def classify_with_network3(
# alignment files
group_1, group_2, group_3, # these arguments should be strings that are used as the file suffix
# which data to use
strand, motif_start_positions, preprocess, events_per_pos, feature_set, title,
# training params
learning_algorithm, train_test_split, iterations, epochs, max_samples, batch_size,
# model params
learning_rate, L1_reg, L2_reg, hidden_dim, model_type, model_dir=None, extra_args=None,
# output params
out_path="./"):
# checks and file IO
assert(len(motif_start_positions) >= 3)
out_file = open(out_path + title + ".tsv", 'wa')
if model_dir is not None:
print("looking for model in {}".format(os.path.abspath(model_dir)))
model_file = find_model_path(os.path.abspath(model_dir), title)
else:
model_file = None
# bin to hold accuracies for each iteration
scores = []
collect_data_vectors_args = {
"events_per_pos": events_per_pos,
"portion": train_test_split,
"strand": strand,
"max_samples": max_samples,
"feature_set": feature_set,
"kmer_length": 6
}
for i in xrange(iterations):
list_of_datasets = [] # [((g1, g1l), (xg1, xg1l), (tg1, tg1l)), ... ]
add_to_list = list_of_datasets.append
for n, group in enumerate((group_1, group_2, group_3)):
train_set, xtrain_set, test_set = collect_data_vectors2(label=n,
files=group,
motif_starts=motif_start_positions[n],
dataset_title=title + "_group{}".format(n),
**collect_data_vectors_args)
add_to_list((train_set, xtrain_set, test_set))
# unpack to make things easier, list_of_datasets[group][set_idx][vector/labels]
c_train, c_tr_labels = list_of_datasets[0][0][0], list_of_datasets[0][0][1]
c_xtr, c_xtr_targets = list_of_datasets[0][1][0], list_of_datasets[0][1][1]
c_test, c_test_targets = list_of_datasets[0][2][0], list_of_datasets[0][2][1]
mc_train, mc_tr_labels = list_of_datasets[1][0][0], list_of_datasets[1][0][1]
mc_xtr, mc_xtr_targets = list_of_datasets[1][1][0], list_of_datasets[1][1][1]
mc_test, mc_test_targets = list_of_datasets[1][2][0], list_of_datasets[1][2][1]
hmc_train, hmc_tr_labels = list_of_datasets[2][0][0], list_of_datasets[2][0][1]
hmc_xtr, hmc_xtr_targets = list_of_datasets[2][1][0], list_of_datasets[2][1][1]
hmc_test, hmc_test_targets = list_of_datasets[2][2][0], list_of_datasets[2][2][1]
nb_c_train, nb_c_xtr, nb_c_test = len(c_train), len(c_xtr), len(c_test)
nb_mc_train, nb_mc_xtr, nb_mc_test = len(mc_train), len(mc_xtr), len(mc_test)
nb_hmc_train, nb_hmc_xtr, nb_hmc_test = len(hmc_train), len(hmc_xtr), len(hmc_test)
assert(nb_c_train > 0 and nb_mc_train > 0 and nb_hmc_train > 0), "got zero training vectors"
# level training events so that the model gets equal exposure
tr_level = np.min([nb_c_train, nb_mc_train, nb_hmc_train])
xtr_level = np.min([nb_c_xtr, nb_mc_xtr, nb_hmc_xtr])
test_level = np.min([nb_c_test, nb_mc_test, nb_hmc_test])
# log how many vectors we got
print("{motif}: got {C} C, {mC} mC, and {hmC} hmC, training vectors, leveled to {level}"
.format(motif=title, C=nb_c_train, mC=nb_mc_train, hmC=nb_hmc_train, level=tr_level), file=sys.stderr)
print("{motif}: got {xC} C, {xmC} mC, and {xhmC} hmC, cross-training vectors, leveled to {xlevel}"
.format(motif=title, xC=nb_c_xtr, xmC=nb_mc_xtr, xhmC=nb_hmc_xtr, xlevel=xtr_level), file=sys.stderr)
print("{motif}: got {xC} C, {xmC} mC, and {xhmC} hmC, test vectors, leveled to {tstLevel}"
.format(motif=title, xC=len(c_test), xmC=len(mc_test), xhmC=len(hmc_test),
tstLevel=test_level), file=sys.stderr)
# stack the data into one object
# training data
training_data = stack_and_level_datasets3(c_train, mc_train, hmc_train, tr_level)
training_labels = append_and_level_labels3(c_tr_labels, mc_tr_labels, hmc_tr_labels, tr_level)
# cross training
xtrain_data = stack_and_level_datasets3(c_xtr, mc_xtr, hmc_xtr, xtr_level)
xtrain_targets = append_and_level_labels3(c_xtr_targets, mc_xtr_targets, hmc_xtr_targets, xtr_level)
# test
test_data = stack_and_level_datasets3(c_test, mc_test, hmc_test, test_level)
test_targets = append_and_level_labels3(c_test_targets, mc_test_targets, hmc_test_targets, test_level)
prc_train, prc_xtrain, prc_test = preprocess_data(training_vectors=training_data,
xtrain_vectors=xtrain_data,
test_vectors=test_data,
preprocess=preprocess)
#if evaluate is True:
# all_test_data = np.vstack((xtrain_data, test_data))
# all_test_targets = np.append(xtrain_targets, test_targets)
# errors, probs = evaluate_network(all_test_data, all_test_targets, model_dir, model_type, batch_size, extra_args)
# return
# shuffle data
X, y = shuffle_and_maintain_labels(prc_train, training_labels)
working_directory_path = "{outpath}/{title}_Models/".format(outpath=out_path, title=title)
if not os.path.exists(working_directory_path):
os.makedirs(working_directory_path)
trained_model_dir = "{workingdirpath}{iteration}/".format(workingdirpath=working_directory_path,
iteration=i)
training_routine_args = {
"motif": title,
"train_data": X,
"labels": y,
"xTrain_data": prc_xtrain,
"xTrain_targets": xtrain_targets,
"learning_rate": learning_rate,
"L1_reg": L1_reg,
"L2_reg": L2_reg,
"epochs": epochs,
"batch_size": batch_size,
"hidden_dim": hidden_dim,
"model_type": model_type,
"model_file": model_file,
"trained_model_dir": trained_model_dir,
"extra_args": extra_args
}
if learning_algorithm == "annealing":
net, summary = mini_batch_sgd_with_annealing(**training_routine_args)
else:
net, summary = mini_batch_sgd(**training_routine_args)
errors, probs = predict(prc_test, test_targets, training_routine_args['batch_size'], net,
model_file=summary['best_model'])
errors = 1 - np.mean(errors)
probs = zip(probs, test_targets)
print("{0}:{1}:{2} test accuracy.".format(title, i, (errors * 100)))
out_file.write("{}\n".format(errors))
scores.append(errors)
with open("{}test_probs.pkl".format(trained_model_dir), 'w') as probs_file:
cPickle.dump(probs, probs_file)
print(">{motif}\t{accuracy}".format(motif=title, accuracy=np.mean(scores), end="\n"), file=out_file)
return net
def classify_with_network2(
# alignment files
group_1, group_2, group_3,
# which data to use
strand, motif_start_positions, preprocess, events_per_pos, feature_set, title,
# training params
learning_algorithm, train_test_split, iterations, epochs, max_samples, batch_size,
# model params
learning_rate, L1_reg, L2_reg, hidden_dim, model_type, model_dir=None, extra_args=None,
# output params
out_path="./"):
print("2 way classification")
assert(len(motif_start_positions) >= 2)
out_file = open(out_path + title + ".tsv", 'wa')
if model_dir is not None:
print("looking for model in {}".format(model_dir))
model_file = find_model_path(model_dir, title)
else:
model_file = None
# bin to hold accuracies for each iteration
scores = []
collect_data_vectors_args = {
"events_per_pos": events_per_pos,
"portion": train_test_split,
"strand": strand,
"max_samples": max_samples,
"feature_set": feature_set,
"kmer_length": 6
}
for i in xrange(iterations):
list_of_datasets = [] # [((g1, g1l), (xg1, xg1l), (tg1, tg1l)), ... ]
add_to_list = list_of_datasets.append
for n, group in enumerate((group_1, group_2)):
train_set, xtrain_set, test_set = collect_data_vectors2(label=n,
files=group,
motif_starts=motif_start_positions[n],
dataset_title=title + "_group{}".format(n),
**collect_data_vectors_args)
add_to_list((train_set, xtrain_set, test_set))
# unpack list
g1_train, g1_tr_labels = list_of_datasets[0][0][0], list_of_datasets[0][0][1]
g1_xtr, g1_xtr_targets = list_of_datasets[0][1][0], list_of_datasets[0][1][1]
g1_test, g1_test_targets = list_of_datasets[0][2][0], list_of_datasets[0][2][1]
g2_train, g2_tr_labels = list_of_datasets[1][0][0], list_of_datasets[1][0][1]
g2_xtr, g2_xtr_targets = list_of_datasets[1][1][0], list_of_datasets[1][1][1]
g2_test, g2_test_targets = list_of_datasets[1][2][0], list_of_datasets[1][2][1]
nb_g1_train, nb_g1_xtr, nb_g1_test = len(g1_train), len(g1_xtr), len(g1_test)
nb_g2_train, nb_g2_xtr, nb_g2_test = len(g2_train), len(g2_xtr), len(g2_test)
assert(nb_g1_train > 0 and nb_g2_train > 0), "got {0} group 1 training and " \
"{1} group 2 training vectors".format(nb_g1_train, nb_g2_train)
# level training and cross-training events so that the model gets equal exposure
tr_level = np.min([nb_g1_train, nb_g2_train])
xtr_level = np.min([nb_g1_xtr, nb_g2_xtr])
test_level = np.min([nb_g1_test, nb_g2_test])
print("{motif}: got {g1} group 1 and {g2} group 2 training vectors, leveled to {level}"
.format(motif=title, g1=nb_g1_train, g2=nb_g2_train, level=tr_level))
print("{motif}: got {g1} group 1 and {g2} group 2 cross-training vectors, leveled to {level}"
.format(motif=title, g1=nb_g1_xtr, g2=nb_g2_xtr, level=xtr_level))
print("{motif}: got {g1} group 1 and {g2} group 2 test vectors, leveled to {level}"
.format(motif=title, g1=nb_g1_test, g2=nb_g2_test, level=test_level))
training_data = stack_and_level_datasets2(g1_train, g2_train, tr_level)
training_labels = append_and_level_labels2(g1_tr_labels, g2_tr_labels, tr_level)
xtrain_data = stack_and_level_datasets2(g1_xtr, g2_xtr, xtr_level)
xtrain_targets = append_and_level_labels2(g1_xtr_targets, g2_xtr_targets, xtr_level)
test_data = stack_and_level_datasets2(g1_test, g2_test, test_level)
test_targets = append_and_level_labels2(g1_test_targets, g2_test_targets, test_level)
prc_train, prc_xtrain, prc_test = preprocess_data(training_vectors=training_data,
xtrain_vectors=xtrain_data,
test_vectors=test_data,
preprocess=preprocess)
# evaluate
X, y = shuffle_and_maintain_labels(prc_train, training_labels)
trained_model_dir = "{0}{1}_Models/".format(out_path, title)
training_routine_args = {
"motif": title,
"train_data": X,
"labels": y,
"xTrain_data": prc_xtrain,
"xTrain_targets": xtrain_targets,
"learning_rate": learning_rate,
"L1_reg": L1_reg,
"L2_reg": L2_reg,
"epochs": epochs,
"batch_size": batch_size,
"hidden_dim": hidden_dim,
"model_type": model_type,
"model_file": model_file,
"trained_model_dir": trained_model_dir,
"extra_args": extra_args
}
if learning_algorithm == "annealing":
net, summary = mini_batch_sgd_with_annealing(**training_routine_args)
else:
net, summary = mini_batch_sgd(**training_routine_args)
errors, probs = predict(prc_test, test_targets, training_routine_args['batch_size'], net,
model_file=summary['best_model'])
errors = 1 - np.mean(errors)
print("{0}: {1} test accuracy.".format(title, (errors * 100)))
out_file.write("{}\n".format(errors))
scores.append(errors)
with open("{}test_probs.pkl".format(trained_model_dir), 'w') as probs_file:
cPickle.dump(probs, probs_file)
print(">{motif}\t{accuracy}".format(motif=title, accuracy=np.mean(scores), end="\n"), file=out_file)
return net
