def create_blobs_if_not_existed(blob_names):
existd_names = set(workspace.Blobs())
for xx in blob_names:
if xx not in existd_names:
workspace.CreateBlob(str(xx))
init_params=False)
AddModel(deploy_model, "data")
# The parameter initialization network only needs to be run once.
# Now all the parameter blobs are going to be initialized in the workspace.
workspace.RunNetOnce(train_model.param_init_net)
# overwrite=True allows you to run this cell several times and avoid errors
workspace.CreateNet(train_model.net, overwrite=True)
# Set the iterations number and track the accuracy & loss
total_iters = 200
accuracy = np.zeros(total_iters)
loss = np.zeros(total_iters)
print("The blobs in the workspace pre-train: {}".format(workspace.Blobs()))
# Now, we will manually run the network for 200 iterations.
for i in range(total_iters):
workspace.RunNet(train_model.net)
accuracy[i] = workspace.blobs["accuracy"]
loss[i] = workspace.blobs["loss"]
print("The blobs in the workspace post-train: {}".format(workspace.Blobs()))
# param_init_net here will only create a data reader
# Other parameters won't be re-created because we selected
# init_params=False before
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net, overwrite=True)
test_accuracy = np.zeros(100)
def initialize_gpu_from_weights_file(model, weights_file, gpu_id=0):
"""Initialize a network with ops on a specific GPU.
If you use CUDA_VISIBLE_DEVICES to target specific GPUs, Caffe2 will
automatically map logical GPU ids (starting from 0) to the physical GPUs
specified in CUDA_VISIBLE_DEVICES.
"""
logger.info('Loading weights from: {}'.format(weights_file))
ws_blobs = workspace.Blobs()
src_blobs = load_object(weights_file)
if 'cfg' in src_blobs:
saved_cfg = load_cfg(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
# Backwards compat--dictionary used to be only blobs, now they are
# stored under the 'blobs' key
src_blobs = src_blobs['blobs']
# Initialize weights on GPU gpu_id only
unscoped_param_names = OrderedDict() # Print these out in model order
for blob in model.params:
unscoped_param_names[c2_utils.UnscopeName(str(blob))] = True
with c2_utils.NamedCudaScope(gpu_id):
for unscoped_param_name in unscoped_param_names.keys():
if (unscoped_param_name.find(']_') >= 0
and unscoped_param_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_param_name[unscoped_param_name.find(']_') +
2:]
else:
src_name = unscoped_param_name
if src_name not in src_blobs:
logger.info('{:s} not found'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_param_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum]' if has_momentum else ''
logger.info(
'{:s}{:} loaded from weights file into {:s}: {}'.format(
src_name, has_momentum_str, dst_name,
src_blobs[src_name].shape))
if dst_name in ws_blobs:
# If the blob is already in the workspace, make sure that it
# matches the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
assert ws_blob.shape == src_blobs[src_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
src_name,
ws_blob.shape,
src_blobs[src_name].shape)
workspace.FeedBlob(
dst_name, src_blobs[src_name].astype(np.float32, copy=False))
if has_momentum:
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(np.float32,
copy=False))
# We preserve blobs that are in the weights file but not used by the current
# model. We load these into CPU memory under the '__preserve__/' namescope.
# These blobs will be stored when saving a model to a weights file. This
# feature allows for alternating optimization of Faster R-CNN in which blobs
# unused by one step can still be preserved forward and used to initialize
# another step.
for src_name in src_blobs.keys():
if (src_name not in unscoped_param_names
and not src_name.endswith('_momentum')
and src_blobs[src_name] is not None):
with c2_utils.CpuScope():
workspace.FeedBlob('__preserve__/{:s}'.format(src_name),
src_blobs[src_name])
logger.info(
'{:s} preserved in workspace (unused)'.format(src_name))
def initialize_gpu_0_from_weights_file(model, weights_file):
logger.info('Loading from: {}'.format(weights_file))
is_first_init = 'trainedCOCO' in weights_file
ws_blobs = workspace.Blobs()
with open(weights_file, 'r') as f:
src_blobs = pickle.load(f)
if 'cfg' in src_blobs:
saved_cfg = yaml.load(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
# Backwards compat--dictionary used to be only blobs, now they are
# stored under the 'blobs' key
src_blobs = src_blobs['blobs']
# Initialize weights on GPU 0 only
unscoped_param_names = OrderedDict() # Print these out in model order
for blob in model.params:
unscoped_param_names[utils.blob.unscope_name(str(blob))] = True
with core.NameScope('gpu_0'):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
for unscoped_param_name in unscoped_param_names.keys():
if (unscoped_param_name.find(']_') >= 0 and
unscoped_param_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_param_name[
unscoped_param_name.find(']_') + 2:]
else:
src_name = unscoped_param_name
if src_name not in src_blobs:
logger.info('{:s} not found'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_param_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum]' if has_momentum else ''
logger.info('{:s}{:} loaded from weights file into {:s}: {}'.
format(
src_name, has_momentum_str,
dst_name, src_blobs[src_name].shape))
pretrained_w = src_blobs[src_name]
if dst_name in ws_blobs:
# If the blob is already in the workspace, make sure that it
# matches the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
if ws_blob.shape != src_blobs[src_name].shape:
pretrained_w = inflate_weights(
pretrained_w, ws_blob, src_name, src_blobs)
workspace.FeedBlob(
dst_name,
pretrained_w.astype(np.float32, copy=False))
if has_momentum and not is_first_init:
# when feeding momentum, we're probably resuming from
# previous checkpoint. So all the inflated stuff won't be
# needed in that case
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(
np.float32, copy=False))
# Add _rm/_riv BN mean/var params, in case the pre-trained model contains it.
# Needed to test the scratch trained models.
for src_name in src_blobs.keys():
if src_name.endswith('_rm') or src_name.endswith('_riv'):
with core.NameScope('gpu_0'):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
dst_name = core.ScopedName(src_name)
workspace.FeedBlob(dst_name, src_blobs[src_name])
logger.info('Loaded BN param {}'.format(src_name))
# We preserve blobs that are in the weights file but not used by the current
# model. We load these into CPU memory under the '__preserve__/' namescope.
# These blobs will be stored when saving a model to a weights file. This
# feature allows for alternating optimization of Faster R-CNN in which blobs
# unused by one step can still be preserved forward and used to initialize
# another step.
for src_name in src_blobs.keys():
if (src_name not in unscoped_param_names and
not src_name.endswith('_momentum') and
src_blobs[src_name] is not None):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU)):
workspace.FeedBlob(
'__preserve__/{:s}'.format(src_name), src_blobs[src_name])
logger.info(
'{:s} preserved in workspace (unused)'.format(src_name))
def run(self, index, inputFeatures, accumulate=True, image_path=None):
"""
index - index of the dataset entry
inputFeatures - features input to the head
accumulate - whether to save to predictions in self.all_... members
image_path - path to the annotated image, to which the predictions correspond
"""
timers = self.timers
# Format the inputs to the mask rcnn head
features = {}
for k, v in inputFeatures.iteritems():
assert v.dim() == 3, 'Batch mode not allowed'
features[k] = np.expand_dims(v.data.cpu().numpy(), axis=0)
gpu_dev = caffe2_core.DeviceOption(caffe2_pb2.CUDA, self.gpu_id)
name_scope = 'gpu_{}'.format(self.gpu_id)
# Clean the workspace to make damn sure that nothing comes from the
# possible forwarding of target features, depending on the use of this
# module
parameters = [str(s) for s in self.model.params] + [
str(s) + '_momentum' for s in self.model.TrainableParams()
]
for b in workspace.Blobs():
if not b in parameters:
workspace.FeedBlob(b, np.array([]))
# Produce the top level of the pyramid of features
with caffe2_core.NameScope(name_scope):
with caffe2_core.DeviceScope(gpu_dev):
workspace.FeedBlob(
caffe2_core.ScopedName("predicted_fpn_res5_2_sum"),
features['fpn_res5_2_sum'])
workspace.RunOperatorOnce(self.subsampler)
features[
u'fpn_res5_2_sum_subsampled_2x'] = workspace.FetchBlob(
caffe2_core.ScopedName(
"predicted_fpn_res5_2_sum_subsampled_2x"))
# Forward the rest of the features in the head of the model
im_info = np.array([[1024., 2048., 1.]], dtype=np.float32)
im_scales = np.array([1.])
im_shape = (1024, 2048, 3)
with caffe2_core.NameScope(name_scope):
with caffe2_core.DeviceScope(gpu_dev):
cls_boxes_i, cls_segms_i = im_detect_all_given_features(
self.model, self.subsampler, features, im_info, im_scales,
im_shape, timers)
# If required, store the results in the class's members
if accumulate:
extend_results(index, self.all_boxes_ann_frame, cls_boxes_i)
if cls_segms_i is not None and accumulate:
extend_results(index, self.all_segms_ann_frame, cls_segms_i)
if image_path is not None and accumulate:
self.id_sequences.append(image_path)
if index % 10 == 0:
ave_total_time = np.sum([t.average_time for t in timers.values()])
det_time = (timers['im_detect_bbox'].average_time +
timers['im_detect_mask'].average_time)
misc_time = (timers['misc_bbox'].average_time +
timers['misc_mask'].average_time)
print(('im_detect: '
'{:d}/{:d} {:.3f}s + {:.3f}s => avg total time: {:.3f}s'
).format(index, self.num_images, det_time, misc_time,
ave_total_time))
return cls_boxes_i, cls_segms_i
def testBlobNameOverrides(self):
original_names = ['blob_a', 'blob_b', 'blob_c']
new_names = ['x', 'y', 'z']
blobs = [np.random.permutation(6) for i in range(3)]
for i, blob in enumerate(blobs):
self.assertTrue(workspace.FeedBlob(original_names[i], blob))
self.assertTrue(workspace.HasBlob(original_names[i]))
self.assertEqual(len(workspace.Blobs()), 3)
try:
# Saves the blobs to a local db.
tmp_folder = tempfile.mkdtemp()
with self.assertRaises(RuntimeError):
workspace.RunOperatorOnce(
core.CreateOperator("Save",
original_names, [],
absolute_path=1,
strip_prefix='.temp',
blob_name_overrides=new_names,
db=os.path.join(tmp_folder, "db"),
db_type=self._db_type))
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator("Save",
original_names, [],
absolute_path=1,
blob_name_overrides=new_names,
db=os.path.join(tmp_folder, "db"),
db_type=self._db_type)))
self.assertTrue(workspace.ResetWorkspace())
self.assertEqual(len(workspace.Blobs()), 0)
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator("Load", [], [],
absolute_path=1,
db=os.path.join(tmp_folder, "db"),
db_type=self._db_type,
load_all=1)))
self.assertEqual(len(workspace.Blobs()), 3)
for i, name in enumerate(new_names):
self.assertTrue(workspace.HasBlob(name))
self.assertTrue((workspace.FetchBlob(name) == blobs[i]).all())
# moved here per @cxj's suggestion
load_new_names = ['blob_x', 'blob_y', 'blob_z']
# load 'x' into 'blob_x'
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator("Load", [],
load_new_names[0:1],
absolute_path=1,
db=os.path.join(tmp_folder, "db"),
db_type=self._db_type,
source_blob_names=new_names[0:1])))
# we should have 'blob_a/b/c/' and 'blob_x' now
self.assertEqual(len(workspace.Blobs()), 4)
for i, name in enumerate(load_new_names[0:1]):
self.assertTrue(workspace.HasBlob(name))
self.assertTrue((workspace.FetchBlob(name) == blobs[i]).all())
self.assertTrue(
workspace.RunOperatorOnce(
core.CreateOperator("Load", [],
load_new_names[0:3],
absolute_path=1,
db=os.path.join(tmp_folder, "db"),
db_type=self._db_type,
source_blob_names=new_names[0:3])))
# we should have 'blob_a/b/c/' and 'blob_x/y/z' now
self.assertEqual(len(workspace.Blobs()), 6)
for i, name in enumerate(load_new_names[0:3]):
self.assertTrue(workspace.HasBlob(name))
self.assertTrue((workspace.FetchBlob(name) == blobs[i]).all())
finally:
# clean up temp folder.
try:
shutil.rmtree(tmp_folder)
except OSError as e:
if e.errno != errno.ENOENT:
raise
print("The deploy model is saved to: " + root_folder + "/mnist_model.minidb")
# Now we can load the model back and run the prediction to verify it works.
# In[18]:
# we retrieve the last input data out and use it in our prediction test before we scratch the workspace
blob = workspace.FetchBlob("data")
pyplot.figure()
_ = visualize.NCHW.ShowMultiple(blob)
# reset the workspace, to make sure the model is actually loaded
workspace.ResetWorkspace(root_folder)
# verify that all blobs are destroyed.
print("The blobs in the workspace after reset: {}".format(workspace.Blobs()))
# load the predict net
predict_net = pe.prepare_prediction_net(
os.path.join(root_folder, "mnist_model.minidb"), "minidb")
# verify that blobs are loaded back
print("The blobs in the workspace after loading the model: {}".format(
workspace.Blobs()))
# feed the previously saved data to the loaded model
workspace.FeedBlob("data", blob)
# predict
workspace.RunNetOnce(predict_net)
softmax = workspace.FetchBlob("softmax")
final_image = transposed_image
print("Shape of final_image: " + str(np.array(final_image).shape))
with open(MODEL_ROOT + "/init_net.pb", "rb") as f:
init_net = f.read()
with open(MODEL_ROOT + "/predict_net.pb", "rb") as f:
predict_net = f.read()
workspace.ResetWorkspace()
blob_name = model_props[MODEL].blob_name
workspace.FeedBlob(blob_name, final_image)
print("The blobs in the workspace after FeedBlob: {}".format(workspace.Blobs()))
# Create a predictor using the loaded model.
p = workspace.Predictor(init_net, predict_net)
start = time.time()
for i in range(0, args.iterations):
results = p.run([final_image])
end = time.time()
if args.time:
print('Wall time per iteration (s): {:0.4f}'.format(
(end - start) / args.iterations))
max_idx = np.argmax(results[0][0])
sum_probability = sum(results[0][0])
def initialize_gpu_from_weights_file(model, weights_file, gpu_id=0):
"""Initialize a network with ops on a specific GPU.
If you use CUDA_VISIBLE_DEVICES to target specific GPUs, Caffe2 will
automatically map logical GPU ids (starting from 0) to the physical GPUs
specified in CUDA_VISIBLE_DEVICES.
"""
logger.info('Loading weights from: {}'.format(weights_file))
ws_blobs = workspace.Blobs()
with open(weights_file, 'r') as f:
src_blobs = pickle.load(f)
if 'cfg' in src_blobs:
saved_cfg = load_cfg(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
# Backwards compat--dictionary used to be only blobs, now they are
# stored under the 'blobs' key
src_blobs = src_blobs['blobs']
# Initialize weights on GPU gpu_id only
unscoped_param_names = OrderedDict() # Print these out in model order
for blob in model.params:
unscoped_param_names[c2_utils.UnscopeName(str(blob))] = True
with c2_utils.NamedCudaScope(gpu_id):
for unscoped_param_name in unscoped_param_names.keys():
if (unscoped_param_name.find(']_') >= 0
and unscoped_param_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_param_name[unscoped_param_name.find(']_') +
2:]
else:
src_name = unscoped_param_name
if src_name not in src_blobs:
logger.info('{:s} not found'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_param_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum]' if has_momentum else ''
logger.debug(
'{:s}{:} loaded from weights file into {:s}: {}'.format(
src_name, has_momentum_str, dst_name,
src_blobs[src_name].shape))
if dst_name in ws_blobs:
print("dst_name:" + dst_name)
# If the blob is already in the workspace, make sure that it
# matches the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
print("xhpan:ws_blob.shape:" + str(ws_blob.shape))
print("xhpan:src_blobs[src_name].shape:" +
str(src_blobs[src_name].shape))
classes_layers_list_w = [
'gpu_0/cls_score_w', 'gpu_0/bbox_pred_w',
'gpu_0/mask_fcn_logits_w'
]
classes_layers_list_b = [
'gpu_0/cls_score_b', 'gpu_0/bbox_pred_b',
'gpu_0/mask_fcn_logits_b'
]
# -----------------(11, 1024) - --------------(10, 1024)
# -----------------(11,) - --------------(10,)
# -----------------(44, 1024) - --------------(40, 1024)
# -----------------(44,) - --------------(40,)
# -----------------(11, 256, 1, 1) - --------------(10, 256, 1, 1)
# -----------------(11,) - --------------(10,)
# if ws_blob.shape != src_blobs[src_name].shape:
# if dst_name is 'gpu_0/cls_score_w':
# if ws_blob.shape[0] > src_blobs[src_name].shape[0]:#(10, 1024)
# src_blobs[src_name].extend(0.0001 * np.random.randn(*(ws_blob.shape[0] - src_blobs[src_name].shape[0], ws_blob.shape[1])))
# else:
# num = src_blobs[src_name].shape[0] - ws_blob.shape[0]
# src_blobs[src_name] = src_blobs[src_name][-num]
# elif dst_name is
cfg.MODEL.NUM_CLASSES
if ws_blob.shape != src_blobs[src_name].shape:
print("ws_blob.shape != src_blobs[src_name].shape")
print("-----------------" + str(ws_blob.shape) +
"---------------" + str(src_blobs[src_name].shape))
if dst_name in classes_layers_list_w or dst_name in classes_layers_list_b:
if dst_name in classes_layers_list_w:
target_shape = [
ws_blob.shape[0] - src_blobs[src_name].shape[0]
]
target_shape.extend(list(ws_blob.shape[1:]))
init_weight = 0.0001 * np.random.randn(
*(tuple(target_shape)))
src_blobs[src_name] = np.append(
src_blobs[src_name], init_weight, axis=0)
else:
target_shape = [
ws_blob.shape[0] - src_blobs[src_name].shape[0]
]
target_shape.extend(list(ws_blob.shape[1:]))
init_weight = -np.log(
(1 - 0.00001) / 0.00001) * np.ones(
*(tuple(target_shape)))
src_blobs[src_name] = np.append(
src_blobs[src_name], init_weight, axis=0)
target_shape = [
ws_blob.shape[0] -
src_blobs[src_name + '_momentum'].shape[0]
]
target_shape.extend(list(ws_blob.shape[1:]))
init_weight = np.zeros(target_shape)
src_blobs[src_name + '_momentum'] = np.append(
src_blobs[src_name + '_momentum'],
init_weight,
axis=0)
# if ws_blob.shape != src_blobs[src_name].shape:
# print ("ws_blob.shape != src_blobs[src_name].shape")
# print ("-----------------" + str(ws_blob.shape) + "---------------" + str(src_blobs[src_name].shape))
# if dst_name in classes_layers_list_w or dst_name in classes_layers_list_b:
# if dst_name in classes_layers_list_w :
# src_blobs[src_name] = 0.0001 * np.random.randn(*(ws_blob.shape))
# else:
# src_blobs[src_name] = -np.log((1 - 0.00001) / 0.00001) * np.ones(*(ws_blob.shape))
#
# src_blobs[src_name + '_momentum'] = np.zeros(ws_blob.shape)
assert ws_blob.shape == src_blobs[src_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
src_name,
ws_blob.shape,
src_blobs[src_name].shape)
workspace.FeedBlob(
dst_name, src_blobs[src_name].astype(np.float32, copy=False))
if has_momentum:
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(np.float32,
copy=False))
# We preserve blobs that are in the weights file but not used by the current
# model. We load these into CPU memory under the '__preserve__/' namescope.
# These blobs will be stored when saving a model to a weights file. This
# feature allows for alternating optimization of Faster R-CNN in which blobs
# unused by one step can still be preserved forward and used to initialize
# another step.
for src_name in src_blobs.keys():
if (src_name not in unscoped_param_names
and not src_name.endswith('_momentum')
and src_blobs[src_name] is not None):
with c2_utils.CpuScope():
workspace.FeedBlob('__preserve__/{:s}'.format(src_name),
src_blobs[src_name])
logger.debug(
'{:s} preserved in workspace (unused)'.format(src_name))
def print_all(self):
# approach 1: all
print(workspace.Blobs(), end='\n')
for _, l in enumerate(workspace.Blobs()):
print(l)
print(self.FetchBlobWrapper(l))
def initialize_master_gpu_model_params(
model, weights_file, load_momentum=True):
ws_blobs = workspace.Blobs()
logger.info("Initializing model params from file: {}".format(weights_file))
with open(weights_file, 'r') as fopen:
blobs = pickle.load(fopen)
if 'blobs' in blobs:
blobs = blobs['blobs']
unscoped_blob_names = OrderedDict()
# Return the model iter from which training should start
model_iter = 0
if 'model_iter' in blobs:
model_iter = blobs['model_iter']
if 'lr' in blobs:
prev_lr = float(blobs['lr'])
elif cfg.TRAIN.RESET_START_ITER:
prev_lr = 1.
else:
raise Exception('No lr blob found.')
# initialize params, params momentum, computed params
if 'test' not in model.net.Name() and load_momentum:
for param in model.params:
if param in model.TrainableParams():
unscoped_blob_names[misc.unscope_name(
str(param) + '_momentum')] = True
for blob in model.GetAllParams():
unscoped_blob_names[misc.unscope_name(str(blob))] = True
root_gpu_id = cfg.ROOT_GPU_ID
with core.NameScope('gpu_{}'.format(root_gpu_id)):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, root_gpu_id)):
for unscoped_blob_name in unscoped_blob_names.keys():
scoped_blob_name = misc.scoped_name(unscoped_blob_name)
if unscoped_blob_name not in blobs:
logger.info('{:s} not found'.format(unscoped_blob_name))
continue
if scoped_blob_name in ws_blobs:
ws_blob = workspace.FetchBlob(scoped_blob_name)
if 'pred' in unscoped_blob_name:
if np.prod(ws_blob.shape) \
!= np.prod(blobs[unscoped_blob_name].shape):
logger.info(('{:s} (classifier) found but ' +
'unmatching (not loaded):' +
'{} ---> {}')
.format(
unscoped_blob_name,
blobs[unscoped_blob_name].shape,
ws_blob.shape))
continue
else:
blobs[unscoped_blob_name] = np.reshape(
blobs[unscoped_blob_name], ws_blob.shape)
if len(ws_blob.shape) != \
len(blobs[unscoped_blob_name].shape):
# inflate if so
assert ws_blob.shape[:2] == \
blobs[unscoped_blob_name].shape[:2], \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
unscoped_blob_name, ws_blob.shape,
blobs[unscoped_blob_name].shape)
assert ws_blob.shape[-2:] == \
blobs[unscoped_blob_name].shape[-2:], \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
unscoped_blob_name, ws_blob.shape,
blobs[unscoped_blob_name].shape)
logger.info(
('{:s} loaded from weights file into: {:s}' +
' inflated {} ---> {}').format(
unscoped_blob_name, scoped_blob_name,
blobs[unscoped_blob_name].shape,
ws_blob.shape))
# inflate
num_inflate = ws_blob.shape[2]
blobs[unscoped_blob_name] = np.stack(
[blobs[unscoped_blob_name]] * num_inflate,
axis=2) / float(num_inflate)
else:
logger.info(
('{:s} loaded from weights file into: {:s}' +
' {}').format(
unscoped_blob_name, scoped_blob_name,
ws_blob.shape))
assert ws_blob.shape == blobs[unscoped_blob_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
unscoped_blob_name, ws_blob.shape,
blobs[unscoped_blob_name].shape)
data = blobs[unscoped_blob_name].astype(np.float32, copy=False)
workspace.FeedBlob(scoped_blob_name, data)
# hack fix: load and broadcast lr to all gpus
for i in range(cfg.NUM_GPUS):
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, i)):
workspace.FeedBlob(
'gpu_{}/lr'.format(i), np.array(prev_lr, dtype=np.float32))
return model_iter, prev_lr
def initialize_master_xpu_model_params(model, weights_file, opts, reset_epoch):
log.info("Initializing model params from file: {}".format(weights_file))
with open(weights_file, 'r') as fopen:
blobs = pickle.load(fopen)
if 'blobs' in blobs:
blobs = blobs['blobs']
start_epoch = 0
best_metric = float('-inf')
if 'epoch' in blobs:
log.info('epoch {} is found in model file'.format(blobs['epoch']))
if not reset_epoch:
start_epoch = blobs['epoch']
else:
log.info('Reset epoch')
else:
log.info('no epoch is found in model file')
lr = opts['model_param']['base_learning_rate']
if 'lr' in blobs:
lr = blobs['lr']
if 'best_metric' in blobs and not reset_epoch:
best_metric = blobs['best_metric']
if model is not None:
log.info('initialize model parameters using weights file: {}'.format(
weights_file
))
ws_blobs = workspace.Blobs()
unscoped_blob_names = OrderedDict()
for blob in model.GetAllParams():
unscoped_blob_names[unscope_name(str(blob))] = True
root_xpu_id = opts['distributed']['first_xpu_id']
device = opts['distributed']['device']
caffe2_pb2_DEVICE =\
caffe2_pb2.CUDA if opts['distributed']['device'] == 'gpu'\
else caffe2_pb2.CPU
with core.NameScope('{}_{}'.format(device, root_xpu_id)):
with core.DeviceScope(core.DeviceOption(caffe2_pb2_DEVICE, 0)):
for unscoped_blob_name in unscoped_blob_names.keys():
scoped_blob_name = scoped_name(unscoped_blob_name)
if unscoped_blob_name not in blobs:
log.info('{:s} not found'.format(unscoped_blob_name))
continue
log.info(
'{:s} loaded from weights file into: {:s}'.format(
unscoped_blob_name, scoped_blob_name
)
)
if scoped_blob_name in ws_blobs:
ws_blob = workspace.FetchBlob(scoped_blob_name)
if not ws_blob.shape == blobs[unscoped_blob_name].shape:
log.info(
('Workspace blob {} with shape {} does '
'not match weights file shape {}').format(
unscoped_blob_name, ws_blob.shape,
blobs[unscoped_blob_name].shape)
)
else:
workspace.FeedBlob(
scoped_blob_name,
blobs[unscoped_blob_name].astype(
np.float32, copy=False))
else:
log.info('Skip initializing model parameters from file: {}'.format(
weights_file
))
log.info('Complete initialize_master_xpu_model_params')
return start_epoch, lr, best_metric
print("Model saved as " + full_init_net_out + " and " +
full_predict_net_out)
# [end-20181001-ben-add] #
#
# load model
#
blob = workspace.FetchBlob("data")
plt.figure("Load data")
plt.title("Batch of Testing Data")
_ = visualize.NCHW.ShowMultiple(blob)
# reset the workspace, to make sure the model is actually loaded
print("The blobs in the workspace before reset: {}".format(
workspace.Blobs()))
workspace.ResetWorkspace(root_folder)
print("The blobs in the workspace after reset: {}".format(
workspace.Blobs())) # all blobs are destroyed
# load the predict net and verify the blobs
predict_net = pe.prepare_prediction_net(
os.path.join(root_folder, "mnist_model.minidb"), "minidb")
print("The blobs in the workspace after loading the model: {}".format(
workspace.Blobs()))
# feed the previously saved data to the loaded model
workspace.FeedBlob("data", blob)
workspace.RunNetOnce(predict_net)
softmax = workspace.FetchBlob("softmax")
def main():
# Initialize C2
workspace.GlobalInit(
['caffe2', '--caffe2_log_level=0', '--caffe2_gpu_memory_tracking=1']
)
# Set up logging and load config options
logger = setup_logging(__name__)
logging.getLogger('detectron.roi_data.loader').setLevel(logging.INFO)
args = parse_args()
logger.info('Called with args:')
logger.info(args)
if args.cfg_file is not None:
merge_cfg_from_file(args.cfg_file)
if args.opts is not None:
merge_cfg_from_list(args.opts)
assert_and_infer_cfg()
smi_output, cuda_ver, cudnn_ver = c2_utils.get_nvidia_info()
logger.info("cuda version : {}".format(cuda_ver))
logger.info("cudnn version: {}".format(cudnn_ver))
logger.info("nvidia-smi output:\n{}".format(smi_output))
logger.info('Training with config:')
logger.info(pprint.pformat(cfg))
# Note that while we set the numpy random seed network training will not be
# deterministic in general. There are sources of non-determinism that cannot
# be removed with a reasonble execution-speed tradeoff (such as certain
# non-deterministic cudnn functions).
np.random.seed(cfg.RNG_SEED)
# test model
logger.info("creat test model ...")
test_model = test_engine.initialize_model_from_cfg(cfg.TEST.WEIGHTS, gpu_id=0)
logger.info("created test model ...")
#cfg.TRAIN.IMS_PER_BATCH = 1
train_data = DataLoader(root, "val_id.txt", cfg, test_model, is_train=False)
# creat mode
model, weights_file, start_iter, checkpoints = create_model(False, cfg, output_dir)
# test blob
print(workspace.Blobs())
# create input blob
blob_names = ['data_stage2']
for gpu_id in range(cfg.NUM_GPUS):
with c2_utils.NamedCudaScope(gpu_id):
for blob_name in blob_names:
workspace.CreateBlob(core.ScopedName(blob_name))
# Override random weight initialization with weights from a saved model
if weights_file:
nu.initialize_gpu_from_weights_file(model, weights_file, gpu_id=0)
# Even if we're randomly initializing we still need to synchronize
# parameters across GPUs
nu.broadcast_parameters(model)
workspace.CreateNet(model.net)
logger.info('Outputs saved to: {:s}'.format(os.path.abspath(output_dir)))
logger.info("start test ...")
save_root = os.path.join(output_dir, 'fusion')
if not os.path.exists(save_root):
os.makedirs(save_root)
for cur_iter in range(10000):
# feed data
# print("{} iter starting feed data...".format(cur_iter))
data_stage2, gt_label, meta = train_data.next_batch()
'''#
print('input0-20 sungalsses max score:', np.max(data_stage2[0, 4, :, :]))
print('input20-40 sungalsses max score:', np.max(data_stage2[0, 24, :, :]))
print('input0-20 glovess max score:', np.max(data_stage2[0, 3, :, :]))
print('input20-40 glovess max score:', np.max(data_stage2[0, 23, :, :]))
#'''
with c2_utils.NamedCudaScope(gpu_id):
workspace.FeedBlob(core.ScopedName('data_stage2'), data_stage2)
# print("workspace.RunNet(model.net.Proto().name)")
with c2_utils.NamedCudaScope(gpu_id):
workspace.RunNet(model.net.Proto().name)
batch_probs = workspace.FetchBlob(core.ScopedName('probs_human_NCHW_stage2'))
batch_probs = batch_probs.transpose((0, 2, 3, 1))
assert len(meta) == batch_probs.shape[0]
#print('batch_probs shape:', batch_probs.shape)
for i in range(len(meta)):
probs = cv2.resize(batch_probs[i], (meta[i]['width'], meta[i]['height']), interpolation=cv2.INTER_LINEAR)
probs = probs.transpose((2,0,1))
print('sungalsses max score:', np.max(probs[4, :, :]))
print('glovess max score:', np.max(probs[3, :, :]))
#print('probs shape:', probs.shape)
cv2.imwrite(os.path.join(save_root, meta[i]['id']+'.png'), probs.argmax(0))
print("prossed ", cur_iter)
def LoadModelFromPickleFile(
model,
pkl_file,
use_gpu=True,
root_gpu_id=0,
bgr2rgb=False,
inflating=True,
collapsing=True,
center_init=False,
):
ws_blobs = workspace.Blobs()
with open(pkl_file, 'r') as fopen:
blobs = pickle.load(fopen)
if 'blobs' in blobs:
blobs = blobs['blobs']
unscoped_blob_names = OrderedDict()
for blob in model.GetAllParams():
unscoped_blob_names[unscope_name(str(blob))] = True
if use_gpu:
device_opt = caffe2_pb2.CUDA
else:
device_opt = caffe2_pb2.CPU
with core.NameScope('gpu_{}'.format(root_gpu_id)):
with core.DeviceScope(core.DeviceOption(device_opt, root_gpu_id)):
for unscoped_blob_name in unscoped_blob_names.keys():
scoped_blob_name = scoped_name(unscoped_blob_name)
if unscoped_blob_name not in blobs:
log.info('{} not found'.format(unscoped_blob_name))
continue
if scoped_blob_name in ws_blobs:
ws_blob = workspace.FetchBlob(scoped_blob_name)
target_shape = ws_blob.shape
if target_shape == blobs[unscoped_blob_name].shape:
log.info('copying {}'.format(unscoped_blob_name))
if bgr2rgb and unscoped_blob_name == 'conv1_w':
feeding_blob = FlipBGR2RGB(
blobs[unscoped_blob_name]
)
else:
feeding_blob = blobs[unscoped_blob_name]
elif ws_blob.ndim == 5:
# inflate from FC to 1x1x1 conv
if blobs[unscoped_blob_name].ndim == 2:
log.info('convolutionalize {}'.format(
unscoped_blob_name)
)
feeding_blob = blobs[unscoped_blob_name]
feeding_blob = np.reshape(
feeding_blob,
feeding_blob.shape + (1, 1, 1)
)
else:
# may need to inflate
if not inflating:
log.info(
'{} found, but inflating is ignored'.format(
unscoped_blob_name
)
)
continue
feeding_blob = InflateBlob(
blobs[unscoped_blob_name],
target_shape,
unscoped_blob_name,
(0 if center_init else 1)
)
elif ws_blob.ndim == 4:
# may need to collapse
if not collapsing:
log.info(
'{} found, but collapsing is ignored'.format(
unscoped_blob_name
)
)
continue
feeding_blob = CollapseBlob(
blobs[unscoped_blob_name],
target_shape,
unscoped_blob_name
)
# either copy, inflate, or collapse blob
workspace.FeedBlob(
scoped_blob_name,
feeding_blob.astype(np.float32, copy=False)
)
nesterov=1,
policy='poly',
power=1.,
max_iter=MAX_ITER,
)
# initialization
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
print("hello foo")
# ================= DEBUG PRINT =======================
# print(model.net.Proto())
# print(model.param_init_net.Proto())
# i = 0
# for param in model.param_to_grad:
# print("{} : {}".format(i , param))
# i += 1
# i = 1
# for param in model.params:
# print("{} : {}".format(i , param))
# i += 1
i = 1
for blob in workspace.Blobs():
print("{} : {}".format(i, blob))
i += 1
# Let's show all plots inline.
# You might see a warning saying that caffe2 does not have GPU support. That means you are running a CPU-only build. Don't be alarmed - anything CPU is still runnable without problem.
# ## Workspaces
#
# Let's cover workspaces first, where all the data reside.
#
# If you are familiar with Matlab, workspace consists of blobs you create and store in memory. For now, consider a blob to be a N-dimensional Tensor similar to numpy's ndarray, but is contiguous. Down the road, we will show you that a blob is actually a typed pointer that can store any type of C++ objects, but Tensor is the most common type stored in a blob. Let's show what the interface looks like.
#
# `Blobs()` prints out all existing blobs in the workspace.
# `HasBlob()` queries if a blob exists in the workspace. For now, we don't have anything yet.
# In[ ]:
print("Current blobs in the workspace: {}".format(workspace.Blobs()))
print("Workspace has blob 'X'? {}".format(workspace.HasBlob("X")))
# We can feed blobs into the workspace using `FeedBlob()`.
# In[3]:
X = np.random.randn(2, 3).astype(np.float32)
print("Generated X from numpy:\n{}".format(X))
workspace.FeedBlob("X", X)
# Now, let's take a look what blobs there are in the workspace.
# In[4]:
print("Current blobs in the workspace: {}".format(workspace.Blobs()))
merge_cfg_from_file(cfg_file)
cfg.TRAIN.WEIGHTS = '' # NOTE: do not download pretrained model weights
cfg.TEST.WEIGHTS = weights_file
cfg.NUM_GPUS = 1
assert_and_infer_cfg()
#according the cfg to bulid model
model = initialize_model_from_cfg(weights_file)
return model
if __name__ == '__main__':
workspace.GlobalInit(['caffe2', '--caffe2_log_level=0'])
args = parse_args()
model = get_model(args.cfg, args.wts)
img = cv2.imread(args.img)
#im_scale = im_conv_body_only(model,img,cfg.TEST.SCALE, cfg.TEST.MAX_SIZE)
im_blob, im_scale, _im_info = blob_utils.get_image_blob(
img, cfg.TEST.SCALE, cfg.TEST.MAX_SIZE)
with c2_utils.NamedCudaScope(0):
# workspace.FeedBlob(core.ScopedName('data'), im_blob)
# workspace.RunNet(model.net.Proto().name)
# blob = workplace.FetchBlob('rois')
# print 1
cls_b, _, _ = infer_engine.im_detect_all(model, img, None)
blobs = workspace.Blobs()
print blobs
mask_logits = workspace.FetchBlob(core.ScopedName('mask_logits'))
#print mask_logits
print mask_logits.shape
np.save('/data1/shuai/adas/code/mask_logits.npy', mask_logits)
def _compare(self, model, forward_only):
# Store list of blobs that exist in the beginning
workspace.RunNetOnce(model.param_init_net)
init_ws = {k: workspace.FetchBlob(k) for k in workspace.Blobs()}
# Run with executor
for enable_executor in [0, 1]:
self.enable_rnn_executor(model.net, enable_executor, forward_only)
workspace.ResetWorkspace()
# Reset original state
for k, v in init_ws.items():
workspace.FeedBlob(k, v)
np.random.seed(10022015)
ws = {}
for j in range(len(self.Tseq)):
input_shape = [self.Tseq[j], self.batch_size, self.input_dim]
workspace.FeedBlob(
"input",
np.random.rand(*input_shape).astype(np.float32))
workspace.FeedBlob(
"target",
np.random.rand(self.Tseq[j], self.batch_size,
self.hidden_dim).astype(np.float32))
if j == 0:
workspace.CreateNet(model.net, overwrite=True)
workspace.RunNet(model.net.Proto().name)
# Store results for each iteration
for k in workspace.Blobs():
ws[k + "." + str(j)] = workspace.FetchBlob(k)
if enable_executor:
rnn_exec_ws = ws
else:
non_exec_ws = ws
# Test that all blobs are equal after running with executor
# or without.
self.assertEqual(list(non_exec_ws.keys()), list(rnn_exec_ws.keys()))
mismatch = False
for k in rnn_exec_ws.keys():
non_exec_v = non_exec_ws[k]
rnn_exec_v = rnn_exec_ws[k]
if type(non_exec_v) is np.ndarray:
if not np.allclose(non_exec_v, rnn_exec_v):
print("Mismatch: {}".format(k))
nv = non_exec_v.flatten()
rv = rnn_exec_v.flatten()
c = 0
for j in range(len(nv)):
if rv[j] != nv[j]:
print(j, rv[j], nv[j])
c += 1
if c == 10:
break
mismatch = True
self.assertFalse(mismatch)
workspace.RunNetOnce(c2_net.param_init_net)
workspace.CreateNet(c2_net.net)
# load pretrained weights
wts = pickle.load(open('pretrained/i3d_baseline_32x2_IN_pretrain_400k.pkl',
'rb'),
encoding='latin')['blobs']
for key in wts:
if type(wts[key]) == np.ndarray:
workspace.FeedBlob(key, wts[key])
workspace.FeedBlob('data', data)
workspace.RunNet(c2_net.net.Proto().name)
c2_blobs = {key: workspace.FetchBlob(key) for key in workspace.Blobs()}
#-----------------------------------------------------------------------------------------------#
torch.backends.cudnn.enabled = False
from models import resnet
data = torch.from_numpy(data).cuda()
pth_net = resnet.i3_res50(num_classes=400).cuda().eval()
def hook(module, input, output):
setattr(module, "_value_hook", output)
for name, module in pth_net.named_modules():
module.register_forward_hook(hook)
def test_lstm_extract_predictor_net(self):
model = ModelHelper(name="lstm_extract_test")
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
output, _, _, _ = rnn_cell.LSTM(
model=model,
input_blob="input",
seq_lengths="seqlengths",
initial_states=("hidden_init", "cell_init"),
dim_in=20,
dim_out=40,
scope="test",
drop_states=True,
return_last_layer_only=True,
)
# Run param init net to get the shapes for all inputs
shapes = {}
workspace.RunNetOnce(model.param_init_net)
for b in workspace.Blobs():
shapes[b] = workspace.FetchBlob(b).shape
# But export in CPU
(predict_net, export_blobs) = ExtractPredictorNet(
net_proto=model.net.Proto(),
input_blobs=["input"],
output_blobs=[output],
device=core.DeviceOption(caffe2_pb2.CPU, 1),
)
# Create the net and run once to see it is valid
# Populate external inputs with correctly shaped random input
# and also ensure that the export_blobs was constructed correctly.
workspace.ResetWorkspace()
shapes['input'] = [10, 4, 20]
shapes['cell_init'] = [1, 4, 40]
shapes['hidden_init'] = [1, 4, 40]
print(predict_net.Proto().external_input)
self.assertTrue('seqlengths' in predict_net.Proto().external_input)
for einp in predict_net.Proto().external_input:
if einp == 'seqlengths':
workspace.FeedBlob("seqlengths",
np.array([10] * 4, dtype=np.int32))
else:
workspace.FeedBlob(
einp,
np.zeros(shapes[einp]).astype(np.float32),
)
if einp != 'input':
self.assertTrue(einp in export_blobs)
print(str(predict_net.Proto()))
self.assertTrue(workspace.CreateNet(predict_net.Proto()))
self.assertTrue(workspace.RunNet(predict_net.Proto().name))
# Validate device options set correctly for the RNNs
import google.protobuf.text_format as protobuftx
for op in predict_net.Proto().op:
if op.type == 'RecurrentNetwork':
for arg in op.arg:
if arg.name == "step_net":
step_proto = caffe2_pb2.NetDef()
protobuftx.Merge(arg.s, step_proto)
for step_op in step_proto.op:
self.assertEqual(0,
step_op.device_option.device_type)
self.assertEqual(1,
step_op.device_option.cuda_gpu_id)
elif arg.name == 'backward_step_net':
self.assertEqual("", arg.s)
def initialize_gpu_from_weights_file(model, weights_file, gup_id=0):
logger.info('Loading weights from: {}'.format(weights_file))
ws_blobs = workspace.Blobs()
src_blobs = load_object(weights_file)
if 'cfg' in src_blobs:
saved_cfg = load_cfg(src_blobs['cfg'])
configure_bbox_reg_weights(model, saved_cfg)
if 'blobs' in src_blobs:
src_blobs = src_blobs['blobs']
unscoped_para_names = OrderedDict()
for blob in model.params:
unscoped_para_names[c2_utils.UnscopeName(str(blob))] = True
with c2_utils.NamedCudaScope(gup_id):
for unscoped_para_name in unscoped_para_names.keys():
if (unscoped_para_name.find(']_') >= 0
and unscoped_para_name not in src_blobs):
# Special case for sharing initialization from a pretrained
# model:
# If a blob named '_[xyz]_foo' is in model.params and not in
# the initialization blob dictionary, then load source blob
# 'foo' into destination blob '_[xyz]_foo'
src_name = unscoped_para_name[unscoped_para_name.find((']_') +
2)]
else:
src_name = unscoped_para_name
if src_name not in src_blobs:
logger.info('{:s} not found.'.format(src_name))
continue
dst_name = core.ScopedName(unscoped_para_name)
has_momentum = src_name + '_momentum' in src_blobs
has_momentum_str = ' [+ momentum] ' if has_momentum else ''
logger.info(
'{:s}{:} loaded form weights file into {:s}: {}'.format(
src_name, has_momentum_str, dst_name,
src_blobs[src_name].shape))
if dst_name in ws_blobs:
# if the blob is already in the workspace, make sure that it matches
# the shape of the loaded blob
ws_blob = workspace.FetchBlob(dst_name)
assert ws_blob.shape == src_blobs[src_name].shape, \
('Workspace blob {} with shape {} does not match '
'weights file shape {}').format(
src_name,
ws_blob.shape,
src_blobs[src_name].shape)
workspace.FeedBlob(
dst_name, src_blobs[src_name].astype(np.float32, copy=False))
if has_momentum:
workspace.FeedBlob(
dst_name + '_momentum',
src_blobs[src_name + '_momentum'].astype(np.float32,
copy=False))
for src_name in src_blobs.keys():
if (src_name not in unscoped_para_names
and not src_name.endswith('_momentum')
and src_blobs[src_name] is not None):
with c2_utils.CpuScope():
workspace.FeedBlob('__presever__/{:s}'.format(src_name),
src_blobs[src_name])
logger.info(
'{:s} preserved in workspace (unused)'.format(src_name))
def test_meta_net_def_net_runs(self):
for param, value in self.params.items():
workspace.FeedBlob(param, value)
extra_init_net = core.Net('extra_init')
extra_init_net.ConstantFill('data', 'data', value=1.0)
pem = pe.PredictorExportMeta(
predict_net=self.predictor_export_meta.predict_net,
parameters=self.predictor_export_meta.parameters,
inputs=self.predictor_export_meta.inputs,
outputs=self.predictor_export_meta.outputs,
shapes=self.predictor_export_meta.shapes,
extra_init_net=extra_init_net,
net_type='dag',
)
db_type = 'minidb'
db_file = tempfile.NamedTemporaryFile(
delete=False, suffix=".{}".format(db_type))
pe.save_to_db(
db_type=db_type,
db_destination=db_file.name,
predictor_export_meta=pem)
workspace.ResetWorkspace()
meta_net_def = pe.load_from_db(
db_type=db_type,
filename=db_file.name,
)
self.assertTrue("data" not in workspace.Blobs())
self.assertTrue("y" not in workspace.Blobs())
init_net = pred_utils.GetNet(meta_net_def, pc.PREDICT_INIT_NET_TYPE)
# 0-fills externalblobs blobs and runs extra_init_net
workspace.RunNetOnce(init_net)
self.assertTrue("data" in workspace.Blobs())
self.assertTrue("y" in workspace.Blobs())
print(workspace.FetchBlob("data"))
np.testing.assert_array_equal(
workspace.FetchBlob("data"), np.ones(shape=(1, 5)))
np.testing.assert_array_equal(
workspace.FetchBlob("y"), np.zeros(shape=(1, 10)))
# Load parameters from DB
global_init_net = pred_utils.GetNet(meta_net_def,
pc.GLOBAL_INIT_NET_TYPE)
workspace.RunNetOnce(global_init_net)
# Run the net with a reshaped input and verify we are
# producing good numbers (with our custom implementation)
workspace.FeedBlob("data", np.random.randn(2, 5).astype(np.float32))
predict_net = pred_utils.GetNet(meta_net_def, pc.PREDICT_NET_TYPE)
self.assertEqual(predict_net.type, 'dag')
workspace.RunNetOnce(predict_net)
np.testing.assert_array_almost_equal(
workspace.FetchBlob("y"),
workspace.FetchBlob("data").dot(self.params["y_w"].T) +
self.params["y_b"])
loss = np.zeros(total_iters)
# Now, we will manually run the network for 200 iterations.
data_array = []
drop1_array = []
fc2_array = []
for i in range(total_iters):
#for i in range(1):
workspace.RunNet(train_model.net)
accuracy[i] = workspace.FetchBlob('accuracy')
loss[i] = workspace.FetchBlob('loss')
print('iter {0} loss = {1} '.format(i, loss[i]))
print(' accuracy = {0} '.format(accuracy[i]))
print("Current blobs in the workspace: {}".format(workspace.Blobs()))
print("Workspace has blob 'data'? {}".format(workspace.HasBlob("data")))
#print("Fetched data:\n{}".format(workspace.FetchBlob("data")))
data_array.append(workspace.FetchBlob("data"))
print('data_array', np.shape(data_array))
print("Workspace has blob 'drop1'? {}".format(workspace.HasBlob("drop1")))
#print("Fetched drop1:\n{}".format(workspace.FetchBlob("drop1")))
drop1_array.append(workspace.FetchBlob("drop1"))
print('drop1_array', np.shape(drop1_array))
print("Workspace has blob 'fc2'? {}".format(workspace.HasBlob("fc2")))
#print("Fetched fc2:\n{}".format(workspace.FetchBlob("fc2")))
fc2_array.append(workspace.FetchBlob("fc2"))
print('fc2_array', np.shape(fc2_array))
def from_caffe2(self, init_net, predict_net):
"""Construct Relay expression from caffe2 graph.
Parameters
----------
init_net : protobuf object
predict_net : protobuf object
Returns
-------
mod : tvm.relay.Module
The module that optimizations will be performed on.
params : dict
A dict of name: tvm.nd.array pairs, used as pretrained weights
"""
from caffe2.python import workspace
workspace.RunNetOnce(init_net)
# Input
input_name = predict_net.op[0].input[0]
# Params
self._params = {}
used_blobs = set()
for c2_op in predict_net.op:
for i in c2_op.input:
used_blobs.add(i)
for blob in workspace.Blobs():
if blob in used_blobs and blob != input_name:
self._params[blob] = _nd.array(workspace.FetchBlob(blob))
# Variables
self._nodes = {}
for blob in predict_net.external_input:
if blob in self._params:
self._nodes[blob] = new_var(blob,
shape=self._params[blob].shape,
dtype=self._params[blob].dtype)
else:
shape = self._shape[blob] if blob in self._shape else ()
if isinstance(self._dtype, dict) and blob in self._dtype:
dtype = str(self._dtype[blob])
elif isinstance(self._dtype, str):
dtype = self._dtype
else:
dtype = "float32"
self._nodes[blob] = new_var(blob, shape=shape, dtype=dtype)
# Ops
for c2_op in predict_net.op:
for blob in c2_op.output:
self._ops[blob] = c2_op
for c2_op in predict_net.op:
self._process_op(c2_op)
# Outputs
out = []
for blob in predict_net.external_output:
out.append(self._nodes[blob])
if len(out) > 1:
outputs = _expr.Tuple(out)
else:
outputs = out[0]
func = _expr.Function(ir_pass.free_vars(outputs), outputs)
self._mod[self._mod.entry_func] = func
return self._mod, self._params
final_image = transposed_image
print("Shape of final_image: " + str(np.array(final_image).shape))
with open(MODEL_ROOT + "/init_net.pb", "rb") as f:
init_net = f.read()
with open(MODEL_ROOT + "/predict_net.pb", "rb") as f:
predict_net = f.read()
workspace.ResetWorkspace()
blob_name = model_props[MODEL].blob_name
workspace.FeedBlob(blob_name, final_image)
print("The blobs in the workspace after FeedBlob: {}".format(
workspace.Blobs()))
# Create a predictor using the loaded model.
p = workspace.Predictor(init_net, predict_net)
start = time.time()
for i in range(0, args.iterations):
results = p.run([final_image])
end = time.time()
if args.time:
print("Wall time per iteration (s): {:0.4f}".format(
(end - start) / args.iterations))
max_idx = np.argmax(results[0][0])
sum_probability = sum(results[0][0])
def InferTensorRunAndCompare(self, model):
'''
Runs shape inference, and then the model to check
that the inferred shapes agree with the actual ones
'''
(shapes, types) = workspace.InferShapesAndTypes(
[model.param_init_net, model.net], )
# .. Create net
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net, True)
workspace.RunNet(model.Proto().name)
# ... and then check the shapes mismatch
correct_shapes = {}
correct_types = {}
for b in workspace.Blobs():
arr = workspace.FetchBlob(b)
correct_shapes[b] = arr.shape
if type(arr) is np.ndarray:
if arr.dtype == np.dtype('float32'):
correct_types[b] = caffe2_pb2.TensorProto.FLOAT
elif arr.dtype == np.dtype('int32'):
correct_types[b] = caffe2_pb2.TensorProto.INT32
# BYTE
# STRING
elif arr.dtype == np.dtype('bool'):
correct_types[b] = caffe2_pb2.TensorProto.BOOL
elif arr.dtype == np.dtype('uint8'):
correct_types[b] = caffe2_pb2.TensorProto.UINT8
elif arr.dtype == np.dtype('int8'):
correct_types[b] = caffe2_pb2.TensorProto.INT8
elif arr.dtype == np.dtype('uint16'):
correct_types[b] = caffe2_pb2.TensorProto.UINT16
elif arr.dtype == np.dtype('int16'):
correct_types[b] = caffe2_pb2.TensorProto.INT16
elif arr.dtype == np.dtype('int64'):
correct_types[b] = caffe2_pb2.TensorProto.INT64
elif arr.dtype == np.dtype('float16'):
correct_types[b] = caffe2_pb2.TensorProto.FLOAT16
elif arr.dtype == np.dtype('float64'):
correct_types[b] = caffe2_pb2.TensorProto.DOUBLE
else:
correct_types[b] = "unknown {}".format(arr.dtype)
else:
correct_types[b] = str(type(arr))
for b in correct_shapes:
self.assertTrue(
np.array_equal(
np.array(shapes[b]).astype(np.int32),
np.array(correct_shapes[b]).astype(np.int32)),
"Shape {} mismatch: {} vs. {}".format(b, shapes[b],
correct_shapes[b]))
self.assertFalse(
b not in types and b in correct_types,
"Type for {} not defined".format(b),
)
self.assertEqual(
types[b], correct_types[b],
"Type {} mismatch: {} vs. {}".format(
b,
types[b],
correct_types[b],
))
from caffe2.python.cnn import CNNModelHelper
import unittest
import numpy as np
m, k, n = (1, 28 * 28, 10) # [m][k] * [k][n] = [m][n]
x = np.random.rand(m, k).astype(np.float32) - 0.5 # x = m*k 2D tensor
workspace.ResetWorkspace() # clear workspace
workspace.FeedBlob("x", x) # feed x as a blob
model = ModelHelper(name="test_model") # create model
model.Proto() # print model's protocol buffer before add operator
brew.fc(
model, "x", "y", k, n
) # fully connected NN, weight = k*n 2D tensor /// bias, y = m*n 2D tensor
brew.softmax(model, "y", "z")
model.Validate()
model.Proto() # print model's protocol buffer after add operator
workspace.RunNetOnce(
model.param_init_net) # init [y_w(weight), y_b(bias) (randomize)]
# weight is 2D array, bias is 1D array
workspace.Blobs() # print workspace's blobs
# workspace.FetchBlob("y_w")
# workspace.FetchBlob("y_b")
workspace.RunNetOnce(model.net)
# y = workspace.FetchBlob("y")
# z = workspace.FetchBlob("z")
def load_save(self, src_device_type, src_gpu_id,
dst_device_type, dst_gpu_id):
workspace.ResetWorkspace()
dtypes = [np.float16, np.float32, np.float64, np.bool, np.int8,
np.int16, np.int32, np.int64, np.uint8, np.uint16]
arrays = [np.random.permutation(6).reshape(2, 3).astype(T)
for T in dtypes]
src_device_option = core.DeviceOption(
src_device_type, src_gpu_id)
dst_device_option = core.DeviceOption(
dst_device_type, dst_gpu_id)
for i, arr in enumerate(arrays):
self.assertTrue(workspace.FeedBlob(str(i), arr, src_device_option))
self.assertTrue(workspace.HasBlob(str(i)))
try:
# Saves the blobs to a local db.
tmp_folder = tempfile.mkdtemp()
op = core.CreateOperator(
"Save",
[str(i) for i in range(len(arrays))], [],
absolute_path=1,
db=os.path.join(tmp_folder, "db"), db_type=self._db_type)
self.assertTrue(workspace.RunOperatorOnce(op))
# Reset the workspace so that anything we load is surely loaded
# from the serialized proto.
workspace.ResetWorkspace()
self.assertEqual(len(workspace.Blobs()), 0)
def _LoadTest(keep_device, device_type, gpu_id, blobs, loadAll):
"""A helper subfunction to test keep and not keep."""
op = core.CreateOperator(
"Load",
[], blobs,
absolute_path=1,
db=os.path.join(tmp_folder, "db"), db_type=self._db_type,
device_option=dst_device_option,
keep_device=keep_device,
load_all=loadAll)
self.assertTrue(workspace.RunOperatorOnce(op))
for i, arr in enumerate(arrays):
self.assertTrue(workspace.HasBlob(str(i)))
fetched = workspace.FetchBlob(str(i))
self.assertEqual(fetched.dtype, arr.dtype)
np.testing.assert_array_equal(
workspace.FetchBlob(str(i)), arr)
proto = caffe2_pb2.BlobProto()
proto.ParseFromString(workspace.SerializeBlob(str(i)))
self.assertTrue(proto.HasField('tensor'))
self.assertEqual(proto.tensor.device_detail.device_type,
device_type)
if device_type == caffe2_pb2.CUDA:
self.assertEqual(proto.tensor.device_detail.cuda_gpu_id,
gpu_id)
blobs = [str(i) for i in range(len(arrays))]
# Load using device option stored in the proto, i.e.
# src_device_option
_LoadTest(1, src_device_type, src_gpu_id, blobs, 0)
# Load again, but this time load into dst_device_option.
_LoadTest(0, dst_device_type, dst_gpu_id, blobs, 0)
# Load back to the src_device_option to see if both paths are able
# to reallocate memory.
_LoadTest(1, src_device_type, src_gpu_id, blobs, 0)
# Reset the workspace, and load directly into the dst_device_option.
workspace.ResetWorkspace()
_LoadTest(0, dst_device_type, dst_gpu_id, blobs, 0)
# Test load all which loads all blobs in the db into the workspace.
workspace.ResetWorkspace()
_LoadTest(1, src_device_type, src_gpu_id, [], 1)
# Load again making sure that overwrite functionality works.
_LoadTest(1, src_device_type, src_gpu_id, [], 1)
# Load again with different device.
_LoadTest(0, dst_device_type, dst_gpu_id, [], 1)
workspace.ResetWorkspace()
_LoadTest(0, dst_device_type, dst_gpu_id, [], 1)
finally:
# clean up temp folder.
try:
shutil.rmtree(tmp_folder)
except OSError as e:
if e.errno != errno.ENOENT:
raise
def main():
workspace.GlobalInit(['caffe2', '--caffe2_log_level=0'])
args = parse_args()
logger.info('Called with args:')
logger.info(args)
# with open(args.init_net_path) as f:
# init_net = f.read()
# with open(args.predict_net_path) as f:
# predict_net = f.read()
# p = workspace.Predictor(init_net, predict_net)
# img = np.zeros((1,3,256,256), dtype=np.float32)
# workspace.FeedBlob('data', img)
# results = p.run({'data': img})
init_def = caffe2_pb2.NetDef()
with open(args.init_net_path, 'r') as f:
init_def.ParseFromString(f.read())
# init_def.device_option.CopyFrom(device_options)
net_def = caffe2_pb2.NetDef()
with open(args.predict_net_path, 'r') as f:
net_def.ParseFromString(f.read())
# net_def.device_option.CopyFrom(device_options)
# model = model_helper.ModelHelper(arg_scope=arg_scope)
# model = cnn.CNNModelHelper()
model = detector.DetectionModelHelper(name=net_def.name,
train=True,
num_classes=1000,
init_params=True)
predict_net = core.Net(net_def)
init_net = core.Net(init_def)
model.param_init_net.AppendNet(init_net)
model.net.AppendNet(predict_net)
model.params.extend([
core.BlobReference(x) for x in predict_net.Proto().external_input
if x != 'data'
])
# add_training_operators(model, 'pred', 'label')
blob_names = ['data', 'label']
for gpu_id in range(1):
with c2_utils.NamedCudaScope(gpu_id):
for blob_name in blob_names:
workspace.CreateBlob(core.ScopedName(blob_name))
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net, overwrite=True)
out_file_name = os.path.join(args.out_dir, net_def.name + '.pkl')
net_utils.save_model_to_weights_file(out_file_name, model)
# workspace.CreateNet(init_def)
# workspace.CreateNet(net_def)
# workspace.RunNet(net_def)
# workspace.RunNet(init_def)
print(type(init_def))
print(net_def.name)
print(workspace.blobs)
print(len(workspace.blobs))
print(workspace.Blobs())
