def test_build_dataset_vectorizing():
phi = lambda text: Counter(text.split())
split_df = sst.dev_reader(sst_home)
dataset = sst.build_dataset(split_df, phi, vectorizer=None, vectorize=True)
assert len(dataset['X']) == split_df.shape[0]
assert len(dataset['y']) == len(dataset['X'])
assert len(dataset['raw_examples']) == len(dataset['X'])
def test_build_dataset_not_vectorizing():
phi = lambda text: text
split_df = sst.dev_reader(sst_home)
dataset = sst.build_dataset(split_df,
phi,
vectorizer=None,
vectorize=False)
assert len(dataset['X']) == split_df.shape[0]
assert dataset['X'] == dataset['raw_examples']
assert len(dataset['y']) == len(dataset['X'])
def test_build_dataset_not_vectorizing():
phi = lambda tree: tree
class_func = None
reader = sst.train_reader
dataset = sst.build_dataset(reader,
phi,
class_func,
vectorizer=None,
vectorize=False)
assert len(dataset['X']) == len(list(reader()))
assert dataset['X'] == dataset['raw_examples']
assert len(dataset['y']) == len(dataset['X'])
def test_build_dataset_vectorizing():
phi = lambda tree: Counter(tree.leaves())
class_func = None
reader = sst.train_reader
dataset = sst.build_dataset(reader,
phi,
class_func,
vectorizer=None,
vectorize=True)
assert len(dataset['X']) == len(list(reader()))
assert len(dataset['y']) == len(dataset['X'])
assert len(dataset['raw_examples']) == len(dataset['X'])
def run_experiment(eta, embed, model, phrase):
print("===================================================")
print("eta: %s, embed_dim: %s, model: %s, phrase-level: %s" %
(eta, embed, model, phrase))
print("===================================================")
if embed == 100:
glove_lookup = glove_lookup_100
elif embed == 200:
glove_lookup = glove_lookup_200
elif embed == 300:
glove_lookup = glove_lookup_300
if model == "lifted":
base_model = tf_lifted_trnn.TfLiftedTreeRNNClassifier
else:
base_model = tf_trnn.TfTreeRNNClassifier
start = time.time()
train = sst.build_dataset(sst.train_reader,
lambda x: x,
sst.ternary_class_func,
vectorizer=None,
vectorize=False,
subtree_labels=phrase)
# Manage the assessment set-up:
X_train = train['X']
y_train = train['y']
X_assess = None
y_assess = None
# Assessment dataset using the training vectorizer:
assess = sst.build_dataset(sst.dev_reader,
lambda x: x,
sst.ternary_class_func,
vectorizer=train['vectorizer'],
vectorize=False,
subtree_labels=phrase)
X_assess, y_assess = assess['X'], assess['y']
test = sst.build_dataset(sst.test_reader,
lambda x: x,
sst.ternary_class_func,
vectorizer=train['vectorizer'],
vectorize=False,
subtree_labels=phrase)
X_test, y_test = test['X'], test['y']
tree_vocab = ["unk"] + sst.get_vocab(map(lambda x: x.leaves(), X_train),
n_words=5000)
embedding = np.asarray(
[glove_lookup.get(k, glove_lookup["unk"]) for k in tree_vocab])
model = base_model(tree_vocab,
eta=eta,
batch_size=16,
embed_dim=embed,
hidden_dim=embed,
max_length=120,
max_iter=10,
embedding=embedding,
reg=0.0,
train_embedding=True,
use_phrases=phrase)
model.fit(X_train, y_train, X_assess=X_assess, y_assess=y_assess)
# Ensure that restoring works properly!
#model.restore(y_assess)
#model.test(X_assess, y_assess)
# Actual test.
model.restore(y_test)
model.test(X_test, y_test)
end = time.time()
print("Time: %s" % (end - start))
import numpy as np
import pandas as pd
from mittens import GloVe
import vsm
import sst
sst.build_dataset()
vsm.ngram_vsm()
pd.DataFrame
def dice_distance(u, v):
np.min(u, v)
np.hstack(u, v)
# ## Building datasets for experiments
#
# The second major phase for our analysis is a kind of set-up phase. Ingredients:
#
# * A reader like `train_reader`
# * A feature function like `unigrams_phi`
# * A class function like `binary_class_func`
#
# The convenience function `sst.build_dataset` uses these to build a dataset for training and assessing a model. See its documentation for details on how it works. Much of this is about taking advantage of `sklearn`'s many functions for model building.
# In[7]:
train_dataset = sst.build_dataset(SST_HOME,
reader=sst.train_reader,
phi=unigrams_phi,
class_func=sst.binary_class_func,
vectorizer=None)
# In[9]:
train_dataset['X'].shape
# In[10]:
print(
"Train dataset with unigram features has {:,} examples and {:,} features".
format(*train_dataset['X'].shape))
# Notice that `sst.build_dataset` has an optional argument `vectorizer`:
#
