

model:

class_path: fip.tasks.amortization.leaf_prediction.LeafPrediction

init_args:

learning_rate: 3e-4

weight_decay: 5e-9

d_model: 128

num_heads: 8

dim_key: 32

num_layers: 4

d_ff: 256

dropout: 0.

max_num_leaf: 100

num_to_keep_training: 10

distributed: false

elimination_type: "self"

trainer:

max_epochs: 2000

accelerator: gpu

check_val_every_n_epoch: 10

log_every_n_steps: 10

profiler: "simple"

devices: 1

accumulate_grad_batches: 1

best_checkpoint_callback:

dirpath: "./outputs/"

filename: "best_model"

save_top_k: 1

mode: "min"

monitor: "val_loss"

every_n_epochs: 1

last_checkpoint_callback:

save_last: true

save_top_k: 0 # only the last checkpoint is saved

early_stopping_callback:

monitor: "val_loss"

min_delta: 0.0001

patience: 500

verbose: False

mode: "min"

class_path: fip.data_modules.numpy_tensor_data_module.NumpyTensorDataModule

init_args:

data_dir : "fip/data/er_linear_dag_scm/total_nodes_5/seed_1/"

train_batch_size: 2000

test_batch_size: 2000

standardize: true

with_true_graph: true

split_data_noise: true

dod: false

num_workers: 0

shuffle: true

num_interventions: 0

seed_everything: 234

model:

class_path: fip.tasks.scm_learning_with_ground_truth.scm_learning_true_graph.SCMLearningTrueGraph

init_args:

lr: 1e-4

weight_decay: 1e-10

d_model: 128

dim_key: 32

num_heads: 8

d_feedforward: 128

total_nodes: 4

total_layers: 2

dropout_prob: 0.

mask_type: "none"

attn_type: "causal"

cost_type: "dot_product"

learnable_loss: false

distributed: false

trainer:

max_epochs: 1000

accelerator: gpu

devices: 1

check_val_every_n_epoch: 1

log_every_n_steps: 10

inference_mode: false

profiler: "simple"

early_stopping_callback:

monitor: "val_loss"

min_delta: 0.0001

patience: 500

verbose: False

mode: "min"

best_checkpoint_callback:

dirpath: "./outputs/"

filename: "best_model"

save_top_k: 1

mode: "min"

monitor: "val_loss"

every_n_epochs: 1

last_checkpoint_callback:

save_last: true

save_top_k: 0 # only the last checkpoint is saved

seed_everything: 5000

model:

class_path: fip.tasks.scm_learning_with_ground_truth.scm_learning_true_perm.SCMLearningTruePerm

init_args:

lr: 1e-4

weight_decay: 1e-10

d_model: 128

dim_key: 32

num_heads: 8

d_feedforward: 128

total_nodes: 2

total_layers: 2

dropout_prob: 0.

mask_type: "triang"

attn_type: "causal"

cost_type: "dot_product"

learnable_loss: false

distributed: false

trainer:

max_epochs: 1000

accelerator: gpu

devices: 1

check_val_every_n_epoch: 1

log_every_n_steps: 10

inference_mode: false

profiler: "simple"

early_stopping_callback:

monitor: "val_loss"

min_delta: 0.0001

patience: 500

verbose: False

mode: "min"

best_checkpoint_callback:

dirpath: "./outputs/"

filename: "best_model"

save_top_k: 1

mode: "min"

monitor: "val_loss"

every_n_epochs: 1

last_checkpoint_callback:

save_last: true

save_top_k: 0 # only the last checkpoint is saved

seed_everything: 5000

model:

class_path: fip.tasks.scm_learning_with_predicted_truth.scm_learning_predicted_leaf.SCMLearningPredLeaf

init_args:

lr: 1e-4

weight_decay: 1e-10

leaf_model_path: "./outputs/amortized_pred_checkpoint/leaf_predicition/best_model.ckpt"

leaf_config_path: "./outputs/amortized_pred_checkpoint/leaf_predicition/config.yaml"

d_model: 128

dim_key: 32

num_heads: 8

d_feedforward: 128

total_nodes: 4

total_layers: 2

dropout_prob: 0.

mask_type: "triang"

attn_type: "causal"

cost_type: "dot_product"

learnable_loss: false

distributed: false

trainer:

max_epochs: 1000

accelerator: gpu

devices: 1

check_val_every_n_epoch: 1

log_every_n_steps: 10

inference_mode: false

early_stopping_callback:

monitor: "val_loss"

min_delta: 0.0001

patience: 500

verbose: False

mode: "min"

best_checkpoint_callback:

dirpath: "./outputs/"

filename: "best_model"

save_top_k: 1

mode: "min"

monitor: "val_loss"

every_n_epochs: 1

last_checkpoint_callback:

save_last: true

save_top_k: 0 # only the last checkpoint is saved

class_path: fip.data_modules.synthetic_data_module.SyntheticDataModule

init_args:

sem_samplers:

class_path: fip.data_generation.sem_factory.SemSamplerFactory

init_args:

node_nums: [10]

noises: ['gaussian']

graphs: ['er', 'sf_in', 'sf_out']

funcs: ['rff']

config_gaussian:

low: 0.2

high: 2.0

config_er:

edges_per_node: [1,2,3]

config_sf:

edges_per_node: [1,2,3]

attach_power: [1.]

config_linear:

weight_low: 1.

weight_high: 3.

bias_low: -3.

bias_high: 3.

config_rff:

num_rf: 100

length_low: 7.

length_high: 10.

out_low: 10.

out_high: 20.

bias_low: -3.

bias_high: 3.

train_batch_size: 2

test_batch_size: 8

sample_dataset_size: 200

standardize: false

num_samples_used: 200

num_workers: 23

pin_memory: true

persistent_workers: true

prefetch_factor: 2

factor_epoch: 16

num_sems: 0

shuffle: true

num_interventions: 0

num_intervention_samples: 0

proportion_treatment: 0.

sample_counterfactuals: false