def test_experiment(assess_reader):
def fit_maxent(X, y):
mod = LogisticRegression(solver='liblinear', multi_class='auto')
mod.fit(X, y)
return mod
nli.experiment(
train_reader=nli.SNLITrainReader(snli_home, samp_percentage=0.01),
phi=lambda x, y: {"$UNK": 1},
train_func=fit_maxent,
assess_reader=assess_reader,
random_state=42)
def test_experiment():
def fit_maxent(X, y):
mod = LogisticRegression()
mod.fit(X, y)
return mod
nli.experiment(train_reader=nli.SNLITrainReader(samp_percentage=0.01),
phi=lambda x, y: {"$UNK": 1},
train_func=fit_maxent,
assess_reader=None,
random_state=42)
def main():
exp = nli.experiment(train_reader=nli.SNLITrainReader(SNLI_HOME,
samp_percentage=1.0),
assess_reader=nli.SNLIDevReader(SNLI_HOME,
samp_percentage=1.0),
phi=sentence_encoding_rnn_phi,
train_func=fit_bilstm_attention,
random_state=None,
vectorize=False)
print(exp)
# In[19]:
dataset = nli.build_dataset(train_reader, word_overlap_phi)
# In[20]:
dataset.keys()
# However, it's more efficient to use `nli.experiment` to bring all these pieces together. This wrapper will work for all the models we consider.
# In[21]:
_ = nli.experiment(train_reader=nli.SNLITrainReader(SNLI_HOME,
samp_percentage=0.10,
random_state=42),
phi=word_overlap_phi,
train_func=fit_softmax_with_crossvalidation,
assess_reader=None,
random_state=42)
# In[12]:
_ = nli.experiment(train_reader=nli.SNLITrainReader(SNLI_HOME,
samp_percentage=0.10,
random_state=42),
phi=word_cross_product_phi,
train_func=fit_softmax_with_crossvalidation,
assess_reader=None,
random_state=42)
# As expected `word_cross_product_phi` is very strong. At this point, one might consider scaling up to `samp_percentage=None`, i.e., the full training set. Such a baseline is very similar to the one established in [the original SNLI paper by Bowman et al.](https://aclanthology.info/papers/D15-1075/d15-1075) for models like this one.