def main():
# Get the dataset from http://ai.stanford.edu/~amaas/data/sentiment/
model, utt_train = load_all_models(model_path)
path = dataset_path
train_x, train_y= grab_data(path+'.train', model)
test_x, test_y = grab_data(path+'.test', model)
f = open('swda_embedded_'+ model_path.split('/')[-1] + '.pkl', 'wb')
pkl.dump((train_x, train_y), f, -1)
pkl.dump((test_x, test_y), f, -1)
f.close()
def evaluate_model(embedding_model_location, with_context=False, f=e_add, aggregated_tagset=False):
# load training and test data
train_utt, train_Y, test_utt, test_Y = load_data()
# uncomment this to find basline scores
# baseline_scores(train_utt, train_Y, test_utt, test_Y)
print "Creating representations"
# Load utterance embedding models
embedding_model, _ = load_all_models(embedding_model_location)
# represent utterances in some way,
# train_X = represent_simple(train_utt, embedding_model)
# test_X = represent_simple(test_utt, embedding_model)
if with_context:
# ---------- lqrz: add or concatenate the previous utterance
# f = concat
# f = e_add
test_X = represent_mix_simple(test_utt, test_Y, embedding_model, f)
train_X = represent_mix_lookup(train_Y, embedding_model, f)
# ----------
else:
train_X = represent_lookup(train_Y, embedding_model)
test_X = represent_simple(test_utt, embedding_model)
# encode tags
train_Y, test_Y = encode_tags(train_Y, test_Y, aggregated_tagset=aggregated_tagset)
# print np.array(train_X).shape, np.array(test_X).shape, np.array(train_Y).shape, np.array(test_Y).shape
print "Training classifiers"
# Train classifiers, print scores
print "Model: ", embedding_model_location
print "KNN Accuracy: ", cl.KNN_classifier(train_X, train_Y, test_X, test_Y)
# print "SVM Accuracy: ", cl.SVM_classifier(train_X, train_Y, test_X, test_Y)
print "NB Accuracy: ", cl.NB_classifier(train_X, train_Y, test_X, test_Y)
print "MLP Accuracy: ", cl.MLP_classifier(train_X, train_Y, test_X, test_Y, n_iter=10)
for tag, tokens in get_utterances_from_file(fname):
# remove id from tag
tag = tag.split("/")[0]
utterances[tag].append(tokens)
return utterances
if __name__ == '__main__':
#TODO: change paths accordingly
embedding_model_filename = 'data/test' # path to doc2vec model
utterance_filename = 'data/swda_utterances.txt' # path to utterance file
model, _ = load_all_models(embedding_model_filename)
utterances = load_utterances(utterance_filename)
#TODO: select tags to plot (max 5)
tags_to_plot = ['qw','ft','ar','fa', 't1']
#TODO: select nr of samples
n_samples = 500
samples = np.empty((n_samples*len(tags_to_plot), model.layer1_size))
for i,tag in enumerate(tags_to_plot):
tag_samples = get_vector_samples(tag, utterances, model, n_samples)
samples[i*n_samples:(i+1)*n_samples,:] = tag_samples
plotGraph(samples, n_samples, tags_to_plot, dimensions='2D')