Release/rc0.1

.gitignore

@@ -107,10 +107,12 @@ venv.bak/
 .vscode/*
 *.code-workspace
 
+# JetBrains IDEA Template
+.idea
+
 # Local storages
-data/*
+datasets/*/
 models/*
 
 # Log files
 *.log
-*.txt

datasets/README.md

@@ -0,0 +1,6 @@
+Using built-in datasets
+=======================
+
+ZeroShotEval is provided with 2 build-in datasest:
+ - AWA2
+ - CUB

configs/BASE-CADA-VAE.yaml → experiment/configs/BASE-CADA-VAE.yaml

@@ -12,4 +12,4 @@ CLS:
     AMSGRAD: True
 DATA:
   FEAT_EMB:
-    PATH: "data/CUB/resnet101/"
+    PATH: "datasets/CUB/resnet101/"

experiment/run_embedding_extractor.py

@@ -0,0 +1,20 @@
+"""
+Script that extracts embeddings from raw data (images, texts) and saves it to
+disk for further Zero-Shot experiments. Run this before
+`run_zeroshot_experiment.py` to generate data for training.
+"""
+
+if __name__ == '__main__':
+    # STEP 0 - DATA PREPARATION
+    # (DO IT MANUALLY BEFORE RUNNING THIS SCRIPT)
+    # Prepare dataset CSV with file names to use for raw data loading.
+    # --------------------------------------------------------------------------
+
+    # STEP 1 - EMBEDDINGS EXTRACTION
+    # Extract embeddings from raw data with specified NN and save it to
+    # the disk. Note that embedding extractor can load data as a whole or
+    # using iterators (to fit into RAM). This step also creates a CSV with
+    # paths to all embedding files.
+    # --------------------------------------------------------------------------
+
+    pass

experiment/run_zeroshot_experiment.py

@@ -0,0 +1,90 @@
+"""Main script-launcher for training of ZSL models."""
+
+import os
+from argparse import Namespace
+from typing import Callable, Tuple
+
+from fvcore.common.config import CfgNode
+from torch.nn import Module
+
+from zeroshoteval.evaluation.classification import classification_procedure
+from zeroshoteval.utils.defaults import default_setup
+from zeroshoteval.utils.parser import load_config, parse_args
+from zeroshoteval.zeroshotnets.build import (
+    build_zeroshot_train,
+    build_zeroshot_inference
+)
+from zeroshoteval.zeroshotnets.cada_vae.cada_vae_inference import CADA_VAE_inference_procedure
+
+# os.chdir("../")
+
+
+def experiment(cfg: CfgNode) -> None:
+    """
+    Start single experiment with the specified configs. Note that this procedure
+    requires data embeddings, extracted previously. If you don't have the
+    extracted embeddings, use the `run_embedding_extractor.py` script.
+
+    Args:
+        cfg(CfgNode): Configs. Details can be found in
+                      zeroshoteval/config/defaults.py
+    """
+
+    # MODEL TRAINING
+    # ==============
+    # Pass extracted embeddings to a zero-shot neural network to train it as an
+    # embedding extractor.
+    # --------------------------------------------------------------------------
+
+    # Embeddings loading from disk
+    # TODO: Replace ModalitiesEmbeddingDataset with loader from separate files
+
+    # Get training procedure function from registry
+    train_procedure: Callable[[CfgNode], Module] = build_zeroshot_train(cfg)
+
+    # Training
+    zsl_model: Module = train_procedure(cfg)
+
+    # MODEL INFERENCE
+    # ===============
+    # Apply trained model to test data and (similar to the prev step) extract
+    # zero-shot embeddings.
+    # --------------------------------------------------------------------------
+
+    # TODO: replace CADA-VAE with general model
+
+    # Get inference procedure function from registry
+    inference_procedure: Callable[[CfgNode, Module], Tuple[Tuple, Tuple]] = build_zeroshot_inference(cfg)
+
+    # Inference
+    train_data, test_data = inference_procedure(cfg, zsl_model)
+
+    # MODEL EVALUATION
+    # ================
+    # Pass extracted zero-shot embeddings to one of evaluation tasks (
+    # classification, clustering, verification, etc.)
+    # --------------------------------------------------------------------------
+
+    # TODO: replace classification procedure with more general evaluation calling
+    classification_procedure(cfg, train_data, test_data)
+
+
+def main() -> None:
+    # Parse arguments
+    experiment_options: Namespace = parse_args()
+
+    # Load default configuration and merge with specific configs from args
+    cfg: CfgNode = load_config(experiment_options)
+
+    # Freeze current config to avoid arbitrary changes
+    cfg.freeze()
+
+    # Perform some basic common setups at the beginning of a job
+    default_setup(cfg, experiment_options)
+
+    # Start the experiment
+    experiment(cfg)
+
+
+if __name__ == "__main__":
+    main()