def get_nearest_neighbors(w, embeddings, k=1000):
"""
For every transition in every pattern, gets the word with the highest
score for that transition.
Only looks at the first `k` words in the vocab (makes sense, assuming
they're sorted by descending frequency).
"""
return argmax(torch.mm(w, embeddings[:k, :]))
def interpret_documents(model, batch_size, dev_data, dev_text, ofile,
max_doc_len):
j = 0
with open(ofile, "w") as ofh:
for batch_idx, chunk in enumerate(chunked(dev_data, batch_size)):
batch = Batch([x for x, y in chunk], model.embeddings,
model.to_cuda)
res, scores = model.forward(batch, 1)
print("ss", scores.size())
output = softmax(res).data
predictions = [int(x) for x in argmax(output)]
num_patts = scores.size()[1]
diffs = np.zeros((num_patts, batch.size()))
# Traversing all patterns.
for i in range(num_patts):
# Copying scores data to numpy array.
scores_data = np.array(scores.data.numpy(), copy=True)
# Zeroing out pattern number i across batch
scores_data[:, i] = 0
# Running mlp.forward() with zeroed out scores.
forwarded = softmax(
model.mlp.forward(Variable(
torch.FloatTensor(scores_data)))).data.numpy()
# Computing difference between forwarded scores and original scores.
for k in range(batch.size()):
# diffs[i,k] = output[k, predictions[k]] - \
# output[k, 1 - predictions[k]] - \
# forwarded[k, predictions[k]] + \
# forwarded[k, 1 - predictions[k]]
diffs[i, k] = forwarded[k, 1 - predictions[k]] - output[
k, 1 - predictions[k]]
# Now, traversing documents in batch
for i in range(batch.size()):
# Document string
text_str = str(" ".join(dev_text[j]).encode('utf-8'))[2:-1]
# Top ten patterns with largest differences between leave-one-out score and original score.
top_ten_deltas = sorted(enumerate(diffs[:, i]),
key=lambda x: x[1],
reverse=True)[:10]
top_ten_neg_deltas = sorted(enumerate(diffs[:, i]),
key=lambda x: x[1])[:10]
# Top ten patterns with largest overall score (regardless of classification)
top_ten_scores = sorted(enumerate(scores.data.numpy()[i, :]),
key=lambda x: x[1],
reverse=True)[:10]
top_scoring_spans = get_top_scoring_spans_for_doc(
model, dev_data[j], max_doc_len)
# Printing out everything.
ofh.write(
"{} {} {} All in, predicted: {:>2,.3f} All in, not-predicted: {:>2,.3f} Leave one out: +res: {} -res: {} Patt scores: {}\n"
.format(
dev_data[j][1], predictions[i], text_str,
output[i, predictions[i]],
output[i, 1 - predictions[i]], " ".join([
"{}:{:>2,.3f}".format(i, x)
for (i, x) in top_ten_deltas
]), " ".join([
"{}:{:>2,.3f}".format(i, x)
for (i, x) in top_ten_neg_deltas
]), " ".join([
"{}:{:>2,.3f}".format(i, x)
for (i, x) in top_ten_scores
])))
ofh.write("Top ten deltas:\n")
for l in top_ten_deltas:
s = top_scoring_spans[l[0]].display(dev_text[j])
ofh.write(
str(int(l[0])) + " " + str(s.encode('utf-8'))[2:-1] +
"\n")
ofh.write("Top ten negative deltas:\n")
for l in top_ten_neg_deltas:
s = top_scoring_spans[l[0]].display(dev_text[j])
ofh.write(
str(int(l[0])) + " " + str(s.encode('utf-8'))[2:-1] +
"\n")
j += 1