def debug(args):
contexts, repsonses, contexts_graph_adjs, response_graph_adjs, labels = load_data_from_file(args.dir + 'data/2011_split_by_idname.csv', tokenize=True, read_case_num=10)
net = GMN(emdedding_dim=768, use_bert=True).to(DEVICE)
optimizer = torch.optim.Adam(net.parameters(), lr=LEARNING_RATE)
loss_function = torch.nn.BCELoss()
print('start training...')
print('dataset size: %s' % labels.shape)
for epoch in range(EPOCH):
net.train()
losses = []
for k in range(len(labels) // BATCH_SIZE):
output_p = net(contexts[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
repsonses[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
contexts_graph_adjs[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
response_graph_adjs[k * BATCH_SIZE:(k+1) * BATCH_SIZE])
loss = loss_function(output_p, labels[k * BATCH_SIZE:(k+1) * BATCH_SIZE])
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.append(loss.item())
print('loss: %.2f' % np.mean(losses))
def spam_filter(train_mode: bool):
labels, texts = dl.load_data_from_file(dataset_path)
word_embeddings = dl.load_word_embeddings_from_file(word_emb_path)
tokenizer = prepare.get_prepared_tokenizer(texts)
texts = tokenizer.texts_to_sequences(texts)
texts = np.array(texts)
texts = pad_sequences(texts, maxlen=config['dataset']['max_seq_len'])
labels = prepare.encode_labels(labels)
labels = np.array(labels)
texts_train, texts_test, labels_train, labels_test = train_test_split(
texts,
labels,
random_state=config['dataset']['random_state'],
test_size=config['dataset']['test_size'])
embeddings_matrix = prepare.map_embeddings_to_word_index(
word_embeddings, tokenizer.word_index)
seq_model = model.get_compiled_model(embeddings_matrix, config)
seq_model.summary()
if train_mode:
print("Entering train mode")
train.train_model(seq_model, config, texts_train, labels_train,
texts_test, labels_test)
else:
print("Loading weights file and entering prediction mode")
load_weights_from_file(
'../checkpoints/weights-improvement-12-0.98.hdf5', seq_model,
texts_test, labels_test)
while True:
try:
text = str(input('>> '))
prediction = get_prediction(seq_model, tokenizer, text)
prediction_index = utils.probability_to_index(prediction)
print("Prediction:", utils.decode_index(prediction_index))
except (KeyboardInterrupt, KeyError):
print()
exit()
def train(args):
eval_contexts, eval_repsonses, eval_contexts_graph_adjs, eval_response_graph_adjs, eval_labels = load_data_from_file(args.dir + 'data/2020_split_by_idname.csv', tokenize=True)
contexts, repsonses, contexts_graph_adjs, response_graph_adjs, labels = load_data_from_file(args.dir + 'data/from_2011_to_2019_split_by_idname.csv', tokenize=True)
# move tensors to the specified devices.
if DEVICE == 'cuda':
for tensor in [eval_contexts, eval_repsonses, eval_contexts_graph_adjs, eval_response_graph_adjs, eval_labels,
contexts, repsonses, contexts_graph_adjs, response_graph_adjs, labels]:
tensor.cuda()
else:
for tensor in [eval_contexts, eval_repsonses, eval_contexts_graph_adjs, eval_response_graph_adjs, eval_labels,
contexts, repsonses, contexts_graph_adjs, response_graph_adjs, labels]:
tensor.cpu()
net = GMN(emdedding_dim=768, use_bert=True).to(DEVICE)
optimizer = torch.optim.Adam(net.parameters(), lr=LEARNING_RATE)
loss_function = torch.nn.BCELoss()
print('start training...')
print('dataset size: %s' % labels.shape)
for epoch in range(EPOCH):
net.train()
losses = []
for k in range(len(labels) // BATCH_SIZE):
output_p = net(contexts[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
repsonses[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
contexts_graph_adjs[k * BATCH_SIZE:(k+1) * BATCH_SIZE],
response_graph_adjs[k * BATCH_SIZE:(k+1) * BATCH_SIZE])
loss = loss_function(output_p, labels[k * BATCH_SIZE:(k+1) * BATCH_SIZE])
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.append(loss.item())
print('loss: %.2f' % np.mean(losses))
evaluate(net, eval_contexts, eval_repsonses, eval_contexts_graph_adjs, eval_response_graph_adjs, eval_labels)
def __getitem__(
self, index
): # because how we design our data loadin, this returns a whole batch
data = data_loader.load_data_from_file(
np.random.choice(np.array(self.file_list))) # random file in here
if self.who_dies_next_mode == True or self.is_validation == True:
X, y, death_times = data_loader.getBatchBalanced(
data,
self.batch_size,
self.feature_indicies,
self.label_indicies,
get_death_times=True)
return X, y, death_times
else:
player_i = np.random.randint(10)
X, y, death_times = data_loader.getBalancedBatchForPlayer(
data,
player_i,
self.batch_size,
self.feature_indicies,
self.label_indicies,
get_death_times=True)
return X, y, death_times, player_i
def make_predictions(file,modelPath):
trainingDataFiles = [file]#glob.glob("/scratch/staff/ak1774/shared_folder/data/train/*.h5")
data = data_loader.load_data_from_file(trainingDataFiles[0])
models = []
for i in range(1,2):
print(i)
models.append( load_pytorch_model('ModelData/' +str(i) +'/' +'model.model',
get_config('/' +str(i) +'/config.json'), data) )
#fullGameData,fullGameLabels = data_loader.getSequencialNaive(data,hero_feature_indicies,label_indicies)
xLims = data['time'].values
#¢health = data['player_4_m_iHealth'].values
#######################
# get original health
######################
norm_stats = None
with open("norm_stats.pickle", 'rb') as f:
norm_stats = pickle.load(f)
for label,min_value,max_value in normalization_stats:
if "_m_iHealth" in label:
health_min = min_value
health_max = max_value
if "m_iMaxHealth" in label:
maxhealth_min = min_value
maxhealth_max = max_value
healthes = []
max_healthes = []
relative_healthes = []
for i in range(0,10):
health_vals = data['player_' + str(i) + '_m_iHealth'].values
maxhealth_vals = data['player_' + str(i) + '_m_iMaxHealth'].values
health_vals = health_vals * (health_max - health_min) + health_min
maxhealth_vals = maxhealth_vals * (maxhealth_max - maxhealth_min) + maxhealth_min
relative_health_vals = health_vals / maxhealth_vals # hopefully maxhealth is never 0
healthes.append(health_vals)
max_healthes.append(maxhealth_vals)
relative_healthes.append(relative_health_vals)
#######################
# get death times
######################
labels = [(i,label) for i,label in enumerate(list(data))]
death_time_indicies = preprocess.labels_to_indicies(preprocess.select_features_by_name("time_until_next_death",labels))
death_times = data.values[:,death_time_indicies].astype(np.float32)
for m in models:
X = [torch.from_numpy(hero_X) for hero_X in m.fullGameData]
pred = model(X)
pred = torch.sigmoid(pred)
pred = pred.cpu().detach().numpy()
y = m.fullGameLabels
currentMeanTrueAccuracy = 0
currentMeanFalseAccuracy=0
numTruePos = 0
numFalsePos = 0
numTrueNeg = 0
numFalseNeg = 0
for i in range(0,data.shape[0]):
predX = 0
for m in models:
y = m.fullGameLabels[i]
y = np.array(y)
y = np.expand_dims(y,0)
X = [torch.from_numpy(hero_X[i:(i+1),:]) for hero_X in m.fullGameData]
print(i)
#predX = averagePred(models,X)
predX = modelPred(m.model,X) +predX
predX = predX/len(models)
'''
true_pos = ((predX > 0.5) == (y > 0.5)).reshape(-1).astype(np.float32)
true_neg = ((predX < 0.5) == (y <0.5)).reshape(-1).astype(np.float32)
false_pos = ((predX > 0.5) == (y < 0.5)).reshape(-1).astype(np.float32)
false_neg = ((predX < 0.5) == (y > 0.5)).reshape(-1).astype(np.float32)
for pos in true_neg:
if pos ==1:
numTrueNeg +=1
for pos in false_neg:
if pos ==1:
numFalseNeg +=1
for pos in true_pos:
if pos ==1:
numTruePos +=1
for pos in false_pos:
if pos ==1:
numFalsePos +=1
'''
prediction = predX
currentMeanTrueAccuracy += np.mean(true_pos)
currentMeanFalseAccuracy += np.mean(false_pos)
prediction = np.squeeze(prediction,0)
if i %3000 ==0:
print('Current true pos ' +str(currentMeanTrueAccuracy/(i+1)))
print('Current false pos ' +str(currentMeanFalseAccuracy/(i+1)))
heroStuff.append(prediction)
labelStuff.append(np.squeeze(y,0))
print()
print(numTruePos)
print(numTrueNeg)
print()
print(numFalsePos)
print(numFalseNeg)
print()
print('True Pos = ' + str(currentMeanTrueAccuracy/19326))
print('False pos = ' + str(currentMeanFalseAccuracy/19326))
heroStuff1 = np.swapaxes(heroStuff,0,1)
labelStuff1= np.swapaxes(labelStuff,0,1)
xLims = xLims - xLims[0] - 90
np.save('hero.npy', np.array(heroStuff1))
np.save('label.npy', np.array(labelStuff1))
np.save('xLims.npy', np.array(xLims))
np.save('health.npy',np.array(healthes))
def train_pytorch():
# is there a config in the current directory?
config_path = "config.json"
if not os.path.isfile("config.json"):
# use default config
config_path = os.path.dirname(
os.path.realpath(__file__)) + "/config/default.json"
with open(config_path) as f:
config = commentjson.load(f)
import pprint
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(config)
sys.stdout.flush()
WHO_DIES_NEXT_MODE = config["predict_who_dies_next"]
use_cuda = config["use_gpu"] == True and torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
print("using device: ", device)
model_type = locate(config["model"])
get_feature_indicies_fn = locate(config["feature_set"])
get_label_indicies_fn = locate(config["lable_set"])
batch_size = config["batch_size"]
print(type(batch_size))
print(type(config["log_at_every_x_sample"]))
epoch_size = int(config["log_at_every_x_sample"] / batch_size)
print("epoch_size: ", epoch_size)
checkpoint_frequency = int(config["chekpoint_at_every_x_sample"] /
(epoch_size * batch_size))
validation_epoch_sice = config["validation_epoch_size"]
if config["optimizer"] == "Adam":
OptimizerType = torch.optim.Adam
elif config["optimizer"] == "SGD":
OptimizerType = torch.optim.SGD
# YARCC
#trainingDataFiles = glob.glob("/scratch/ak1774/data/train/*.h5")
#validationDataFiles = glob.glob("/scratch/ak1774/data/validation/*.h5")
# Viking
trainingDataFiles = glob.glob(
str(Path.cwd().parent / 'randomized_data' / 'train') + '/*.h5'
) #glob.glob("/mnt/lustre/groups/cs-dclabs-2019/esport/death_prediction_data/randomized_data/train/*.h5")
validationDataFiles = glob.glob(
str(Path.cwd().parent / 'randomized_data' / 'validation') + '/*.h5'
) #glob.glob("/mnt/lustre/groups/cs-dclabs-2019/esport/death_prediction_data/randomized_data/validation/*.h5")
#trainingDataFiles = glob.glob("/scratch/staff/ak1774/shared_folder/data/train/*.h5")
#validationDataFiles = glob.glob("/scratch/staff/ak1774/shared_folder/data/validation/*.h5")
example_data = data_loader.load_data_from_file(trainingDataFiles[0])
hero_feature_indicies = get_feature_indicies_fn(example_data)
if WHO_DIES_NEXT_MODE == True:
label_indicies = get_label_indicies_fn(example_data)
else:
#label_indicies = get_label_indicies_fn(example_data,config["label_set_arg"])
label_indicies = get_label_indicies_fn(example_data)
inputFeatureSize = len(hero_feature_indicies[0])
outputFeatureSize = len(label_indicies)
if WHO_DIES_NEXT_MODE == True and outputFeatureSize != 11:
print("error, bad config, label set and prediction mode mismatch")
raise "error, bad config, label set and prediction mode mismatch"
elif WHO_DIES_NEXT_MODE == False and outputFeatureSize != 10:
print("error, bad config, label set and prediction mode mismatch")
raise "error, bad config, label set and prediction mode mismatch"
# the dataset returns a batch when called (because we get the whole batch from one file), the batch size of the data loader thus is set to 1 (default)
# epoch size is how many elements the iterator of the generator will provide, NOTE should not be too small, because it have a significant overhead p=0.05
training_set = DotaDataset(
file_list=trainingDataFiles,
batch_size=batch_size,
epoch_size=epoch_size,
feature_indicies=hero_feature_indicies,
label_indicies=label_indicies,
who_dies_next_mode=WHO_DIES_NEXT_MODE,
is_validation=False) # set is validation to get death times...
training_generator = torch.utils.data.DataLoader(
training_set, num_workers=20, worker_init_fn=worker_init_fn)
validation_set = DotaDataset(
file_list=validationDataFiles,
batch_size=batch_size,
epoch_size=validation_epoch_sice,
feature_indicies=hero_feature_indicies,
label_indicies=label_indicies,
who_dies_next_mode=WHO_DIES_NEXT_MODE,
is_validation=False
) # actually we want the same distribution, so we can compare loss, so dont do anything differently in case of validation
validation_generator = torch.utils.data.DataLoader(
validation_set, num_workers=20, worker_init_fn=worker_init_fn)
#model = models.SimpleFF(inputFeatureSize,outputFeatureSize)
model = model_type(inputFeatureSize, outputFeatureSize,
**config["model_params"])
model.to(device)
print(model.final_layers)
criterion = nn.CrossEntropyLoss()
binary_classification_loss = torch.nn.BCELoss()
optimizer = OptimizerType(model.parameters(), **config["optimizer_params"])
if WHO_DIES_NEXT_MODE == True:
all_train_losses = []
all_train_accuracies = []
all_train_kill_nokill_accuracies = []
all_train_per_second_accuracies = []
all_validation_losses = []
all_validation_accuracies = []
all_validation_kill_nokill_accuracies = []
all_validation_per_second_accuracies = []
for epoch_i in range(50000):
now = time.time()
np.random.seed(
) # reset seed https://github.com/pytorch/pytorch/issues/5059 data loader returns the same values
epoch_losses = []
epoch_overall_accuracies = []
epoch_kill_accuracies = []
epoch_no_kill_accuracies = []
epoch_one_sec_accuracies = []
epoch_two_sec_accuracies = []
epoch_three_sec_accuracies = []
epoch_four_sec_accuracies = []
epoch_five_sec_accuracies = []
for sub_epoch_i, (X, y,
death_times) in enumerate(training_generator):
# since we get a batch of size 1 of batch of real batch size, we take the 0th element
X = [(hero_X[0, :]).to(device) for hero_X in X]
y = torch.argmax(y[0, :], dim=1).to(device)
death_times = death_times[0]
# Forward + Backward + Optimize
optimizer.zero_grad()
output = model(X)
loss = criterion(output, y)
accuracy_vec = (torch.argmax(
output,
1) == y).cpu().numpy().reshape(-1).astype(np.float32)
loss.backward()
optimizer.step()
epoch_losses.append(loss.cpu().detach().numpy().reshape(-1)[0])
(overall_accuracy, (kill_accuracy, no_kill_accuracy),
(one_sec_accuracy, two_sec_accuracy, three_sec_accuracy,
four_sec_accuracy,
five_sec_accuracy)) = calculate_detailed_accuracies(
accuracy_vec, death_times, y)
epoch_overall_accuracies.append(overall_accuracy)
epoch_kill_accuracies.extend(kill_accuracy)
epoch_no_kill_accuracies.extend(no_kill_accuracy)
epoch_one_sec_accuracies.extend(one_sec_accuracy)
epoch_two_sec_accuracies.extend(two_sec_accuracy)
epoch_three_sec_accuracies.extend(three_sec_accuracy)
epoch_four_sec_accuracies.extend(four_sec_accuracy)
epoch_five_sec_accuracies.extend(five_sec_accuracy)
if sub_epoch_i > 0 and (sub_epoch_i % 50) == 0:
print(
epoch_i, " ", sub_epoch_i, " loss: ",
np.array(epoch_losses[-49:]).mean(), " accuracy: ",
np.array(
epoch_overall_accuracies[(-49 *
y.shape[0]):]).mean())
sys.stdout.flush()
all_train_losses.append(np.array(epoch_losses).mean())
all_train_accuracies.append(
np.array(epoch_overall_accuracies).mean())
all_train_kill_nokill_accuracies.append(
(np.array(epoch_kill_accuracies).mean(),
np.array(epoch_no_kill_accuracies).mean()))
all_train_per_second_accuracies.append(
(np.array(epoch_one_sec_accuracies).mean(),
np.array(epoch_two_sec_accuracies).mean(),
np.array(epoch_three_sec_accuracies).mean(),
np.array(epoch_four_sec_accuracies).mean(),
np.array(epoch_five_sec_accuracies).mean()))
# reset logs for validation
epoch_losses = []
epoch_overall_accuracies = []
epoch_kill_accuracies = []
epoch_no_kill_accuracies = []
epoch_one_sec_accuracies = []
epoch_two_sec_accuracies = []
epoch_three_sec_accuracies = []
epoch_four_sec_accuracies = []
epoch_five_sec_accuracies = []
with torch.no_grad():
for X, y, death_times in validation_generator:
X = [(hero_X[0, :]).to(device) for hero_X in X]
y = torch.argmax(y[0, :], dim=1).to(device)
death_times = death_times[0]
output = model(X)
loss = criterion(output, y)
accuracy_vec = (torch.argmax(
output,
1) == y).cpu().numpy().reshape(-1).astype(np.float32)
epoch_losses.append(
loss.cpu().detach().numpy().reshape(-1)[0])
(overall_accuracy, (kill_accuracy, no_kill_accuracy),
(one_sec_accuracy, two_sec_accuracy, three_sec_accuracy,
four_sec_accuracy,
five_sec_accuracy)) = calculate_detailed_accuracies(
accuracy_vec, death_times, y)
epoch_overall_accuracies.append(overall_accuracy)
epoch_kill_accuracies.extend(kill_accuracy)
epoch_no_kill_accuracies.extend(no_kill_accuracy)
epoch_one_sec_accuracies.extend(one_sec_accuracy)
epoch_two_sec_accuracies.extend(two_sec_accuracy)
epoch_three_sec_accuracies.extend(three_sec_accuracy)
epoch_four_sec_accuracies.extend(four_sec_accuracy)
epoch_five_sec_accuracies.extend(five_sec_accuracy)
all_validation_losses.append(np.array(epoch_losses).mean())
all_validation_accuracies.append(
np.array(epoch_overall_accuracies).mean())
all_validation_kill_nokill_accuracies.append(
(np.array(epoch_kill_accuracies).mean(),
np.array(epoch_no_kill_accuracies).mean()))
all_validation_per_second_accuracies.append(
(np.array(epoch_one_sec_accuracies).mean(),
np.array(epoch_two_sec_accuracies).mean(),
np.array(epoch_three_sec_accuracies).mean(),
np.array(epoch_four_sec_accuracies).mean(),
np.array(epoch_five_sec_accuracies).mean()))
# epoch over, checkpoint, report, check validation error
print("Epoch done ", epoch_i, " loss: ",
np.array(epoch_losses).mean(), " accuracy: ",
np.array(epoch_overall_accuracies).mean())
#print("all_train_kill_nokill_accuracies ",len(all_train_kill_nokill_accuracies))
PlotValues((all_train_losses, all_validation_losses), "loss",
["train", "validation"])
PlotValues((all_train_accuracies, all_validation_accuracies),
"accuracy", ["train", "validation"])
PlotValues((*zip(*all_train_kill_nokill_accuracies),
*zip(*all_validation_kill_nokill_accuracies)),
"accuracy_kill", [
"train_kill", "train_no_kill", "validation_kill",
"validation_no_kill"
])
sec_labels = ["1_sec", "2_sec", "3_sec", "4_sec", "5_sec"]
PlotValues(
(*zip(*all_train_per_second_accuracies),
*zip(*all_validation_per_second_accuracies)), "accuracy_sec",
[
*["accuracy_train" + label for label in sec_labels],
*["accuracy_validation" + label for label in sec_labels]
])
#np.save('losses.npy', np.array(mean_losses))
#np.save('accuracies.npy', np.array(mean_accuracies))
print("Epoch took: ", time.time() - now)
sys.stdout.flush()
#PlotValues(mean_validation_accuracies,"valid_accuracy")
#PlotValues(mean_valid_overall_accuracies,"valid_overall_accuracy")
#np.save('mean_valid_overall_accuracies.npy', np.array(mean_valid_overall_accuracies))
#np.save('mean_validation_accuracies.npy', np.array(mean_validation_accuracies))
if (epoch_i % 100) == 99:
torch.save(model.state_dict(),
"model" + str(epoch_i) + ".model")
else: # Per player probability prediction
all_train_losses = []
all_train_accuracies = []
all_train_target_accuracies = []
all_train_die_notdie_accuracies = []
all_train_per_sec_accuracies = [[] for _ in range(20)]
all_train_per_sec_predictions = [[] for _ in range(20)]
all_train_per_sec_predictions_std = [[] for _ in range(20)]
all_validation_losses = []
all_validation_accuracies = []
all_validation_roc_scores = []
all_validation_pr_scores = []
for epoch_i in range(50000):
now = time.time()
np.random.seed(
) # reset seed https://github.com/pytorch/pytorch/issues/5059 data loader returns the same values
epoch_overall_loss = []
epoch_overall_accuracy = []
epoch_target_accuracy = []
epoch_die_accuracy = []
epoch_not_die_accuracy = []
epoch_per_sec_accuracies = [[] for _ in range(20)]
epoch_per_sec_predictions = [[] for _ in range(20)]
for sub_epoch_i, (X, y, death_times,
player_i) in enumerate(training_generator):
# since we get a batch of size 1 of batch of real batch size, we take the 0th element
X = [(hero_X[0, :]).to(device) for hero_X in X]
y = (y[0, :]).to(device)
death_times = death_times[0]
player_i = player_i[0].to(device)
# Forward + Backward + Optimize
optimizer.zero_grad()
output = model(X)
output = torch.sigmoid(output)
output_np = output.cpu().detach().numpy()
# only backpropagate the loss for player_i (so the training data is balanced)
loss = binary_classification_loss(output[:, player_i],
y[:, player_i])
loss.backward()
optimizer.step()
overall_loss = binary_classification_loss(
output, y).cpu().detach().numpy()
epoch_overall_loss.append(overall_loss.reshape(-1).mean())
accuracy_values = ((output > 0.5) == (
y > 0.5)).cpu().numpy().astype(np.float32)
target_accuracy = ((output[:, player_i] > 0.5) == (
y[:, player_i] > 0.5)).cpu().numpy().reshape(-1).astype(
np.float32)
die_accuracy_vec = ((output > 0.5) == (y > 0.5)).view(-1)[
y.view(-1) > 0.5].cpu().numpy().reshape(-1).astype(
np.float32)
not_die_accuracy_vec = ((output > 0.5) == (y > 0.5)).view(-1)[
y.view(-1) < 0.5].cpu().numpy().reshape(-1).astype(
np.float32)
epoch_overall_accuracy.append(
accuracy_values.reshape(-1).mean())
epoch_target_accuracy.append(target_accuracy.mean())
# these have varying size, so calculating the proper mean across batches takes more work
epoch_die_accuracy.extend(die_accuracy_vec)
epoch_not_die_accuracy.extend(not_die_accuracy_vec)
death_times = death_times.cpu().numpy()
#death_times[death_times < 0] = 1000.0 # make invalid death times a large number
for timeslot_i in range(19):
mask_die_in_timeslot = np.logical_and(
(death_times > timeslot_i), (death_times <
(timeslot_i + 1)))
epoch_per_sec_accuracies[timeslot_i].extend(
accuracy_values[mask_die_in_timeslot].reshape(-1))
epoch_per_sec_predictions[timeslot_i].extend(
output_np[mask_die_in_timeslot].reshape(-1))
# and the rest
mask_die_in_timeslot = (death_times > 19)
epoch_per_sec_accuracies[19].extend(
accuracy_values[mask_die_in_timeslot].reshape(-1))
epoch_per_sec_predictions[19].extend(
output_np[mask_die_in_timeslot].reshape(-1))
if sub_epoch_i > 0 and (sub_epoch_i % 50) == 0:
print(epoch_i, " ", sub_epoch_i, " loss: ",
np.array(epoch_overall_loss[-49:]).mean(),
" accuracy: ",
np.array(epoch_target_accuracy[-49:]).mean())
#for timeslot_i in range(19):
# print("epoch_per_sec_predictions ",len(epoch_per_sec_predictions[timeslot_i]))
#print("die accuracy: ",np.array(epoch_die_accuracy[-49:]).mean())
#print("not_die accuracy: ",np.array(epoch_not_die_accuracy[-49:]).mean())
sys.stdout.flush()
if (epoch_i % 10) == 9:
np.save('epoch_per_sec_predictions.npy',
np.array(epoch_per_sec_predictions))
all_train_losses.append(np.array(epoch_overall_loss).mean())
all_train_accuracies.append(
np.array(epoch_overall_accuracy).mean())
all_train_target_accuracies.append(
np.array(epoch_target_accuracy).mean())
all_train_die_notdie_accuracies.append(
(np.array(die_accuracy_vec).mean(),
np.array(not_die_accuracy_vec).mean()))
for timeslot_i in range(20):
all_train_per_sec_accuracies[timeslot_i].append(
np.array(epoch_per_sec_accuracies[timeslot_i]).mean())
all_train_per_sec_predictions[timeslot_i].append(
np.array(epoch_per_sec_predictions[timeslot_i]).mean())
all_train_per_sec_predictions_std[timeslot_i].append(
np.array(epoch_per_sec_predictions[timeslot_i]).std())
# VALIDATION EPOCH
if (epoch_i % 3) == 0:
epoch_overall_loss = []
epoch_overall_accuracy = []
epoch_all_pred = []
epoch_all_y = []
#epoch_die_accuracy = []
#epoch_not_die_accuracy = []
#epoch_per_sec_accuracies = [[] for _ in range(20)]
#epoch_per_sec_predictions = [[] for _ in range(20)]
with torch.no_grad():
for X, y, death_times, player_i in validation_generator:
X = [(hero_X[0, :]).to(device) for hero_X in X]
y = (y[0, :]).to(device)
death_times = death_times[0]
output = model(X)
output = torch.sigmoid(output)
output_np = output.cpu().detach().numpy()
epoch_overall_loss.append(
binary_classification_loss(
output,
y).cpu().detach().numpy().reshape(-1).mean())
accuracy_vec = ((output > 0.5) == (
y > 0.5)).cpu().numpy().reshape(-1).astype(
np.float32)
epoch_overall_accuracy.append(accuracy_vec.mean())
#death_times = death_times.cpu().numpy()
#for timeslot_i in range(19):
# mask_die_in_timeslot = np.logical_and( (death_times > timeslot_i), (death_times < (timeslot_i+1)))
# epoch_per_sec_accuracies[timeslot_i].extend(accuracy_values[mask_die_in_timeslot].reshape(-1))
# epoch_per_sec_predictions[timeslot_i].extend(output_np[mask_die_in_timeslot].reshape(-1))
epoch_all_pred.extend(output_np.reshape(-1))
epoch_all_y.extend(y.cpu().numpy().reshape(-1))
all_validation_roc_scores.append(
roc_auc_score(epoch_all_y, epoch_all_pred))
all_validation_pr_scores.append(
average_precision_score(epoch_all_y, epoch_all_pred))
all_validation_losses.append(
np.array(epoch_overall_loss).mean())
all_validation_accuracies.append(
np.array(epoch_overall_accuracy).mean())
else:
# just copy the previous validation statistics, so we can plot it togeather with training statistics
all_validation_losses.append(all_validation_losses[-1])
all_validation_accuracies.append(all_validation_accuracies[-1])
all_validation_roc_scores.append(all_validation_roc_scores[-1])
all_validation_pr_scores.append(all_validation_pr_scores[-1])
PlotValues((all_train_losses, all_validation_losses), "loss",
["train", "validation"])
PlotValues((all_train_accuracies, all_validation_accuracies),
"accuracy", ["train", "validation"])
PlotValues((all_validation_roc_scores, ), "roc_score", ["roc"])
PlotValues((all_validation_pr_scores, ), "pr_score", ["pr"])
#PlotValues((all_train_losses,),"loss",["train"])
#PlotValues((all_train_accuracies,),"accuracy",["train"])
PlotValues((all_train_target_accuracies, ), "target_accuracy",
["train"])
PlotValues(
([vals[0] for vals in all_train_die_notdie_accuracies
], [vals[1] for vals in all_train_die_notdie_accuracies]),
"all_train_die_notdie_accuracies", ["die", "not_die"])
PlotValues(all_train_per_sec_accuracies,
"all_train_per_sec_accuracies",
[str(time_i + 1) + "_sec" for time_i in range(20)])
PlotWithStd(
values=[vec[-1] for vec in all_train_per_sec_predictions],
stds=[vec[-1] for vec in all_train_per_sec_predictions_std],
legends=["per_sec predictions"],
name="per_sec predictions")
print("Epoch done ", epoch_i, " loss: ",
np.array(epoch_overall_loss).mean(), " accuracy: ",
np.array(epoch_target_accuracy).mean())
print("Epoch took: ", time.time() - now)
sys.stdout.flush()
if (epoch_i % 10) == 9:
np.save('all_train_per_sec_predictions.npy',
np.array(all_train_per_sec_predictions))
np.save('all_train_per_sec_predictions_std.npy',
np.array(all_train_per_sec_predictions_std))
if (epoch_i % 100) == 99:
torch.save(model.state_dict(),
"model" + str(epoch_i) + ".model")
def make_predictions(file, modelPath):
trainingDataFiles = [
'data.h5'
] #glob.glob("/scratch/staff/ak1774/shared_folder/data/train/*.h5")
config = get_config()
get_feature_indicies_fn = locate(config["feature_set"])
get_label_indicies_fn = locate(config["lable_set"])
example_data = data_loader.load_data_from_file(trainingDataFiles[0])
#hero_feature_indicies,label_indicies = get_feature_indicies_fn(example_data)
hero_feature_indicies = get_feature_indicies_fn(example_data)
label_indicies = get_label_indicies_fn(example_data)
print(len(hero_feature_indicies))
print(len(label_indicies))
model = load_pytorch_model(modelPath, hero_feature_indicies,
label_indicies, config)
#model = model.eval()
print(model)
data = data_loader.load_data_from_file(file)
xLims = data['time'].values
fullGameData, fullGameLabels = data_loader.getSequencialNaive(
data, hero_feature_indicies, label_indicies)
print(np.array(fullGameData).shape)
windowData = []
maxWindow = 20
#torch.set_printoptions(precision=10)
currentMeanAccuracy = 0
for i in range(0, 19326):
y = fullGameLabels[i]
y = np.array(y)
y = np.expand_dims(y, 0)
X = [torch.from_numpy(hero_X[i:(i + 1), :]) for hero_X in fullGameData]
predX = model(X)
predX = torch.sigmoid(predX)
predX = predX.cpu().detach().numpy()
accuracy_vec = ((predX > 0.5) == (y > 0.5)).reshape(-1).astype(
np.float32)
currentMeanAccuracy += np.mean(accuracy_vec)
prediction = predX
prediction = np.squeeze(prediction, 0)
#print(prediction)
#print(prediction.shape)
if i % 3000 == 0:
print('Current mean ' + str(currentMeanAccuracy / (i + 1)))
#print('----------------------')
windowData.append(prediction)
print(np.array(windowData).shape)
if len(windowData) > maxWindow:
windowData.pop(0)
#(get_average(windowData,maxWindow,5,i,y))
#(get_average(windowData,maxWindow,10,i,y))
(get_average(windowData, maxWindow, 15, i, y))
#(get_average(windowData,maxWindow,20,i,y))
heroStuff1 = np.swapaxes(heroStuff, 0, 1)
labelStuff1 = np.swapaxes(labelStuff, 0, 1)
#heroStuff1 = heroStuff
#labelStuff1 = labelStuff
#x = arange(0,len(heroStuff1[0]))
#x = np.array(x) /
splitLower = 2500
splitHigher = 3000
xLims = xLims - xLims[0] - 90
print(np.array(heroStuff1).shape)
print(np.array(labelStuff1).shape)
print(np.array(xLims).shape)
'''
heroStuff1 = heroStuff1[:,splitLower:splitHigher]
labelStuff1 = labelStuff1[:,splitLower:splitHigher]
xLims = xLims[splitLower:splitHigher]
'''
#print(heroStuff1[0])
heroStuff1 = (heroStuff1 - 1)
labelStuff1 = (labelStuff1 - 1) * -1
#print(heroStuff1[0])
#heroStuff1[:] = [x - 1 for x in heroStuff1]
#labelStuff1[:] = [(x - 1) * -1 for x in labelStuff1]
# 1:30 game start
print(np.array(heroStuff1).shape)
print(np.array(labelStuff1).shape)
print(np.array(xLims).shape)
#x = np.arange(194)
#plt.subplots_adjust(hspace=100)
#plt.xticks(np.arange(0,1,))
#fig = plt.figure(figsize=(11,8))
#ax1 = fig.add_subplot(111)
#plt.yticks(np.arange(0, 1, step=1))
for i in range(0, 10):
#ax1 = fig.add_subplot(111)
#ax1.plot(heroStuff1[i], label=1)
#ax1.plot(labelStuff1[i], label=2)
plt.subplot(10, 1, (i + 1))
plt.plot(xLims, heroStuff1[i], color='red', linewidth=0.5)
plt.plot(xLims, labelStuff1[i], color='blue', linewidth=0.5)
#plt.title('Player ' + str(i))
#ax1.legend(loc=2)
plt.savefig('smooth_plot.eps')
" num matches is ", num_matches)
sys.stdout.flush()
all_y = [[] for model_path in modelPathList]
all_pred = [[] for model_path in modelPathList]
per_sec_pred = [[[] for _ in range(20)] for model_path in modelPathList]
for i in range(match_per_worker):
match_index = first_match_index_for_this_task + i
if match_index >= num_matches:
continue
print("Loading match ", match_index)
data = data_loader.load_data_from_file(dataPathList[match_index])
# get death times
labels = [(i, label) for i, label in enumerate(list(data))]
death_time_indicies = preprocess.labels_to_indicies(
preprocess.select_features_by_name("time_until_next_death", labels))
death_times = data.values[:, death_time_indicies].astype(np.float32)
for model_i, (model_path,
config_path) in enumerate(zip(modelPathList,
configPathList)):
with open(config_path) as f:
config = commentjson.load(f)
modeldata = test_model.load_pytorch_model(model_path, config, data)
def get_predictions(self, prediction_type="regression"):
if self.model == None:
print 'no evaluation possible since no model was provided.'
return
try:
model = serial.load(self.model)
except Exception as e:
print("error loading {}:".format(self.model))
print(e)
raise Exception("error loading {}:".format(self.model))
X = model.get_input_space().make_theano_batch()
Y = model.fprop(X)
if prediction_type == "classification":
Y = theanoTensor.argmax(Y, axis=1)
else:
assert prediction_type == "regression"
f = theanoFunction([X], Y, allow_input_downcast=True)
print("loading data and predicting...")
data = data_loader.Data('test')
if self.direction == 'reverse':
input_data = data.response_data
output_path = self.stimulus_data_path
m, r, c = input_data.shape
input_data = input_data.reshape((m, r, c, 1))
elif self.direction == 'forward':
input_data = data.stimulus_data
output_path = self.response_data_path
else:
raise Exception(
'Specify either "reverse" or "forward" as direction.')
prediction = f(input_data)
print("writing predictions...")
if output_path[-4:] == '.mat':
scipy.io.savemat(output_path, {'data': prediction})
elif output_path[-4:] == '.txt':
np.savetxt(output_path, prediction)
elif output_path[-4:] == '.npy':
np.save(output_path, prediction)
else:
raise Exception(
'Only ".mat", ".txt" and ".npy" files are supported as data formats.'
)
#------------------------------------------------------------------------------
# for testing
#------------------------------------------------------------------------------
if self.debug:
try:
test_output = np.squeeze(
data_loader.load_data_from_file(
'testUnattendedAudioOrg.mat'))
test_predictions = np.squeeze(
data_loader.load_data_from_file(self.stimulus_data_path))
print 'DEBUG: test_predictions.shape', test_predictions.shape, 'test_output.shape', test_output.shape
print 'DEBUG: test_predictions.max', np.max(
test_predictions), np.argmax(test_predictions)
print 'DEBUG: ', np.corrcoef(test_predictions, test_output)
except:
print 'Detailed correlation analysis only possible for the hello world example.'
pass
def main():
#------------------------------------------------------------------------------
# set parameters
#------------------------------------------------------------------------------
train_network = None
yaml = None
stimulus_data = None
response_data = None
weight_path = None
visualize = False
context_length = 25
direction = 'reverse'
valid_data_entries = -1 # this encodes the default to use all entries
forking = False
verbosity = 1
debug = 0
num_training_epochs = 100
num_neurons = 5
#------------------------------------------------------------------------------
# parse arguments
#------------------------------------------------------------------------------
def usage():
print 'Call python DNNRegression.py with the following arguments: \n', \
'{--train, --predict} \n', \
'-m model.yaml \n', \
'-s stimulus_data [i/o, default text, or .mat file] \n', \
'-r response_data [i/o, default text, or .mat file] \n', \
'-w weights [i/o, default text, or .pkl] \n', \
'[--visual visualize.png] \n', \
'[--context N for N>=1 *25] \n', \
'[--numEpochs N default is 100]', \
'[--numNeurons N default is 5]', \
'[--dir forward/reverse*] \n', \
'[--valid which parts are valid, in case of concatenating trials, default all valid.] \n', \
'[--forking for matlab sake] \n'
try:
opts, _ = getopt.getopt(sys.argv[1:], "htpm:s:r:w:", [
"help", "train", "predict", "model=", "stimulus=", "response=",
"weights=", "visual=", "context=", "dir=", "valid=", "forking=",
"verbosity=", "debug", "numEpochs=", "numNeurons=", "numChannels="
])
except getopt.GetoptError as err:
print str(err) # will print something like "option -a not recognized"
usage()
sys.exit(2)
for opt, arg in opts:
if opt in ("-h", "--help"):
usage()
sys.exit()
elif opt in ("-p", "--predict"):
train_network = 0
elif opt in ("-t", "--train"):
train_network = 1
elif opt in ("-m", "--model"):
yaml = arg
elif opt in ("-s", "--stimulus"):
stimulus_data = arg
elif opt in ("-r", "--response"):
response_data = arg
elif opt in ("-w", "--weights"):
weight_path = arg
elif opt in ("--visual"):
visualize = arg
elif opt in ("--context"):
context_length = int(arg)
elif opt in ("--numEpochs"):
num_training_epochs = int(arg)
elif opt in ("--numNeurons"):
num_neurons = int(arg)
elif opt in ("--dir"):
direction = arg
elif opt in ("--valid"):
valid_data_entries = arg
elif opt in ("--forking"):
print "forking is not supported for now"
# forking = True
elif opt in ("--verbosity"):
verbosity = arg
elif opt in ("--debug"):
debug = 1
else:
assert False, "unhandled option"
if train_network is None:
print 'You need to define wheter to predict outputs or to train the network.'
usage()
sys.exit(2)
if yaml is None:
print 'You need to define a valid model path / yaml file.'
usage()
sys.exit(2)
if stimulus_data is None:
print 'You need to define a valid stimulus data path.'
usage()
sys.exit(2)
if response_data is None:
print 'You need to define a valid response data path.'
usage()
sys.exit(2)
if weight_path is None:
print 'You need to define a valid path containing the weights of the mode.'
usage()
sys.exit(2)
#------------------------------------------------------------------------------
# Save environment for other parts of the DNN
#------------------------------------------------------------------------------
os.environ["EEGTOOLS_TRAIN_NETWORK"] = str(train_network)
os.environ["EEGTOOLS_STIMULUS_DATA_PATH"] = stimulus_data
os.environ["EEGTOOLS_RESPONSE_DATA_PATH"] = response_data
os.environ["EEGTOOLS_VALID_DATA_PATH"] = str(valid_data_entries)
os.environ["EEGTOOLS_CONTEXT_LENGTH"] = str(context_length)
os.environ["EEGTOOLS_DIRECTION"] = direction
os.environ["EEGTOOLS_DEBUG"] = str(debug)
# determine number of channels from data
if direction == 'forward':
input_path = stimulus_data
else:
input_path = response_data
input_data = data_loader.load_data_from_file(input_path)
num_channels = input_data.shape[1]
os.environ["EEGTOOLS_NUM_CHANNELS"] = str(num_channels)
#------------------------------------------------------------------------------
# Save imports for the YAML file
#------------------------------------------------------------------------------
pkl.dump(context_length, open('context_length.pkl', 'wb'))
pkl.dump(num_training_epochs, open('num_training_epochs.pkl', 'wb'))
pkl.dump(num_neurons, open('num_neurons.pkl', 'wb'))
pkl.dump(num_channels, open('num_channels.pkl', 'wb'))
#------------------------------------------------------------------------------
# train / run DNN
#------------------------------------------------------------------------------
net = DNNregression(train_network, yaml, stimulus_data, response_data,
weight_path, visualize, context_length, direction,
valid_data_entries, forking, verbosity, debug)
if net.train_network:
net.train()
net.save_weights()
else:
net.get_predictions()
if net.visualize:
net.show_weights()
def get_predictions(self, prediction_type="regression"):
if self.model == None:
print 'no evaluation possible since no model was provided.'
return
try:
model = serial.load(self.model)
except Exception as e:
print("error loading {}:".format(self.model))
print(e)
raise Exception("error loading {}:".format(self.model))
X = model.get_input_space().make_theano_batch()
Y = model.fprop(X)
if prediction_type == "classification":
Y = theanoTensor.argmax(Y, axis=1)
else:
assert prediction_type == "regression"
f = theanoFunction([X], Y, allow_input_downcast=True)
print("loading data and predicting...")
data = data_loader.Data('test')
if self.direction == 'reverse':
input_data = data.response_data
output_path = self.stimulus_data_path
m, r, c = input_data.shape
input_data = input_data.reshape((m, r, c , 1))
elif self.direction == 'forward':
input_data = data.stimulus_data
output_path = self.response_data_path
else:
raise Exception('Specify either "reverse" or "forward" as direction.')
prediction = f(input_data)
print("writing predictions...")
if output_path[-4:] == '.mat':
scipy.io.savemat(output_path, {'data': prediction})
elif output_path[-4:] == '.txt':
np.savetxt(output_path, prediction)
elif output_path[-4:] == '.npy':
np.save(output_path, prediction)
else:
raise Exception('Only ".mat", ".txt" and ".npy" files are supported as data formats.')
#------------------------------------------------------------------------------
# for testing
#------------------------------------------------------------------------------
if self.debug:
try:
test_output = np.squeeze(data_loader.load_data_from_file('testUnattendedAudioOrg.mat'))
test_predictions = np.squeeze(data_loader.load_data_from_file(self.stimulus_data_path))
print 'DEBUG: test_predictions.shape', test_predictions.shape, 'test_output.shape', test_output.shape
print 'DEBUG: test_predictions.max', np.max(test_predictions), np.argmax(test_predictions)
print 'DEBUG: ', np.corrcoef(test_predictions, test_output)
except:
print 'Detailed correlation analysis only possible for the hello world example.'
pass
# setup jupiter kernel with requirements
# import sys
# !{sys.executable} -m pip install numpy scipy eikon matplotlib cvxopt
"""Una volta predisposto l'ambiente, carichiamo i dati precedentemente scaricati da Eikon ed organizzati in file .json. Sono presenti quattro diversi scenari:
1. Scenario 'assets-test.json': file utilizzato ai fini del testing del codice, 3 soli simboli, 6 soli periodi mensili, date corrette ma fittizie. Questo file è utile solo ai fini del testing del codice.
2. Scenario 'assets-small.json': file di piccole dimensioni (periodo semestrale, simboli provenienti da DAX30, CAC40 ed IBEX 30, frequenza mensile)
3. Scenario 'assets-large.json': file di medie dimensioni (periodo decennale, simboli provenienti da DAX30, CAC40 ed IBEX 30, frequenza mensile)
4. Scenario 'assets-production.json': file di grandi dimensioni (periodo ventennale, simboli provenienti da DAX30, CAC40 ed IBEX 30, frequenza mensile)
Nel codice sono presenti funzionalità dedicate alla creazione e manipolazione di ulteriori insiemi di simboli e relativi performance temporali.
"""
# load data from file
ek_data_processed = dl.load_data_from_file('assets-test.json')
target_assets = ek_data_processed['target_assets']
print("Assets list: ")
print(target_assets)
print("Date from: " + ek_data_processed["start_date"])
print("Date to: " + ek_data_processed["end_date"])
print("Number of assets: " + str(len(target_assets)))
print("Number of time samples: " +
str(len(ek_data_processed[target_assets[0]])))
"""## Assets
Vengono qui caricati il RIC dell'asset e la relativa serie temporale di ritorni menili %.
```
assets.add_or_modify_asset(asset_name, ek_data_processed[asset_name])
```
Una volta caricati tutti gli asset viene calcolata la matrice di covarianza.
```
def main():
#------------------------------------------------------------------------------
# set parameters
#------------------------------------------------------------------------------
train_network = None
yaml = None
stimulus_data = None
response_data = None
weight_path = None
visualize = False
context_length = 25
direction = 'reverse'
valid_data_entries = -1 # this encodes the default to use all entries
forking = False
verbosity = 1
debug = 0
num_training_epochs = 100
num_neurons = 5
#------------------------------------------------------------------------------
# parse arguments
#------------------------------------------------------------------------------
def usage():
print 'Call python DNNRegression.py with the following arguments: \n', \
'{--train, --predict} \n', \
'-m model.yaml \n', \
'-s stimulus_data [i/o, default text, or .mat file] \n', \
'-r response_data [i/o, default text, or .mat file] \n', \
'-w weights [i/o, default text, or .pkl] \n', \
'[--visual visualize.png] \n', \
'[--context N for N>=1 *25] \n', \
'[--numEpochs N default is 100]', \
'[--numNeurons N default is 5]', \
'[--dir forward/reverse*] \n', \
'[--valid which parts are valid, in case of concatenating trials, default all valid.] \n', \
'[--forking for matlab sake] \n'
try:
opts, _ = getopt.getopt(sys.argv[1:],
"htpm:s:r:w:", ["help", "train", "predict", "model=",
"stimulus=", "response=", "weights=",
"visual=", "context=", "dir=", "valid=",
"forking=", "verbosity=", "debug",
"numEpochs=", "numNeurons=", "numChannels="])
except getopt.GetoptError as err:
print str(err) # will print something like "option -a not recognized"
usage()
sys.exit(2)
for opt, arg in opts:
if opt in ("-h", "--help"):
usage()
sys.exit()
elif opt in ("-p", "--predict"):
train_network = 0
elif opt in ("-t", "--train"):
train_network = 1
elif opt in ("-m", "--model"):
yaml = arg
elif opt in ("-s", "--stimulus"):
stimulus_data = arg
elif opt in ("-r", "--response"):
response_data = arg
elif opt in ("-w", "--weights"):
weight_path = arg
elif opt in ("--visual"):
visualize = arg
elif opt in ("--context"):
context_length = int(arg)
elif opt in ("--numEpochs"):
num_training_epochs = int(arg)
elif opt in ("--numNeurons"):
num_neurons = int(arg)
elif opt in ("--dir"):
direction = arg
elif opt in ("--valid"):
valid_data_entries = arg
elif opt in ("--forking"):
print "forking is not supported for now"
# forking = True
elif opt in ("--verbosity"):
verbosity = arg
elif opt in ("--debug"):
debug = 1
else:
assert False, "unhandled option"
if train_network is None:
print 'You need to define wheter to predict outputs or to train the network.'
usage()
sys.exit(2)
if yaml is None:
print 'You need to define a valid model path / yaml file.'
usage()
sys.exit(2)
if stimulus_data is None:
print 'You need to define a valid stimulus data path.'
usage()
sys.exit(2)
if response_data is None:
print 'You need to define a valid response data path.'
usage()
sys.exit(2)
if weight_path is None:
print 'You need to define a valid path containing the weights of the mode.'
usage()
sys.exit(2)
#------------------------------------------------------------------------------
# Save environment for other parts of the DNN
#------------------------------------------------------------------------------
os.environ["EEGTOOLS_TRAIN_NETWORK"] = str(train_network)
os.environ["EEGTOOLS_STIMULUS_DATA_PATH"] = stimulus_data
os.environ["EEGTOOLS_RESPONSE_DATA_PATH"] = response_data
os.environ["EEGTOOLS_VALID_DATA_PATH"] = str(valid_data_entries)
os.environ["EEGTOOLS_CONTEXT_LENGTH"] = str(context_length)
os.environ["EEGTOOLS_DIRECTION"] = direction
os.environ["EEGTOOLS_DEBUG"] = str(debug)
# determine number of channels from data
if direction == 'forward':
input_path = stimulus_data
else:
input_path = response_data
input_data = data_loader.load_data_from_file(input_path)
num_channels = input_data.shape[1]
os.environ["EEGTOOLS_NUM_CHANNELS"] = str(num_channels)
#------------------------------------------------------------------------------
# Save imports for the YAML file
#------------------------------------------------------------------------------
pkl.dump(context_length, open( 'context_length.pkl', 'wb'))
pkl.dump(num_training_epochs, open( 'num_training_epochs.pkl', 'wb'))
pkl.dump(num_neurons, open( 'num_neurons.pkl', 'wb'))
pkl.dump(num_channels, open( 'num_channels.pkl', 'wb'))
#------------------------------------------------------------------------------
# train / run DNN
#------------------------------------------------------------------------------
net = DNNregression(train_network, yaml, stimulus_data, response_data,
weight_path, visualize, context_length, direction,
valid_data_entries, forking, verbosity, debug)
if net.train_network:
net.train()
net.save_weights()
else:
net.get_predictions()
if net.visualize:
net.show_weights()