def train(args):
X_train = args["X_train"]
X_train = X_train.reshape((X_train.shape[0], 1, 28, 28))
X_train /= 255
y_train = np.asarray(args["y_train"].flatten(), dtype="int32")
num_epochs = 1 if "num_epochs" not in args else args["num_epochs"]
net = get_net(num_epochs)
net.fit(X_train, y_train)
if "vis" in args and args["vis"] == True:
helper.plot_conv_activity(net.layers_[1], X_train[0:1])
return net.get_all_params_values()
def train(args):
X_train = args["X_train"]
X_train = X_train.reshape( (X_train.shape[0], 1, 28, 28) )
X_train /= 255
y_train = np.asarray( args["y_train"].flatten(), dtype="int32" )
num_epochs = 1 if "num_epochs" not in args else args["num_epochs"]
net = get_net(num_epochs)
net.fit(X_train, y_train)
if "vis" in args and args["vis"] == True:
helper.plot_conv_activity( net.layers_[1], X_train[0:1] )
return net.get_all_params_values()
def train(arg):
global args, symbols, best_weights, SEED
args = arg
filenames = [ (args["dir"] + os.path.sep + args["attr_values"]["filename"][int(x)]) \
for x in args["X_train"].flatten().tolist() ]
symbols = prepare()
alpha = args["alpha"]
momentum = 0.9
if alpha != -1:
if "rmsprop" not in args:
updates = lasagne.updates.momentum(symbols.loss, symbols.all_params, alpha, momentum)
else:
updates = lasagne.updates.rmsprop(symbols.loss, symbols.all_params, alpha)
else:
updates = lasagne.updates.adagrad(symbols.loss, symbols.all_params, 1.0)
iter_train = theano.function(
[symbols.X],
[symbols.loss, symbols.loss_flat],
updates=updates
)
if "batch_size" in args:
bs = args["batch_size"]
else:
bs = 128
best_valid_accuracy = -1
for e in range(0, args["epochs"]):
np.random.seed(SEED)
np.random.shuffle(filenames)
SEED += 1
np.random.seed(SEED)
sys.stderr.write("Epoch #%i:\n" % e)
batch_train_losses = []
batch_train_accuracies = []
batch_train_alt_losses = []
for b in range(0, len(filenames)):
if b*bs >= len(filenames):
break
X_train_batch = get_batch(filenames, b*bs, (b+1)*bs)
#print (X_train_batch.shape)
#sys.stderr.write(" Batch #%i (%i-%i)\n" % ((b+1), (b*bs), ((b+1)*bs) ))
loss, loss_flat = iter_train( X_train_batch )
batch_train_losses.append(loss)
print (loss, loss_flat, sum(loss_flat > 0))
helper.plot_conv_activity( symbols.conv_layer, X_train_batch[1:2] )
sys.stderr.write( " train_loss = %f\n" % \
(np.mean(batch_train_losses)) )
current_weights = lasagne.layers.get_all_param_values(symbols.output_layer)
return (print_network(symbols.output_layer), current_weights)
def train(arg):
global args, symbols, best_weights, SEED
args = arg
filenames = [ (args["dir"] + os.path.sep + args["attr_values"]["filename"][int(x)]) \
for x in args["X_train"].flatten().tolist() ]
symbols = prepare()
alpha = args["alpha"]
momentum = 0.9
if alpha != -1:
if "rmsprop" not in args:
updates = lasagne.updates.momentum(symbols.loss,
symbols.all_params, alpha,
momentum)
else:
updates = lasagne.updates.rmsprop(symbols.loss, symbols.all_params,
alpha)
else:
updates = lasagne.updates.adagrad(symbols.loss, symbols.all_params,
1.0)
iter_train = theano.function([symbols.X],
[symbols.loss, symbols.loss_flat],
updates=updates)
if "batch_size" in args:
bs = args["batch_size"]
else:
bs = 128
best_valid_accuracy = -1
for e in range(0, args["epochs"]):
np.random.seed(SEED)
np.random.shuffle(filenames)
SEED += 1
np.random.seed(SEED)
sys.stderr.write("Epoch #%i:\n" % e)
batch_train_losses = []
batch_train_accuracies = []
batch_train_alt_losses = []
for b in range(0, len(filenames)):
if b * bs >= len(filenames):
break
X_train_batch = get_batch(filenames, b * bs, (b + 1) * bs)
#print (X_train_batch.shape)
#sys.stderr.write(" Batch #%i (%i-%i)\n" % ((b+1), (b*bs), ((b+1)*bs) ))
loss, loss_flat = iter_train(X_train_batch)
batch_train_losses.append(loss)
print(loss, loss_flat, sum(loss_flat > 0))
helper.plot_conv_activity(symbols.conv_layer, X_train_batch[1:2])
sys.stderr.write( " train_loss = %f\n" % \
(np.mean(batch_train_losses)) )
current_weights = lasagne.layers.get_all_param_values(symbols.output_layer)
return (print_network(symbols.output_layer), current_weights)
