def get_trans():
configs = get_configs()
data = get_data()
nn, ngram = get_models()
# if bigram nn, it's from t to t-1, then need to transpose?
nn_trans_dict = nn.predict_identity()
nn_trans = nn_trans_dict.values()[0]
print '...nn_trans type', type(nn_trans), nn_trans.shape
nn_trans = nn_trans_dict['1']
print nn_trans_dict.keys()
print '...nn_trans type', type(nn_trans), nn_trans.shape
nn_trans = np.exp(nn_trans)
ngram_trans = ngram.ngram
return ngram_trans, nn_trans, ngram, nn, configs, data
def get_models(data=None, configs=None, save=False):
if configs is None:
configs = get_configs()
if data is None:
data = get_data(configs)
# TODO: remove hack
if configs["bigram"]:
reduced_keys = [configs["reduced_key"]]
data.keys = reduced_keys
test_data = data.get_test_data()
test_data.keys = reduced_keys
retrieve_model = configs["retrieve_model"]
model = SkipGramNN(data, configs)
print "SkipGramNN, # of syms", len(model.syms)
if not retrieve_model:
model_loss = model.train()
if save:
model.save("skipgram-%s.pkl" % (configs["corpus"]))
plt.clf()
plt.plot(model.loss_curve)
plt.savefig("losses-%s.png" % configs.name)
print "=== train loss ==="
print "loss: %.2f" % model_loss
loss = model.check_loss()
if not configs["regularize"]:
assert np.allclose(loss, model_loss)
if save:
model_weights = model.weights.value
fname = "w1-%s.pkl" % configs.name
print fname
with open(fname, "wb") as p:
pickle.dump(model.W1.value, p)
pickle.dump(data.syms, p)
pickle.dump(model_loss, p)
pickle.dump(model_weights, p)
pickle.dump(configs, p)
fname = "skipgram-bach.pkl"
model.save(fname)
else:
fname = os.path.join("data", "test_skipgram_model.pkl")
print fname
assert fname is not None, "Error: no model to retrieve in the time being"
with open(fname, "rb") as p:
w1 = pickle.load(p)
syms = pickle.load(p)
model_loss = pickle.load(p)
model_weights = pickle.load(p)
configs_reloaded = pickle.load(p)
for key in configs.keys():
if key not in configs_reloaded.keys():
print "no key", key
for key in configs.keys():
if key in configs_reloaded.keys():
if configs[key] != configs_reloaded[key]:
print configs[key], configs_reloaded[key]
# assert configs == configs_reloaded
model.init_weights(model_weights, model_loss)
train_seq_data = data.get_train_seqs_data()
train_seqs = [seq for seq in train_seq_data.seqs]
syms = data.syms
# ngram_model = NGram(train_seqs, syms, 2, configs)
ngram_model = NGram(data.seqs, syms, 2, configs)
print "\n\nNgram, # of syms", len(ngram_model.syms)
if save:
ngram_model.save("bigram-%s.pkl" % (configs["corpus"]))
print len(ngram_model.syms), len(model.data.syms)
assert ngram_model.syms == model.data.syms
return model, ngram_model
