def load_network(loadfile, multiview=None, sort_layers=False):
from convdata import DataProvider
from python_util.gpumodel import IGPUModel
load_dic = IGPUModel.load_checkpoint(loadfile)
layers = load_dic['model_state']['layers']
op = load_dic['op']
if sort_layers:
depths = get_depths(layers)
layers = collections.OrderedDict(
sorted(layers.items(), key=lambda item: depths[item[0]]))
options = {}
for o in load_dic['op'].get_options_list():
options[o.name] = o.value
dp_params = {}
for v in ('color_noise', 'multiview_test', 'inner_size', 'scalar_mean',
'minibatch_size'):
dp_params[v] = options[v]
lib_name = "cudaconvnet._ConvNet"
print("Importing %s C++ module" % lib_name)
libmodel = __import__(lib_name, fromlist=['_ConvNet'])
dp_params['libmodel'] = libmodel
if multiview is not None:
dp_params['multiview_test'] = multiview
dp = DataProvider.get_instance(options['data_path'],
batch_range=options['test_batch_range'],
type=options['dp_type'],
dp_params=dp_params,
test=True)
epoch, batchnum, data = dp.get_next_batch()
images, labels = data[:2]
images = images.T
images.shape = (images.shape[0], dp.num_colors, dp.inner_size,
dp.inner_size)
labels.shape = (-1, )
labels = labels.astype('int')
assert images.shape[0] == labels.shape[0]
if 1:
rng = np.random.RandomState(8)
i = rng.permutation(images.shape[0])
images = images[i]
labels = labels[i]
data = [images, labels] + list(data[2:])
# data['data_mean'] = dp.data_mean
# data['label_names'] = dp.batch_meta['label_names']
return layers, data, dp
def plain_pickle(loadfile, savefile):
load_dic = IGPUModel.load_checkpoint(loadfile)
options = {}
for o in load_dic['op'].get_options_list():
options[o.name] = o.value
load_dic['op'] = options
with open(savefile, 'wb') as f:
pickle.dump(load_dic, f, protocol=-1)
print("Wrote %r" % savefile)
def loadbcp(name, shape, params=None):
load_dic = IGPUModel.load_checkpoint(params[0])
rows, cols = shape
return load_dic['model_state']['layers'][params[1]]['biases'].reshape(rows,cols)
def loadwcp(name, idx, shape, params=None):
load_dic = IGPUModel.load_checkpoint(params[0])
rows, cols = shape
return load_dic['model_state']['layers'][params[1]]['weights'][idx].reshape(rows,cols)
def loadbcp(name, shape, params=None):
load_dic = IGPUModel.load_checkpoint(params[0])
rows, cols = shape
return load_dic['model_state']['layers'][params[1]]['biases'].reshape(
rows, cols)
def loadwcp(name, idx, shape, params=None):
load_dic = IGPUModel.load_checkpoint(params[0])
rows, cols = shape
return load_dic['model_state']['layers'][
params[1]]['weights'][idx].reshape(rows, cols)
if not donames:
load_convnets = sys.argv[1:] # just the paths
else:
load_convnets = sys.argv[1::2]
legend_names = sys.argv[2::2] # the paths and a shorter legend name
print 'Loading convnets ' + ', '.join(load_convnets)
for cost_idx in range(ncosts):
pl.figure(cost_idx + basefig)
if navg_groups > 1:
allte = None
alltr = None
for cv, i in zip(load_convnets, range(len(load_convnets))):
cp = IGPUModel.load_checkpoint(cv)
numbatches, train_errors, test_errors = get_errors(
cp,
show_cost,
cost_idx,
norm=cost_idx == 0,
smooth=smooth,
interp_test=dointerp)
#plot_errors(numbatches, train_errors, test_errors)
tr = np.array(train_errors, dtype=np.double)
te = np.array(test_errors, dtype=np.double)
if navg_groups > 1:
if allte is None:
allte = np.zeros(list(te.shape) + [len(load_convnets)],
dtype=np.double)
alltr = np.zeros(list(tr.shape) + [len(load_convnets)],
avg_groups = ['warp','nowarp']
navg_groups = len(avg_groups)
if not donames:
load_convnets = sys.argv[1:] # just the paths
else:
load_convnets = sys.argv[1::2]; legend_names = sys.argv[2::2] # the paths and a shorter legend name
print 'Loading convnets ' + ', '.join(load_convnets)
for cost_idx in range(ncosts):
pl.figure(cost_idx + basefig)
if navg_groups > 1: allte = None; alltr = None
for cv,i in zip(load_convnets, range(len(load_convnets))):
cp = IGPUModel.load_checkpoint(cv)
numbatches, train_errors, test_errors = get_errors(cp, show_cost,cost_idx, norm=cost_idx == 0,
smooth=smooth, interp_test=dointerp)
#plot_errors(numbatches, train_errors, test_errors)
tr = np.array(train_errors, dtype=np.double); te = np.array(test_errors, dtype=np.double)
if navg_groups > 1:
if allte is None:
allte = np.zeros(list(te.shape)+[len(load_convnets)], dtype=np.double)
alltr = np.zeros(list(tr.shape)+[len(load_convnets)], dtype=np.double)
allte[:,i] = te; alltr[:,i] = tr
else:
if dolog: tr = np.log10(tr); te = np.log10(te)
pl.plot(tr, te, label=os.path.basename(legend_names[i] if donames else os.path.normpath(cv)))
if navg_groups > 1:
groups = np.arange(len(load_convnets),dtype=np.int64)
groups = groups.reshape([navg_groups,len(load_convnets)//navg_groups])
