def result_callback(id_to_label, output_file, res):
"""
When TF inference completes, assemble the results to write to disk.
"""
data = {}
order = [id_to_label[i] for i in range(len(id_to_label))]
correct, total = 0, 0
for (gold, cid), (probs, ) in res:
total += 1
if gold == order[probs.flatten().argmax()]:
correct += 1
data[cid] = (gold, {k: v for k, v in zip(order, probs.flatten())})
logging.info("Accuracy: %.3f", float(correct) / total)
write_probabilities(data, options.output)
acc, preds, prob = sess.run(
[
model.accuracy, model.predictions,
model.log_probs
],
feed_dict={
model.compiler.loom_input_tensor: batch_feed,
model._keep_probability: 1.0
})
accs.append(acc)
probs.append(prob)
probs = numpy.concatenate(probs, axis=0)
logging.info("Accuracy: %f", sum(accs) / len(accs))
data = {}
order = [id_to_label[i] for i in range(len(id_to_label))]
for dist, (lid, _, cid) in zip(probs, instances):
cid = id_to_cid.get(cid, str(cid))
g = id_to_label[lid]
data[cid] = (g,
{k: v
for k, v in zip(order, dist.flatten())})
write_probabilities(data, options.output)
else:
logging.error(
"You must specify the input and output files, and either a training file, or testing and model files!"
)
finally:
shutil.rmtree(temppath)
# for (cid, label), text in zip(gold, instances):
# total_probs = {k : v for k, v in prior.items()}
# seq = [None for i in range(options.max_ngram)]
# for c in list(text) + [None]:
# seq = seq[1:]
# seq.append(c)
# single_probs = {}
# for n in range(1, options.max_ngram + 1):
# b = backoffs[n]
# num_gram = tuple(seq[-n:])
# den_gram = num_gram[:-1]
# for l, v in model.items():
# num = v[n].get(num_gram, 0)
# #print(num_gram, den_gram, l)
# if num > 0:
# den = v[n - 1][den_gram]
# #print(num, den)
# prob = num / den
# #print(prob, b)
# single_probs[l] = single_probs.get(l, 0.0) + (prob * b)
# for l, p in single_probs.items():
# if p != 0.0:
# total_probs[l] += math.log(p)
# scores[cid] = (label, total_probs)
# #print(prob)
# #prob += model[n][gram] * b
# #model[n][gram] = model[n].get(gram, 0) + 1
write_probabilities(scores, options.output)
