I am carrying out some research on sentiment analysis and I was looking to reproduce the results of Subtask A, specified in the paper here. Now, I should add that with the source code as is provided, and with TensorFlow backend, I get very similar results. However, upon closer inspection, I noticed some inconsistincies between what's in the code and in the paper.
For example, the paper states that:
We train all of our networks to minimize the
cross-entropy loss, using back-propagation with
stochastic gradient descent and mini-batches of
size 128.
However, in the code, the batch size is 50:
history = nn_model.fit(training[0], training[1], validation_data=validation if not FINAL else testing, epochs=50, batch_size=50, class_weight=class_weights, callbacks=_callbacks)
There are few other inconsistencies as well. My question is, should I change the hyper parameters to the ones specified in the paper in order to exactly replicate the results? Or is exact replication not possible? I should also add that when I run the system multiple times, I get similar, but slightly different results for each epoch. Maybe this has something to with using TensorFlow as a backend? I've been reading that some people have had issues getting the same results when using TensorFlow backend.
Not really an issue, more of a question.
I am carrying out some research on sentiment analysis and I was looking to reproduce the results of Subtask A, specified in the paper here. Now, I should add that with the source code as is provided, and with TensorFlow backend, I get very similar results. However, upon closer inspection, I noticed some inconsistincies between what's in the code and in the paper.
For example, the paper states that:
However, in the code, the batch size is 50:
history = nn_model.fit(training[0], training[1], validation_data=validation if not FINAL else testing, epochs=50, batch_size=50, class_weight=class_weights, callbacks=_callbacks)There are few other inconsistencies as well. My question is, should I change the hyper parameters to the ones specified in the paper in order to exactly replicate the results? Or is exact replication not possible? I should also add that when I run the system multiple times, I get similar, but slightly different results for each epoch. Maybe this has something to with using TensorFlow as a backend? I've been reading that some people have had issues getting the same results when using TensorFlow backend.