Add simple save model test (#1227)

* add save model test

jiant/utils/torch_utils.py

@@ -144,3 +144,24 @@ def get_model_for_saving(model: nn.Module) -> nn.Module:
         return model.module
     else:
         return model
+
+
+def eq_state_dicts(state_dict_1, state_dict_2):
+    """Checks if the model weights in state_dict_1 and state_dict_2 are equal.
+
+    Args:
+        state_dict_1 (dict): state_dict of a PyTorch model
+        state_dict_2 (dict): state_dict of a PyTorch model
+
+    Requires:
+        state_dict_1 and state_dict_2 to be from the same model
+
+    Returns:
+        bool: Returns True if all model weights are equal in state_dict_1 and state_dict_2
+    """
+    for key_item_1, key_item_2 in zip(state_dict_1.items(), state_dict_2.items()):
+        if torch.equal(key_item_1[1], key_item_2[1]):
+            pass
+        else:
+            return False
+    return True

tests/proj/simple/test_runscript.py

@@ -1,9 +1,11 @@
 import os
 import pytest
+import torch
 
 import jiant.utils.python.io as py_io
 from jiant.proj.simple import runscript as run
 import jiant.scripts.download_data.runscript as downloader
+import jiant.utils.torch_utils as torch_utils
 
 
 @pytest.mark.parametrize("task_name", ["copa"])
@@ -29,3 +31,83 @@ def test_simple_runscript(tmpdir, task_name, model_type):
 
     val_metrics = py_io.read_json(os.path.join(exp_dir, "runs", RUN_NAME, "val_metrics.json"))
     assert val_metrics["aggregated"] > 0
+
+
+@pytest.mark.gpu
+@pytest.mark.parametrize("task_name", ["copa"])
+@pytest.mark.parametrize("model_type", ["roberta-large"])
+def test_simple_runscript_save(tmpdir, task_name, model_type):
+    run_name = f"{test_simple_runscript.__name__}_{task_name}_{model_type}_save"
+    data_dir = str(tmpdir.mkdir("data"))
+    exp_dir = str(tmpdir.mkdir("exp"))
+
+    downloader.download_data([task_name], data_dir)
+
+    args = run.RunConfiguration(
+        run_name=run_name,
+        exp_dir=exp_dir,
+        data_dir=data_dir,
+        model_type=model_type,
+        tasks=task_name,
+        max_steps=1,
+        train_batch_size=32,
+        do_save=True,
+        eval_every_steps=10,
+        learning_rate=0.01,
+        num_train_epochs=5,
+    )
+    run.run_simple(args)
+
+    # check best_model and last_model exist
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.p"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.metadata.json"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.p"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.metadata.json"))
+
+    # assert best_model not equal to last_model
+    best_model_weights = torch.load(
+        os.path.join(exp_dir, "runs", run_name, "best_model.p"), map_location=torch.device("cpu")
+    )
+    last_model_weights = torch.load(
+        os.path.join(exp_dir, "runs", run_name, "last_model.p"), map_location=torch.device("cpu")
+    )
+    assert not torch_utils.eq_state_dicts(best_model_weights, last_model_weights)
+
+    run_name = f"{test_simple_runscript.__name__}_{task_name}_{model_type}_save_best"
+    args = run.RunConfiguration(
+        run_name=run_name,
+        exp_dir=exp_dir,
+        data_dir=data_dir,
+        model_type=model_type,
+        tasks=task_name,
+        max_steps=1,
+        train_batch_size=16,
+        do_save_best=True,
+    )
+    run.run_simple(args)
+
+    # check only best_model saved
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.p"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.metadata.json"))
+    assert not os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.p"))
+    assert not os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.metadata.json"))
+
+    # check output last model
+    run_name = f"{test_simple_runscript.__name__}_{task_name}_{model_type}_save_last"
+    args = run.RunConfiguration(
+        run_name=run_name,
+        exp_dir=exp_dir,
+        data_dir=data_dir,
+        model_type=model_type,
+        tasks=task_name,
+        max_steps=1,
+        train_batch_size=16,
+        do_save_last=True,
+    )
+    run.run_simple(args)
+
+    # check only last_model saved
+    assert not os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.p"))
+    assert not os.path.exists(os.path.join(exp_dir, "runs", run_name, "best_model.metadata.json"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.p"))
+    assert os.path.exists(os.path.join(exp_dir, "runs", run_name, "last_model.metadata.json"))