def implicit_model(train_val_implicit_data, gpu_count):
train, val = train_val_implicit_data
model = MatrixFactorizationModel(train=train,
val=val,
embedding_dim=10,
lr=1e-1)
model_trainer = CollieTrainer(model=model,
gpus=gpu_count,
max_epochs=10,
deterministic=True,
logger=False,
checkpoint_callback=False)
model_trainer.fit(model)
model.freeze()
return model
def test_instantiation_of_sparse_model_with_weight_decay(
train_val_implicit_data, capfd):
train, val = train_val_implicit_data
model_1 = MatrixFactorizationModel(train=train,
val=val,
sparse=False,
weight_decay=100)
assert model_1.hparams.weight_decay == 100
with pytest.warns(UserWarning):
model_2 = MatrixFactorizationModel(train=train,
val=val,
sparse=True,
weight_decay=100)
assert model_2.hparams.weight_decay == 0
def test_loading_and_saving_implicit_model(implicit_model,
untrained_implicit_model, tmpdir):
expected = implicit_model.get_item_predictions(user_id=42,
unseen_items_only=False)
# set up TemporaryDirectory for writing and reading all files in this test
temp_dir_name = str(tmpdir)
save_model_path = os.path.join(temp_dir_name, 'test_mf_model_save.pth')
implicit_model.save_model(save_model_path)
loaded_implicit_model = MatrixFactorizationModel(
load_model_path=save_model_path)
actual = loaded_implicit_model.get_item_predictions(
user_id=42, unseen_items_only=False)
assert expected.equals(actual)
# now, test that this is not equal to a randomly initialized model's output
new_preds = untrained_implicit_model.get_item_predictions(
user_id=42, unseen_items_only=False)
assert not expected.equals(new_preds)
def test_okay_mismatched_train_and_val_loaders(train_val_implicit_data):
train, val = train_val_implicit_data
train = copy.copy(train)
val = copy.copy(val)
train.num_negative_samples = 2
val.num_negative_samples = 3
model = MatrixFactorizationModel(train=train, val=val)
trainer = CollieTrainer(model=model,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer.fit(model)
def test_instantiation_of_model_loss(train_val_implicit_data):
train, val = train_val_implicit_data
train = copy.copy(train)
val = copy.copy(val)
train.num_negative_samples = 1
val.num_negative_samples = 1
model_1 = MatrixFactorizationModel(train=train, val=val, loss='hinge')
assert model_1.loss_function == hinge_loss
model_2 = MatrixFactorizationModel(train=train, val=val, loss='bpr')
assert model_2.loss_function == bpr_loss
with pytest.raises(ValueError):
MatrixFactorizationModel(train=train, val=val, loss='warp')
train.num_negative_samples = 2
val.num_negative_samples = 2
model_4 = MatrixFactorizationModel(train=train, val=val, loss='hinge')
assert model_4.loss_function == adaptive_hinge_loss
model_5 = MatrixFactorizationModel(train=train, val=val, loss='bpr')
assert model_5.loss_function == adaptive_bpr_loss
model_6 = MatrixFactorizationModel(train=train, val=val, loss='warp')
assert model_6.loss_function == warp_loss
def custom_loss_function(*args, **kwargs):
return 42
model_7 = MatrixFactorizationModel(train=train,
val=val,
loss=custom_loss_function)
assert model_7.loss_function == custom_loss_function
with pytest.raises(ValueError):
MatrixFactorizationModel(train=train, val=val, loss='nonexistent_loss')
def test_mismatched_train_and_val_loaders(train_val_implicit_data,
change_to_make):
train, val = train_val_implicit_data
train = copy.copy(train)
val = copy.copy(val)
expected_error = AssertionError
if change_to_make == 'num_users':
train.num_users = 3
val.num_users = 2
elif change_to_make == 'num_items':
train.num_items = 1
val.num_items = 2
elif change_to_make == 'num_negative_samples':
train.num_negative_samples = 1
val.num_negative_samples = 2
elif change_to_make == 'bad_train_num_negative_samples':
train.num_negative_samples = 0
expected_error = ValueError
with pytest.raises(expected_error):
MatrixFactorizationModel(train=train, val=val)
def models_trained_for_one_step(request,
train_val_implicit_data,
movielens_metadata_df,
movielens_implicit_df,
train_val_implicit_pandas_data,
gpu_count):
train, val = train_val_implicit_data
if request.param == 'mf_hdf5':
# create, fit, and return the model all at once so we can close the HDF5 file
train_pandas_df, val_pandas_df = train_val_implicit_pandas_data
with tempfile.TemporaryDirectory() as temp_dir:
pandas_df_to_hdf5(df=train_pandas_df,
out_path=os.path.join(temp_dir, 'train.h5'),
key='interactions')
pandas_df_to_hdf5(df=val_pandas_df,
out_path=os.path.join(temp_dir, 'val.h5'),
key='interactions')
train_loader = HDF5InteractionsDataLoader(hdf5_path=os.path.join(temp_dir, 'train.h5'),
user_col='user_id',
item_col='item_id',
num_users=train.num_users,
num_items=train.num_items,
batch_size=1024,
shuffle=True)
val_loader = HDF5InteractionsDataLoader(hdf5_path=os.path.join(temp_dir, 'val.h5'),
user_col='user_id',
item_col='item_id',
num_users=val.num_users,
num_items=val.num_items,
batch_size=1024,
shuffle=False)
model = MatrixFactorizationModel(train=train_loader,
val=val_loader,
embedding_dim=15,
dropout_p=0.1,
lr=1e-1,
bias_lr=1e-2,
optimizer='adam',
bias_optimizer='sgd',
weight_decay=1e-7,
loss='bpr',
sparse=False)
model_trainer = CollieTrainer(model=model,
gpus=gpu_count,
max_steps=1,
deterministic=True,
logger=False,
checkpoint_callback=False)
model_trainer.fit(model)
model.freeze()
return model
elif request.param == 'sparse_mf':
model = MatrixFactorizationModel(train=train,
val=val,
embedding_dim=15,
dropout_p=0.1,
lr=1e-1,
bias_lr=1e-2,
optimizer='sparse_adam',
bias_optimizer='sgd',
weight_decay=0,
loss='hinge',
sparse=True)
elif request.param == 'mf_no_val':
model = MatrixFactorizationModel(train=train, val=None)
elif request.param == 'mf_non_approximate' or request.param == 'mf_approximate':
if request.param == 'mf_non_approximate':
train_loader = InteractionsDataLoader(interactions=train, batch_size=1024, shuffle=True)
val_loader = InteractionsDataLoader(interactions=val, batch_size=1024, shuffle=False)
else:
train_loader = ApproximateNegativeSamplingInteractionsDataLoader(interactions=train,
batch_size=1024,
shuffle=True)
val_loader = ApproximateNegativeSamplingInteractionsDataLoader(interactions=val,
batch_size=1024,
shuffle=False)
model = MatrixFactorizationModel(train=train_loader,
val=val_loader,
embedding_dim=15,
dropout_p=0.1,
lr=1e-1,
bias_lr=1e-2,
optimizer='adam',
bias_optimizer='sgd',
weight_decay=1e-7,
loss='bpr',
sparse=False)
elif request.param == 'mf_with_y_range':
model = MatrixFactorizationModel(train=train,
val=val,
y_range=(0, 4))
elif request.param == 'nonlinear_mf':
model = NonlinearMatrixFactorizationModel(train=train,
val=val,
user_embedding_dim=15,
item_embedding_dim=15,
user_dense_layers_dims=[15, 10],
item_dense_layers_dims=[15, 10],
embedding_dropout_p=0.05,
dense_dropout_p=0.1,
lr=1e-1,
bias_lr=1e-2,
optimizer='adam',
bias_optimizer='sgd',
weight_decay=1e-7,
loss='bpr')
elif request.param == 'nonlinear_mf_with_y_range':
model = NonlinearMatrixFactorizationModel(train=train,
val=val,
y_range=(0, 4))
elif request.param == 'neucf':
model = NeuralCollaborativeFiltering(train=train,
val=val,
embedding_dim=10,
num_layers=1,
dropout_p=0.1,
lr=1e-3,
weight_decay=0.,
optimizer='adam',
loss='adaptive')
elif request.param == 'neucf_sigmoid':
model = NeuralCollaborativeFiltering(train=train,
val=val,
final_layer='sigmoid')
elif request.param == 'neucf_relu':
model = NeuralCollaborativeFiltering(train=train,
val=val,
final_layer='relu')
elif request.param == 'neucf_leaky_rulu':
model = NeuralCollaborativeFiltering(train=train,
val=val,
final_layer='leaky_relu')
elif request.param == 'neucf_custom':
model = NeuralCollaborativeFiltering(train=train,
val=val,
final_layer=torch.tanh)
elif (
request.param == 'hybrid_pretrained' or request.param == 'hybrid_pretrained_metadata_layers'
):
implicit_model = MatrixFactorizationModel(train=train,
val=val,
embedding_dim=10,
lr=1e-1,
optimizer='adam')
implicit_model_trainer = CollieTrainer(model=implicit_model,
gpus=gpu_count,
max_steps=1,
deterministic=True,
logger=False,
checkpoint_callback=False)
implicit_model_trainer.fit(implicit_model)
implicit_model.freeze()
genres = (
torch.tensor(movielens_metadata_df[
[c for c in movielens_metadata_df.columns if 'genre' in c]
].values)
.topk(1)
.indices
.view(-1)
)
if request.param == 'hybrid_pretrained_metadata_layers':
metadata_layers_dims = [16, 12]
else:
metadata_layers_dims = None
model_frozen = HybridPretrainedModel(train=train,
val=val,
item_metadata=movielens_metadata_df,
trained_model=implicit_model,
metadata_layers_dims=metadata_layers_dims,
freeze_embeddings=True,
dropout_p=0.15,
loss='warp',
lr=.01,
optimizer=torch.optim.Adam,
metadata_for_loss={'genre': genres},
metadata_for_loss_weights={'genre': .4},
weight_decay=0.0)
model_frozen_trainer = CollieTrainer(model=model_frozen,
gpus=gpu_count,
max_steps=1,
deterministic=True,
logger=False,
checkpoint_callback=False)
model_frozen_trainer.fit(model_frozen)
model = HybridPretrainedModel(train=train,
val=val,
item_metadata=movielens_metadata_df,
trained_model=implicit_model,
metadata_layers_dims=metadata_layers_dims,
freeze_embeddings=False,
dropout_p=0.15,
loss='bpr',
lr=1e-4,
optimizer=torch.optim.Adam,
metadata_for_loss={'genre': genres},
metadata_for_loss_weights={'genre': .4},
weight_decay=0.0)
model.load_from_hybrid_model(model_frozen)
model_trainer = CollieTrainer(model=model,
gpus=gpu_count,
max_steps=1,
deterministic=True,
logger=False,
checkpoint_callback=False)
if request.param == 'mf_no_val':
with pytest.warns(UserWarning):
model_trainer.fit(model)
else:
model_trainer.fit(model)
model.freeze()
return model
def untrained_implicit_model_no_val_data(train_val_implicit_data):
train, _ = train_val_implicit_data
model = MatrixFactorizationModel(train=train, val=None)
return model
def untrained_implicit_model(train_val_implicit_data):
train, val = train_val_implicit_data
model = MatrixFactorizationModel(train=train, val=val)
return model
def test_instantiation_of_model_optimizer(train_val_implicit_data):
train, val = train_val_implicit_data
model_1 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer=None)
trainer_1 = CollieTrainer(model=model_1,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_1.fit(model_1)
assert not isinstance(model_1.optimizers(), list)
model_1_lr_schedulers = [s['scheduler'] for s in trainer_1.lr_schedulers]
assert len(model_1_lr_schedulers) == 1
model_2 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer=None,
lr_scheduler_func=None)
trainer_2 = CollieTrainer(model=model_2,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_2.fit(model_2)
assert not isinstance(model_2.optimizers(), list)
model_2_lr_schedulers = [s['scheduler'] for s in trainer_2.lr_schedulers]
assert len(model_2_lr_schedulers) == 0
model_3 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer='infer',
bias_lr='infer')
trainer_3 = CollieTrainer(model=model_3,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_3.fit(model_3)
assert len(model_3.optimizers()) == 2
assert model_3.bias_optimizer == model_3.optimizer
assert model_3.hparams.bias_lr == model_3.hparams.lr
model_3_lr_schedulers = [s['scheduler'] for s in trainer_3.lr_schedulers]
assert len(model_3_lr_schedulers) == 2
model_4 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer='infer',
bias_lr='infer',
lr_scheduler_func=None)
trainer_4 = CollieTrainer(model=model_4,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_4.fit(model_4)
assert len(model_4.optimizers()) == 2
assert model_4.bias_optimizer == model_4.optimizer
assert model_4.hparams.bias_lr == model_4.hparams.lr
model_4_lr_schedulers = [s['scheduler'] for s in trainer_4.lr_schedulers]
assert len(model_4_lr_schedulers) == 0
model_5 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer='infer',
bias_lr=10,
lr_scheduler_func=None)
trainer_5 = CollieTrainer(model=model_5,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_5.fit(model_5)
assert len(model_5.optimizers()) == 2
assert model_5.bias_optimizer == model_5.optimizer
assert model_5.hparams.bias_lr != model_5.hparams.lr
model_5_lr_schedulers = [s['scheduler'] for s in trainer_5.lr_schedulers]
assert len(model_5_lr_schedulers) == 0
model_6 = MatrixFactorizationModel(train=train,
val=val,
optimizer='fake_optimizer')
trainer_6 = CollieTrainer(model=model_6,
logger=False,
checkpoint_callback=False,
max_steps=1)
with pytest.raises(ValueError):
trainer_6.fit(model_6)
model_7 = MatrixFactorizationModel(train=train,
val=val,
bias_optimizer='fake_optimizer')
trainer_7 = CollieTrainer(model=model_7,
logger=False,
checkpoint_callback=False,
max_steps=1)
with pytest.raises(ValueError):
trainer_7.fit(model_7)
# ``Adadelta`` accepts ``weight_decay`` parameter
model_8 = MatrixFactorizationModel(train=train,
val=val,
optimizer=torch.optim.Adadelta)
trainer_8 = CollieTrainer(model=model_8,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_8.fit(model_8)
# ``LBFGS`` does not accept ``weight_decay`` parameter
model_9 = MatrixFactorizationModel(train=train,
val=val,
optimizer=torch.optim.LBFGS,
sparse=True)
trainer_9 = CollieTrainer(model=model_9,
logger=False,
checkpoint_callback=False,
max_steps=1)
trainer_9.fit(model_9)
def run_movielens_example(epochs: int = 20, gpus: int = 0) -> None:
"""
Retrieve and split data, train and evaluate a model, and save it.
From the terminal, you can run this script with:
.. code-block:: bash
python collie_recs/movielens/run.py --epochs 20
Parameters
-------------
epochs: int
Number of epochs for model training
gpus: int
Number of gpus to train on
"""
t = Timer()
t.timecheck(' 1.0 - retrieving MovieLens 100K dataset')
df = read_movielens_df(decrement_ids=True)
t.timecheck(' 1.0 complete')
t.timecheck(' 2.0 - splitting data')
df_imp = convert_to_implicit(df)
interactions = Interactions(users=df_imp['user_id'],
items=df_imp['item_id'],
allow_missing_ids=True)
train, val, test = stratified_split(interactions, val_p=0.1, test_p=0.1)
train_loader = InteractionsDataLoader(train, batch_size=1024, shuffle=True)
val_loader = InteractionsDataLoader(val, batch_size=1024, shuffle=False)
t.timecheck(' 2.0 complete')
t.timecheck(' 3.0 - training the model')
model = MatrixFactorizationModel(train=train_loader,
val=val_loader,
dropout_p=0.05,
loss='adaptive',
lr=5e-2,
embedding_dim=10,
optimizer='adam',
weight_decay=1e-7)
trainer = CollieTrainer(
model=model,
gpus=gpus,
max_epochs=epochs,
deterministic=True,
logger=False,
checkpoint_callback=False,
callbacks=[EarlyStopping(monitor='val_loss_epoch', mode='min')],
weights_summary='full',
terminate_on_nan=True)
trainer.fit(model)
model.eval()
t.timecheck('\n 3.0 complete')
t.timecheck(' 4.0 - evaluating model')
auc_score, mrr_score, mapk_score = evaluate_in_batches([auc, mrr, mapk],
test,
model,
k=10)
print(f'AUC: {auc_score}')
print(f'MRR: {mrr_score}')
print(f'MAP@10: {mapk_score}')
t.timecheck(' 4.0 complete')
t.timecheck(' 5.0 - saving model')
absolute_data_path = DATA_PATH / 'fitted_model'
model.save_model(absolute_data_path)
t.timecheck(' 5.0 complete')