def training(num_batches, batch_size, x_train, label_train, mask_train):
for i in range(num_batches):
idx = range(i * batch_size, (i + 1) * batch_size)
x_batch = x_train[idx]
y_batch = label_train[idx]
mask_batch = mask_train[idx]
loss, out, batch_norm = train(x_batch, y_batch, mask_batch)
norms.append(batch_norm)
preds.append(out)
losses.append(loss)
predictions = np.concatenate(preds, axis=0)
loss_train = np.mean(losses)
all_losses_train.append(loss_train)
acc_train = utils.proteins_acc(predictions,
label_train[0:num_batches * batch_size],
mask_train[0:num_batches * batch_size])
print('acc_train: ', acc_train)
all_accuracy_train.append(acc_train)
mean_norm = np.mean(norms)
all_mean_norm.append(mean_norm)
print " average training loss: %.5f" % loss_train
print " average training accuracy: %.5f" % acc_train
print " average norm: %.5f" % mean_norm
def testing(num_batches, batch_size, X, y, mask):
for i in range(num_batches):
idx = range(i * batch_size, (i + 1) * batch_size)
x_batch = X[idx]
y_batch = y[idx]
mask_batch = mask[idx]
loss, out = evaluate(x_batch, y_batch, mask_batch)
preds.append(out)
losses.append(loss)
predictions = np.concatenate(preds, axis=0)
loss_eval = np.mean(losses)
all_losses.append(loss_eval)
acc_eval = utils.proteins_acc(predictions, y, mask)
all_accuracy.append(acc_eval)
print("Average evaluation loss ({}): {:.5f}".format(subset, loss_eval))
print("Average evaluation accuracy ({}): {:.5f}".format(
subset, acc_eval))
return i
def validate(sess):
gen = data_gen.gen_valid
valid_masks = []
valid_outs = []
valid_targets = []
sum = 0
for batch, i in gen():
valid_fetches = [prediction]
valid_feed_dict = {
X_input: batch['X'],
t_input: batch['t'],
X_length: batch['length'],
t_mask: batch['mask'],
is_training_pl: False
}
valid_out = sess.run(fetches=valid_fetches,
feed_dict=valid_feed_dict)[0]
h_out = np.zeros((i, 700, 8), dtype="float32")
h_out[:, :valid_out.shape[1], :] = valid_out
h_mask = np.zeros((i, 700), dtype="float32")
h_mask[:, :valid_out.shape[1]] = batch['mask']
h_targets = np.zeros((i, 700), dtype="int32")
h_targets[:, :valid_out.shape[1]] = batch['t']
valid_masks.append(h_mask)
valid_targets.append(h_targets)
valid_outs.append(h_out)
sum += i
valid_outs = np.concatenate(valid_outs, axis=0)[:sum]
valid_targets = np.concatenate(valid_targets, axis=0)[:sum]
valid_masks = np.concatenate(valid_masks, axis=0)[:sum]
valid_accs = utils.proteins_acc(valid_outs, valid_targets,
valid_masks)
print(" valid_accs,", valid_accs)
sum_fetches = [val_summaries, global_step]
sum_feed_dict = {
valid_accs_pl: valid_accs,
}
summaries, i = sess.run(sum_fetches, sum_feed_dict)
summary_writer.add_summary(summaries, i)
import utils
if len(sys.argv) < 2:
sys.exit("Usage: python eval_avrg.py <predictions_path> [subset=test]")
predictions_path_all = glob.glob(sys.argv[1] + "*")
mybool = False
for predictions_path in predictions_path_all:
print(predictions_path)
if not mybool:
predictions = np.load(predictions_path)
mybool = True
else:
predictions = predictions + np.load(predictions_path)
import Data_Manipulator
if len(sys.argv) == 3:
subset = sys.argv[2]
assert subset in ['train', 'valid', 'test', 'test_valid']
else:
subset = 'test'
if subset == "test":
_, mask, y, _ = Data_Manipulator.get_test()
acc = utils.proteins_acc(predictions, y, mask)
print "Accuracy (%s) is: %.5f" % (subset, acc)
else:
subset = 'test'
if subset == "test":
_, mask, y, _ = data.get_test()
elif subset == "train":
y = data.labels_train
mask = data.mask_train
elif subset == "train_valid":
y = data.labels
mask = data.mask
else:
y = data.labels_valid
mask = data.mask_valid
acc = utils.proteins_acc(predictions, y, mask)
print "Accuracy (%s) is: %.5f" % (subset,acc)
## Alternative model avrg!! ##
john = np.zeros((640,700,8))
for predictions_path in predictions_path_all:
print(predictions_path)
predictions = np.load(predictions_path)#.ravel()
predictions = np.argmax(predictions, axis=2)
for i in range(640):
for j in range(700):
num = predictions[i, j]
john[i, j, num] = john[i, j, num] + 1
nn.layers.set_all_param_values(l_out, metadata['param_values'])
print "Compile functions"
predict = theano.function([sym_x, sym_mask], inference)
print "Predict"
predictions = []
batch_size = config.batch_size
num_batches = np.size(X, axis=0) // batch_size
for i in range(num_batches):
idx = range(i * batch_size, (i + 1) * batch_size)
x_batch = X[idx]
mask_batch = mask[idx]
p = predict(x_batch, mask_batch)
predictions.append(p)
predictions = np.concatenate(predictions, axis=0)
predictions_path = os.path.join(
"predictions",
os.path.basename(metadata_path).replace("dump_",
"predictions_").replace(
".pkl", ".npy"))
print(utils.proteins_acc(predictions, data.labels_test, data.mask_test))
print "Storing predictions in %s" % predictions_path
np.save(predictions_path, predictions)
def main():
sym_y = T.imatrix('target_output')
sym_mask = T.matrix('mask')
sym_x = T.tensor3()
TOL = 1e-5
num_epochs = config.epochs
batch_size = config.batch_size
#### DATA ####
# print "@@@@TESTING@@@@"
# l_in = nn.layers.InputLayer(shape=(None, 700, 42))
# l_dim_a = nn.layers.DimshuffleLayer(
# l_in, (0,2,1))
# l_conv_a = nn.layers.Conv1DLayer(
# incoming=l_dim_a, num_filters=42, border_mode='same',
# filter_size=3, stride=1, nonlinearity=nn.nonlinearities.rectify)
# l_dim_b = nn.layers.DimshuffleLayer(
# l_conv_a, (0,2,1))
# out = nn.layers.get_output(l_dim_b, sym_x)
# testvar = np.ones((128, 700, 42)).astype('float32')
# print "@@@@EVAL@@@@"
# john = out.eval({sym_x: testvar})
# print("Johns shape")
# print(john.shape)
print("Building network ...")
##########################DEBUG##########################
l_in, l_out = config.build_model()
##########################DEBUG##########################
all_layers = nn.layers.get_all_layers(l_out)
num_params = nn.layers.count_params(l_out)
print(" number of parameters: %d" % num_params)
print(" layer output shapes:")
for layer in all_layers:
name = layer.__class__.__name__
print(" %s %s" % (name, nn.layers.get_output_shape(layer)))
print("Creating cost function")
# lasagne.layers.get_output produces a variable for the output of the net
out_train = nn.layers.get_output(
l_out, sym_x, deterministic=False)
# testvar = np.ones((128, 700, 42)).astype('float32')
# john = out_train.eval({sym_x: testvar})
# print("@@@@@JOHN@@@@@")
# print(john.shape)
# print(john.reshape((-1, num_classes)).shape)
print("Creating eval function")
out_eval = nn.layers.get_output(
l_out, sym_x, deterministic=True)
probs_flat = out_train.reshape((-1, num_classes))
lambda_reg = config.lambda_reg
all_params = nn.layers.get_all_params(l_out)
for i, p in enumerate(all_params):
if p.ndim == 3:
values = p.get_value()
if side == 'right':
values[..., int(values.shape[2] / 2.0 - 0.5):] = 0
p.set_value(values)
all_params[i] = p[..., : int(values.shape[2] / 2.0 - 0.5)]
else:
values[..., : int(values.shape[2] / 2.0 + 0.5)] = 0
p.set_value(values)
all_params[i] = p[..., int(values.shape[2] / 2.0 + 0.5):]
params = [el for el in all_params if el.name == "W" or el.name == "gamma"]
reg_term = sum(T.sum(p ** 2) for p in params)
cost = T.nnet.categorical_crossentropy(T.clip(probs_flat, TOL, 1 - TOL), sym_y.flatten())
cost = T.sum(cost * sym_mask.flatten()) / T.sum(sym_mask) + lambda_reg * reg_term
# Retrieve all parameters from the network
all_params = [el for el in all_params if el.name == "W" or el.name == "gamma" or el.name == "beta"]
# Setting the weights
if hasattr(config, 'set_weights'):
nn.layers.set_all_param_values(l_out, config.set_weights())
# Compute SGD updates for training
print("Computing updates ...")
if hasattr(config, 'learning_rate_schedule'):
learning_rate_schedule = config.learning_rate_schedule # Import learning rate schedule
else:
learning_rate_schedule = {0: config.learning_rate}
learning_rate = theano.shared(np.float32(learning_rate_schedule[0]))
all_grads = T.grad(cost, all_params)
cut_norm = config.cut_grad
updates, norm_calc = nn.updates.total_norm_constraint(all_grads, max_norm=cut_norm, return_norm=True)
if optimizer == "rmsprop":
updates = nn.updates.rmsprop(updates, all_params, learning_rate)
elif optimizer == "adadelta":
updates = nn.updates.adadelta(updates, all_params, learning_rate)
elif optimizer == "adagrad":
updates = nn.updates.adagrad(updates, all_params, learning_rate)
elif optimizer == "nag":
momentum_schedule = config.momentum_schedule
momentum = theano.shared(np.float32(momentum_schedule[0]))
updates = nn.updates.nesterov_momentum(updates, all_params, learning_rate, momentum)
else:
sys.exit("please choose either <rmsprop/adagrad/adadelta/nag> in configfile")
# Theano functions for training and computing cost
print ("config.batch_size %d" % batch_size)
print ("data.num_classes %d" % num_classes)
if hasattr(config, 'build_model'):
print("has build model")
print("Compiling train ...")
# Use this for training (see deterministic = False above)
train = theano.function(
[sym_x, sym_y, sym_mask], [cost, out_train, norm_calc], updates=updates)
print("Compiling eval ...")
# use this for eval (deterministic = True + no updates)
eval = theano.function([sym_x, sym_y, sym_mask], [cost, out_eval])
# Start timers
start_time = time.time()
prev_time = start_time
all_losses_train = []
all_accuracy_train = []
all_losses_eval_train = []
all_losses_eval_valid = []
all_losses_eval_test = []
all_accuracy_eval_train = []
all_accuracy_eval_valid = []
all_accuracy_eval_test = []
all_mean_norm = []
import data
X_train, X_valid, y_train, y_valid, mask_train, mask_valid, num_seq_train \
= data.get_train()
X_train, X_valid = X_train[..., 21:], X_valid[..., 21:] # Only train with pssm scores
print("y shape")
print(y_valid.shape)
print("X shape")
print(X_valid.shape)
# Start training
for i in range(y_train.shape[0]):
for j in range(y_train.shape[1]):
if y_train[i][j] == 5:
y_train[i][j] = 1
else:
y_train[i][j] = 0
for i in range(y_valid.shape[0]):
for j in range(y_valid.shape[1]):
if y_valid[i][j] == 5:
y_valid[i][j] = 1
else:
y_valid[i][j] = 0
for epoch in range(num_epochs):
if (epoch % 10) == 0:
print ("Epoch %d of %d" % (epoch + 1, num_epochs))
if epoch in learning_rate_schedule:
lr = np.float32(learning_rate_schedule[epoch])
print (" setting learning rate to %.7f" % lr)
learning_rate.set_value(lr)
if optimizer == "nag":
if epoch in momentum_schedule:
mu = np.float32(momentum_schedule[epoch])
print (" setting learning rate to %.7f" % mu)
momentum.set_value(mu)
# print "Shuffling data"
seq_names = np.arange(0, num_seq_train)
np.random.shuffle(seq_names)
X_train = X_train[seq_names]
y_train = y_train[seq_names]
mask_train = mask_train[seq_names]
num_batches = num_seq_train // batch_size
losses = []
preds = []
norms = []
for i in range(num_batches):
idx = range(i * batch_size, (i + 1) * batch_size)
x_batch = X_train[idx]
y_batch = y_train[idx]
mask_batch = mask_train[idx]
loss, out, batch_norm = train(x_batch, y_batch, mask_batch)
# print(batch_norm)
norms.append(batch_norm)
preds.append(out)
losses.append(loss)
# if ((i+1) % config.write_every_batch == 0) | (i == 0):
# if i == 0:
# start_place = 0
# else:
# start_place = i-config.write_every_batch
# print "Batch %d of %d" % (i + 1, num_batches)
# print " curbatch training loss: %.5f" % np.mean(losses[start_place:(i+1)])
# print " curbatch training acc: %.5f" % np.mean(accuracy[start_place:(i+1)])
predictions = np.concatenate(preds, axis=0)
loss_train = np.mean(losses)
all_losses_train.append(loss_train)
acc_train = utils.proteins_acc(predictions, y_train[0:num_batches * batch_size],
mask_train[0:num_batches * batch_size])
all_accuracy_train.append(acc_train)
mean_norm = np.mean(norms)
all_mean_norm.append(mean_norm)
if 1 == 1:
print (" average training loss: %.5f" % loss_train)
print (" average training accuracy: %.5f" % acc_train)
print (" average norm: %.5f" % mean_norm)
sets = [ # ('train', X_train, y_train, mask_train, all_losses_eval_train, all_accuracy_eval_train),
('valid', X_valid, y_valid, mask_valid, all_losses_eval_valid, all_accuracy_eval_valid)]
for subset, X, y, mask, all_losses, all_accuracy in sets:
print (" validating: %s loss" % subset)
preds = []
num_batches = np.size(X, axis=0) // config.batch_size
for i in range(num_batches): ## +1 to get the "rest"
# print(i)
idx = range(i * batch_size, (i + 1) * batch_size)
x_batch = X[idx]
y_batch = y[idx]
mask_batch = mask[idx]
loss, out = eval(x_batch, y_batch, mask_batch)
preds.append(out)
# acc = utils.proteins_acc(out, y_batch, mask_batch)
losses.append(loss)
# accuracy.append(acc)
predictions = np.concatenate(preds, axis=0)
# print " pred"
# print(predictions.shape)
# print(predictions.dtype)
loss_eval = np.mean(losses)
all_losses.append(loss_eval)
# acc_eval = np.mean(accuracy)
acc_eval = utils.proteins_acc(predictions, y, mask)
all_accuracy.append(acc_eval)
print (" average evaluation loss (%s): %.5f" % (subset, loss_eval))
print (" average evaluation accuracy (%s): %.5f" % (subset, acc_eval))
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_prev * (num_epochs - epoch)
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print (" %s since start (%.2f s)" % (utils.hms(time_since_start), time_since_prev))
print (" estimated %s to go (ETA: %s)" % (utils.hms(est_time_left), eta_str))
print()
if (epoch >= config.start_saving_at) and ((epoch % config.save_every) == 0):
print (" saving parameters and metadata")
with open((metadata_path + side + "-%d" % (epoch) + ".pkl"), 'wb') as f:
pickle.dump({
'config_name': config_name,
'param_values': nn.layers.get_all_param_values(l_out),
'losses_train': all_losses_train,
'accuracy_train': all_accuracy_train,
'losses_eval_train': all_losses_eval_train,
'losses_eval_valid': all_losses_eval_valid,
'losses_eval_test': all_losses_eval_test,
'accuracy_eval_valid': all_accuracy_eval_valid,
'accuracy_eval_train': all_accuracy_eval_train,
'accuracy_eval_test': all_accuracy_eval_test,
'mean_norm': all_mean_norm,
'time_since_start': time_since_start,
'i': i,
}, f, pickle.HIGHEST_PROTOCOL)
print (" stored in %s" % metadata_path)
print()
def main():
sym_y = T.imatrix('target_output')
sym_mask = T.matrix('mask')
sym_x = T.tensor3()
TOL = 1e-5
num_epochs = config.epochs
batch_size = config.batch_size
#### DATA ####
# print "@@@@TESTING@@@@"
# l_in = nn.layers.InputLayer(shape=(None, 700, 42))
# l_dim_a = nn.layers.DimshuffleLayer(
# l_in, (0,2,1))
# l_conv_a = nn.layers.Conv1DLayer(
# incoming=l_dim_a, num_filters=42, border_mode='same',
# filter_size=3, stride=1, nonlinearity=nn.nonlinearities.rectify)
# l_dim_b = nn.layers.DimshuffleLayer(
# l_conv_a, (0,2,1))
# out = nn.layers.get_output(l_dim_b, sym_x)
# testvar = np.ones((128, 700, 42)).astype('float32')
# print "@@@@EVAL@@@@"
# john = out.eval({sym_x: testvar})
# print("Johns shape")
# print(john.shape)
print("Building network ...")
l_in, l_out = config.build_model()
all_layers = nn.layers.get_all_layers(l_out)
num_params = nn.layers.count_params(l_out)
print(" number of parameters: %d" % num_params)
print(" layer output shapes:")
for layer in all_layers:
name = string.ljust(layer.__class__.__name__, 32)
print(" %s %s" % (name, nn.layers.get_output_shape(layer)))
print("Creating cost function")
# lasagne.layers.get_output produces a variable for the output of the net
out_train = nn.layers.get_output(
l_out, sym_x, mask=sym_mask, deterministic=False)
# testvar = np.ones((128, 700, 42)).astype('float32')
# john = out_train.eval({sym_x: testvar})
# print("@@@@@JOHN@@@@@")
# print(john.shape)
# print(john.reshape((-1, num_classes)).shape)
out_eval = nn.layers.get_output(
l_out, sym_x, mask=sym_mask, deterministic=True)
probs_flat = out_train.reshape((-1, num_classes))
lambda_reg = config.lambda_reg
params = nn.layers.get_all_params(l_out, regularizable=True)
reg_term = sum(T.sum(p**2) for p in params)
cost = T.nnet.categorical_crossentropy(T.clip(probs_flat, TOL, 1-TOL), sym_y.flatten())
cost = T.sum(cost*sym_mask.flatten()) / T.sum(sym_mask) + lambda_reg * reg_term
# Retrieve all parameters from the network
all_params = nn.layers.get_all_params(l_out, trainable=True)
# Setting the weights
if hasattr(config, 'set_weights'):
nn.layers.set_all_param_values(l_out, config.set_weights())
# Compute SGD updates for training
print("Computing updates ...")
if hasattr(config, 'learning_rate_schedule'):
learning_rate_schedule = config.learning_rate_schedule # Import learning rate schedule
else:
learning_rate_schedule = { 0: config.learning_rate }
learning_rate = theano.shared(np.float32(learning_rate_schedule[0]))
all_grads = T.grad(cost, all_params)
cut_norm = config.cut_grad
updates, norm_calc = nn.updates.total_norm_constraint(all_grads, max_norm=cut_norm, return_norm=True)
if optimizer == "rmsprop":
updates = nn.updates.rmsprop(updates, all_params, learning_rate)
elif optimizer == "adadelta":
updates = nn.updates.adadelta(updates, all_params, learning_rate)
elif optimizer == "adagrad":
updates = nn.updates.adagrad(updates, all_params, learning_rate)
elif optimizer == "nag":
momentum_schedule = config.momentum_schedule
momentum = theano.shared(np.float32(momentum_schedule[0]))
updates = nn.updates.nesterov_momentum(updates, all_params, learning_rate, momentum)
else:
sys.exit("please choose either <rmsprop/adagrad/adadelta/nag> in configfile")
# Theano functions for training and computing cost
print "config.batch_size %d" %batch_size
print "data.num_classes %d" %num_classes
if hasattr(config, 'build_model'):
print("has build model")
print("Compiling functions ...")
# Use this for training (see deterministic = False above)
train = theano.function(
[sym_x, sym_y, sym_mask], [cost, out_train, norm_calc], updates=updates)
# use this for eval (deterministic = True + no updates)
eval = theano.function([sym_x, sym_y, sym_mask], [cost, out_eval])
# Start timers
start_time = time.time()
prev_time = start_time
all_losses_train = []
all_accuracy_train = []
all_losses_eval_train = []
all_losses_eval_valid = []
all_losses_eval_test = []
all_accuracy_eval_train = []
all_accuracy_eval_valid = []
all_accuracy_eval_test = []
all_mean_norm = []
import data
X_train = data.X_train
X_valid = data.X_valid
X_test = data.X_test
y_train = data.labels_train
y_valid = data.labels_valid
y_test = data.labels_test
mask_train = data.mask_train
mask_valid = data.mask_valid
mask_test = data.mask_test
print("y shape")
print(y_valid.shape)
print("X shape")
print(X_valid.shape)
# Start training
if config.batch_norm:
collect_out = nn.layers.get_output(l_out, sym_x, deterministic=True, collect=True)
f_collect = theano.function([sym_x],
[collect_out])
for epoch in range(num_epochs):
if (epoch % 10) == 0:
print "Epoch %d of %d" % (epoch + 1, num_epochs)
if epoch in learning_rate_schedule:
lr = np.float32(learning_rate_schedule[epoch])
print " setting learning rate to %.7f" % lr
learning_rate.set_value(lr)
if optimizer == "nag":
if epoch in momentum_schedule:
mu = np.float32(momentum_schedule[epoch])
print " setting learning rate to %.7f" % mu
momentum.set_value(mu)
print "Shuffling data"
seq_names = np.arange(0,data.num_seq_train)
np.random.shuffle(seq_names)
X_train = X_train[seq_names]
y_train = y_train[seq_names]
mask_train = mask_train[seq_names]
num_batches = data.num_seq_train // batch_size
losses = []
preds = []
norms = []
for i in range(num_batches):
idx = range(i*batch_size, (i+1)*batch_size)
x_batch = X_train[idx]
y_batch = y_train[idx]
mask_batch = mask_train[idx]
loss, out, batch_norm = train(x_batch, y_batch, mask_batch)
print(batch_norm)
norms.append(batch_norm)
preds.append(out)
losses.append(loss)
# if ((i+1) % config.write_every_batch == 0) | (i == 0):
# if i == 0:
# start_place = 0
# else:
# start_place = i-config.write_every_batch
# print "Batch %d of %d" % (i + 1, num_batches)
# print " curbatch training loss: %.5f" % np.mean(losses[start_place:(i+1)])
# print " curbatch training acc: %.5f" % np.mean(accuracy[start_place:(i+1)])
predictions = np.concatenate(preds, axis = 0)
loss_train = np.mean(losses)
all_losses_train.append(loss_train)
acc_train = utils.proteins_acc(predictions, y_train[0:num_batches*batch_size], mask_train[0:num_batches*batch_size])
all_accuracy_train.append(acc_train)
mean_norm = np.mean(norms)
all_mean_norm.append(mean_norm)
if 1==1:
print " average training loss: %.5f" % loss_train
print " average training accuracy: %.5f" % acc_train
print " average norm: %.5f" % mean_norm
if 1==1:#(i + 1) % config.validate_every == 0:
if config.batch_norm:
_ = f_collect(X_train)
sets = [#('train', X_train, y_train, mask_train, all_losses_eval_train, all_accuracy_eval_train),
('valid', X_valid, y_valid, mask_valid, all_losses_eval_valid, all_accuracy_eval_valid),
('test', X_test, y_test, mask_test, all_losses_eval_test, all_accuracy_eval_test)]
for subset, X, y, mask, all_losses, all_accuracy in sets:
print " validating: %s loss" % subset
preds = []
num_batches = np.size(X,axis=0) // config.batch_size
for i in range(num_batches): ## +1 to get the "rest"
print(i)
idx = range(i*batch_size, (i+1)*batch_size)
x_batch = X[idx]
y_batch = y[idx]
mask_batch = mask[idx]
loss, out = eval(x_batch, y_batch, mask_batch)
preds.append(out)
# acc = utils.proteins_acc(out, y_batch, mask_batch)
losses.append(loss)
# accuracy.append(acc)
predictions = np.concatenate(preds, axis = 0)
print " pred"
print(predictions.shape)
print(predictions.dtype)
loss_eval = np.mean(losses)
all_losses.append(loss_eval)
# acc_eval = np.mean(accuracy)
acc_eval = utils.proteins_acc(predictions, y, mask)
all_accuracy.append(acc_eval)
# print " average evaluation loss (%s): %.5f" % (subset, loss_eval)
print " average evaluation accuracy (%s): %.5f" % (subset, acc_eval)
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * num_epochs
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print " %s since start (%.2f s)" % (utils.hms(time_since_start), time_since_prev)
print " estimated %s to go (ETA: %s)" % (utils.hms(est_time_left), eta_str)
print
if (epoch >= config.start_saving_at) and ((epoch % config.save_every) == 0):
print " saving parameters and metadata"
with open((metadata_path + "-%d" % (epoch) + ".pkl"), 'w') as f:
pickle.dump({
'config_name': config_name,
'param_values': nn.layers.get_all_param_values(l_out),
'losses_train': all_losses_train,
'accuracy_train': all_accuracy_train,
'losses_eval_train': all_losses_eval_train,
'losses_eval_valid': all_losses_eval_valid,
'losses_eval_test': all_losses_eval_test,
'accuracy_eval_valid': all_accuracy_eval_valid,
'accuracy_eval_train': all_accuracy_eval_train,
'accuracy_eval_test': all_accuracy_eval_test,
'mean_norm' : all_mean_norm,
'time_since_start': time_since_start,
'i': i,
}, f, pickle.HIGHEST_PROTOCOL)
print " stored in %s" % metadata_path
print
predictions = []
batch_size = config.batch_size
num_batches = np.size(X,axis=0) // batch_size
for i in range(num_batches):
idx = range(i*batch_size, (i+1)*batch_size)
x_batch = X[idx]
mask_batch = mask[idx]
p = predict(x_batch, mask_batch)
predictions.append(p)
predictions = np.concatenate(predictions, axis = 0)
predictions_path = os.path.join("predictions", os.path.basename(metadata_path).replace("dump_", "predictions_").replace(".pkl", ".npy"))
print(utils.proteins_acc(predictions, data.labels_test, data.mask_test))
print "Storing predictions in %s" % predictions_path
np.save(predictions_path, predictions)