def test_model_graph_dimensions(X_sequence, mod_attr, graph_attr):
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab, max_iter=1)
mod.fit(X_train, y_train)
mod_attr_val = getattr(mod, mod_attr)
graph_attr_val = getattr(mod.model.rnn.rnn, graph_attr)
assert mod_attr_val == graph_attr_val
def fit_elmo_rnn(X, y):
mod = TorchRNNClassifier(
vocab=[],
max_iter=50,
use_embedding=False)
mod.fit(X, y)
return mod
def test_predict_functions_honor_device(X_sequence, func):
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab, max_iter=2)
mod.fit(X_train, y_train)
prediction_func = getattr(mod, func)
with pytest.raises(RuntimeError):
prediction_func(X_test, device="FAKE_DEVICE")
def fit_hf_rnn(X, y):
mod = TorchRNNClassifier(
vocab=[],
max_iter=50,
hidden_dim=50,
use_embedding=False) # Pass in the BERT hidden states directly!
mod.fit(X, y)
return mod
def test_model(X_sequence):
"""Just makes sure that this code will run; it doesn't check that
it is creating good models.
"""
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab=vocab, max_iter=100)
mod.fit(X_train, y_train)
mod.predict(X_test)
mod.predict_proba(X_test)
def test_predict_restores_device(X_sequence, func):
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab, max_iter=2)
mod.fit(X_train, y_train)
current_device = mod.device
assert current_device != torch.device("cpu:0")
prediction_func = getattr(mod, func)
prediction_func(X_test, device="cpu:0")
assert mod.device == current_device
def test_torch_rnn_classifier_save_load(X_sequence):
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab=vocab, max_iter=2)
mod.fit(X_train, y_train)
mod.predict(X_test)
with tempfile.NamedTemporaryFile(mode='wb') as f:
name = f.name
mod.to_pickle(name)
mod2 = TorchRNNClassifier.from_pickle(name)
mod2.predict(X_test)
mod2.fit(X_test, y_test)
def test_embedding_update_control(X_sequence, freeze, outcome):
X_train, X_test, y_train, y_test, vocab = X_sequence
embed_dim = 5
embedding = np.ones((len(vocab), embed_dim))
mod = TorchRNNClassifier(vocab,
max_iter=10,
embedding=embedding,
freeze_embedding=freeze)
mod.fit(X_train, y_train)
graph_emb = mod.model.rnn.embedding.weight.detach().cpu().numpy()
assert np.array_equal(embedding, graph_emb) == outcome
def test_pretrained_embedding(X_sequence):
X_train, X_test, y_train, y_test, vocab = X_sequence
embed_dim = 5
embedding = np.ones((len(vocab), embed_dim))
mod = TorchRNNClassifier(vocab,
max_iter=1,
embedding=embedding,
freeze_embedding=True)
mod.fit(X_train, y_train)
graph_emb = mod.model.rnn.embedding.weight.detach().cpu().numpy()
assert np.array_equal(embedding, graph_emb)
def test_cheese_disease(cheese_disease_dataset):
vocab = cheese_disease_dataset['vocab']
X_train = cheese_disease_dataset['X_train']
y_train = cheese_disease_dataset['y_train']
mod = TorchRNNClassifier(vocab=vocab,
embed_dim=50,
hidden_dim=50,
max_iter=200)
mod.fit(X_train, y_train)
X_test = cheese_disease_dataset['X_train']
y_test = cheese_disease_dataset['y_train']
pred = mod.predict(X_test)
acc = accuracy_score(y_test, pred)
assert acc > 0.80
def test_simple_example_params(X_sequence, param, expected):
X_train, X_test, y_train, y_test, vocab = X_sequence
mod = TorchRNNClassifier(vocab, **{param: expected})
if param == "use_embedding" and expected == False:
embedding = np.random.uniform(
low=-1.0, high=1.0, size=(len(vocab), 60))
X_train = [[embedding[vocab.index(w)] for w in ex] for ex in X_train]
X_test = [[embedding[vocab.index(w)] for w in ex] for ex in X_test]
mod.fit(X_train, y_train)
preds = mod.predict(X_test)
acc = accuracy_score(y_test, preds)
if not (param == "max_iter" and expected == 0):
assert acc >= 0.60
def fit_rnn_classifier(X, y):
sst_glove_vocab = get_vocab(X, n_words=10)
# sst_glove_vocab = get_vocab(X, n_words=10000)
mod = TorchRNNClassifier(
sst_glove_vocab,
eta=0.05,
embedding=None,
batch_size=1000,
embed_dim=50,
hidden_dim=50,
max_iter=5,
l2_strength=0.001,
bidirectional=True,
hidden_activation=nn.ReLU())
mod.fit(X, y)
return mod
def fit_simple_chained_rnn(X, y):
vocab = utils.get_vocab(X, n_words=10000)
mod = TorchRNNClassifier(vocab, hidden_dim=50, max_iter=10)
mod.fit(X, y)
return mod
class RNN_Classifier:
'''
Modified torch rnn classifier wrapper class for initial fitting and then fine tuning of weights.
'''
def __init__(self, sent140_train_X_list, sent140_dev_X_list, sent140_train_Y, sent140_dev_Y, sent140_train_embedding, sent140_train_glove_vocab, emoji_train_X_list, emoji_dev_X_list, emoji_test_X_list, emoji_train_Y, emoji_dev_Y, emoji_test_Y, sent140_emoji_train_embedding, sent140_emoji_train_glove_vocab, emojiless_train_X_list, emojiless_dev_X_list, emojiless_test_X_list, emojiless_train_Y, emojiless_dev_Y, emojiless_test_Y, sent140_emojiless_train_embedding, sent140_emojiless_train_glove_vocab, testing):
'''
Pass in initial data for fitting to constructor. Later adding passing logisitic regression
parameters into constructor.
'''
self.testing = testing
self.sent140_train_X = sent140_train_X_list
self.sent140_train_Y = sent140_train_Y
self.sent140_dev_X = sent140_dev_X_list
self.sent140_dev_Y = sent140_dev_Y
self.emoji_train_X = emoji_train_X_list
self.emoji_train_Y = emoji_train_Y
self.emoji_dev_X = emoji_dev_X_list
self.emoji_dev_Y = emoji_dev_Y
if self.testing:
self.emoji_test_X = emoji_test_X_list
self.emoji_test_Y = emoji_test_Y
self.emojiless_train_X = emojiless_train_X_list
self.emojiless_train_Y = emojiless_train_Y
self.emojiless_dev_X = emojiless_dev_X_list
self.emojiless_dev_Y = emojiless_dev_Y
if self.testing:
self.emojiless_test_X = emojiless_test_X_list
self.emojiless_test_Y = emojiless_test_Y
# embeddings and vocabs
self.sent140_train_embedding = sent140_train_embedding
self.sent140_train_glove_vocab = sent140_train_glove_vocab
self.sent140_emoji_train_embedding = sent140_emoji_train_embedding
self.sent140_emoji_train_glove_vocab = sent140_emoji_train_glove_vocab
self.sent140_emojiless_train_embedding = sent140_emojiless_train_embedding
self.sent140_emojiless_train_glove_vocab = sent140_emojiless_train_glove_vocab
# pass in model parameters for to constructor?
def run_sent140(self):
'''
Trained on sent140, predict on emoji
Report score on sent 140 too, just because it's intersting
'''
# model
#self.model_sent140 = TorchRNNClassifier(self.sent140_train_glove_vocab, embedding=self.sent140_train_embedding, bidirectional=True)
self.model_sent140 = TorchRNNClassifier(self.sent140_train_glove_vocab, embedding=self.sent140_train_embedding)
# train
self.model_sent140.fit(self.sent140_train_X, self.sent140_train_Y)
# test on sent140
#sent140_train_preds = self.model_sent140.predict(self.sent140_train_X)
#sent140_dev_preds = self.model_sent140.predict(self.sent140_dev_X)
# test on emoji
emoji_train_preds = self.model_sent140.predict(self.emoji_train_X)
emoji_dev_preds = self.model_sent140.predict(self.emoji_dev_X)
if self.testing:
emoji_test_preds = self.model_sent140.predict(self.emoji_test_X)
else:
emoji_test_preds = None
#return (sent140_train_preds, sent140_dev_preds, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
return (None, None, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
def run_sent140_emojiless(self):
'''
Trained on sent140, fine-tuned on emojiless, predict on emoji
Report score on sent 140 too, just because it's intersting
'''
# model
#self.model_sent140_emojiless = TorchRNNClassifier(self.sent140_emojiless_train_glove_vocab, embedding=self.sent140_emojiless_train_embedding, bidirectional=True)
self.model_sent140_emojiless = TorchRNNClassifier(self.sent140_emojiless_train_glove_vocab, embedding=self.sent140_emojiless_train_embedding)
# train
# combine features
combined_train_X = self.sent140_train_X + self.emojiless_train_X
combined_train_Y = self.sent140_train_Y + self.emojiless_train_Y
self.model_sent140_emojiless.fit(combined_train_X, combined_train_Y)
# test on sent140
#sent140_train_preds = self.model_sent140_emojiless.predict(self.sent140_train_X)
#sent140_dev_preds = self.model_sent140_emojiless.predict(self.sent140_dev_X)
# test on emoji
emoji_train_preds = self.model_sent140_emojiless.predict(self.emoji_train_X)
emoji_dev_preds = self.model_sent140_emojiless.predict(self.emoji_dev_X)
if self.testing:
emoji_test_preds = self.model_sent140_emojiless.predict(self.emoji_test_X)
else:
emoji_test_preds = None
#return (sent140_train_preds, sent140_dev_preds, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
return (None, None, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
def run_sent140_emoji(self):
'''
Trained on sent140, fine-tuned on emoji, predict on emoji
Report score on sent 140 too, just because it's intersting
'''
# model
#self.model_sent140_emoji = TorchRNNClassifier(self.sent140_emoji_train_glove_vocab, embedding=self.sent140_emoji_train_embedding, bidirectional=True)
self.model_sent140_emoji = TorchRNNClassifier(self.sent140_emoji_train_glove_vocab, embedding=self.sent140_emoji_train_embedding)
# train
combined_train_X = self.sent140_train_X + self.emoji_train_X
combined_train_Y = self.sent140_train_Y + self.emoji_train_Y
self.model_sent140_emoji.fit(combined_train_X, combined_train_Y)
# test on sent140
#sent140_train_preds = self.model_sent140_emoji.predict(self.sent140_train_X)
#sent140_dev_preds = self.model_sent140_emoji.predict(self.sent140_dev_X)
# test on emoji
emoji_train_preds = self.model_sent140_emoji.predict(self.emoji_train_X)
emoji_dev_preds = self.model_sent140_emoji.predict(self.emoji_dev_X)
if self.testing:
emoji_test_preds = self.model_sent140_emoji.predict(self.emoji_test_X)
else:
emoji_test_preds = None
#return (sent140_train_preds, sent140_dev_preds, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
return (None, None, emoji_train_preds, emoji_dev_preds, emoji_test_preds)
