def tune(classifier_name, init_shape, final_shape):
kernel = weights_file[classifier_name][classifier_name]['kernel:0'].value
bias = weights_file[classifier_name][classifier_name]['bias:0'].value
if init_shape == kernel.shape:
new_kernel, new_bias = sample_tensors(
weights_list=[kernel, bias],
sampling_instructions=[
final_shape[0], final_shape[1], final_shape[2], final_shape[3]
],
axes=[
[3]
], # The one bias dimension corresponds to the last kernel dimension.
init=['gaussian', 'zeros'],
mean=0.0,
stddev=0.005)
del tuned_weights_file[classifier_name][classifier_name]['kernel:0']
del tuned_weights_file[classifier_name][classifier_name]['bias:0']
tuned_weights_file[classifier_name][classifier_name].create_dataset(
name='kernel:0', data=new_kernel)
tuned_weights_file[classifier_name][classifier_name].create_dataset(
name='bias:0', data=new_bias)
print("{}, shape is tuned:\n\tKernel: {} -> {}\n\tBias: {} -> {}\n".
format(classifier_name, kernel.shape, new_kernel.shape,
bias.shape, new_bias.shape))
def apprentissage_trft():
poids_source_lien = 'C:\\Users\\lamin\\Desktop\\PFE\\traffic_objetct_detection\\VGG_coco_SSD_300x300_iter_400000.h5'
poids_destination_lien='C:\\Users\\lamin\\Desktop\\PFE\\traffic_objetct_detection\\VGG_SSD_300x300_8_classes.h5'
shutil.copy(poids_source_lien, poids_destination_lien)
poids_source_file = h5py.File(poids_source_lien, 'r')
poids_destination_file = h5py.File(poids_destination_lien)
classifier_nom = ['conv4_3_norm_mbox_conf',
'fc7_mbox_conf',
'conv6_2_mbox_conf',
'conv7_2_mbox_conf',
'conv8_2_mbox_conf',
'conv9_2_mbox_conf']
n_classes_source = 81
classes_interet = [0, 3, 8, 1, 2, 10, 4, 6, 12]
for nom in classifier_nom:
kernel = poids_source_file[nom][nom]['kernel:0'].value
bias = poids_source_file[nom][nom]['bias:0'].value
height, width, in_channels, out_channels = kernel.shape
if isinstance(classes_interet, (list, tuple)):
subsampling_indices = []
for i in range(int(out_channels/n_classes_source)):
indices = np.array(classes_interet) + i * n_classes_source
subsampling_indices.append(indices)
subsampling_indices = list(np.concatenate(subsampling_indices))
elif isinstance(classes_interet, int):
subsampling_indices = int(classes_interet * (out_channels/n_classes_source))
else:
raise ValueError("`classes_interet` doit etre entier ou(list/tuple).")
new_kernel, new_bias = sample_tensors(weights_list=[kernel, bias],
sampling_instructions=[height, width, in_channels, subsampling_indices],
axes=[[3]],
init=['gaussian', 'zeros'],
mean=0.0,
stddev=0.005)
del poids_destination_file[name][name]['kernel:0']
del poids_destination_file[name][name]['bias:0']
poids_destination_file[name][name].create_dataset(name='kernel:0', data=new_kernel)
poids_destination_file[name][name].create_dataset(name='bias:0', data=new_bias)
poids_destination_file.flush()
for i in range(int(out_channels/n_classes_source)):
indices = np.array(classes_of_interest) + i * n_classes_source
subsampling_indices.append(indices)
subsampling_indices = list(np.concatenate(subsampling_indices))
elif isinstance(classes_of_interest, int):
subsampling_indices = int(classes_of_interest * (out_channels/n_classes_source))
else:
raise ValueError("`classes_of_interest` must be either an integer or a list/tuple.")
# Sub-sample the kernel and bias.
# The `sample_tensors()` function used below provides extensive
# documentation, so don't hesitate to read it if you want to know
# what exactly is going on here.
new_kernel, new_bias = sample_tensors(weights_list=[kernel, bias],
sampling_instructions=[height, width, in_channels, subsampling_indices],
axes=[[3]], # The one bias dimension corresponds to the last kernel dimension.
init=['gaussian', 'zeros'],
mean=0.0,
stddev=0.005)
# Delete the old weights from the destination file.
del weights_destination_file[name][name]['kernel:0']
del weights_destination_file[name][name]['bias:0']
# Create new datasets for the sub-sampled weights.
weights_destination_file[name][name].create_dataset(name='kernel:0', data=new_kernel)
weights_destination_file[name][name].create_dataset(name='bias:0', data=new_bias)
# Make sure all data is written to our output file before this sub-routine exits.
weights_destination_file.flush()
conv4_3_norm_mbox_conf_kernel = weights_destination_file[classifier_names[0]][classifier_names[0]]['kernel:0']
conv4_3_norm_mbox_conf_bias = weights_destination_file[classifier_names[0]][classifier_names[0]]['bias:0']
def sample_model_for_TR():
weights_source_path = '/Users/royhirsch/Downloads/VGG_coco_SSD_512x512_iter_360000.h5'
weights_destination_path = '/Users/royhirsch/Downloads/VGG_coco_SSD_512x512_iter_360000_2_classes.h5'
# Make a copy of the weights file.
shutil.copy(weights_source_path, weights_destination_path)
# Load both the source weights file and the copy we made.
# We will load the original weights file in read-only mode so that we can't mess up anything.
weights_source_file = h5py.File(weights_source_path, 'r')
weights_destination_file = h5py.File(weights_destination_path)
classifier_names = ['conv4_3_norm_mbox_conf',
'fc7_mbox_conf',
'conv6_2_mbox_conf',
'conv7_2_mbox_conf',
'conv8_2_mbox_conf',
'conv9_2_mbox_conf',
'conv10_2_mbox_conf']
# TODO: Set the number of classes in the source weights file. Note that this number must include
# the background class, so for MS COCO's 80 classes, this must be 80 + 1 = 81.
n_classes_source = 81
# TODO: Set the indices of the classes that you want to pick for the sub-sampled weight tensors.
# In case you would like to just randomly sample a certain number of classes, you can just set
# `classes_of_interest` to an integer instead of the list below. Either way, don't forget to
# include the background class. That is, if you set an integer, and you want `n` positive classes,
# then you must set `classes_of_interest = n + 1`.
classes_of_interest = [0, 6]
# classes_of_interest = 12 # Uncomment this in case you want to just randomly sub-sample the last axis instead of providing a list of indices.
for name in classifier_names:
# Get the trained weights for this layer from the source HDF5 weights file.
kernel = weights_source_file[name][name]['kernel:0'].value
bias = weights_source_file[name][name]['bias:0'].value
# Get the shape of the kernel. We're interested in sub-sampling
# the last dimension, 'o'.
height, width, in_channels, out_channels = kernel.shape
# Compute the indices of the elements we want to sub-sample.
# Keep in mind that each classification predictor layer predicts multiple
# bounding boxes for every spatial location, so we want to sub-sample
# the relevant classes for each of these boxes.
if isinstance(classes_of_interest, (list, tuple)):
subsampling_indices = []
for i in range(int(out_channels / n_classes_source)):
indices = np.array(classes_of_interest) + i * n_classes_source
subsampling_indices.append(indices)
subsampling_indices = list(np.concatenate(subsampling_indices))
elif isinstance(classes_of_interest, int):
subsampling_indices = int(classes_of_interest * (out_channels / n_classes_source))
else:
raise ValueError("`classes_of_interest` must be either an integer or a list/tuple.")
# Sub-sample the kernel and bias.
# The `sample_tensors()` function used below provides extensive
# documentation, so don't hesitate to read it if you want to know
# what exactly is going on here.
new_kernel, new_bias = sample_tensors(weights_list=[kernel, bias],
sampling_instructions=[height, width, in_channels, subsampling_indices],
axes=[[3]],
# The one bias dimension corresponds to the last kernel dimension.
init=['gaussian', 'zeros'],
mean=0.0,
stddev=0.005)
# Delete the old weights from the destination file.
del weights_destination_file[name][name]['kernel:0']
del weights_destination_file[name][name]['bias:0']
# Create new datasets for the sub-sampled weights.
weights_destination_file[name][name].create_dataset(name='kernel:0', data=new_kernel)
weights_destination_file[name][name].create_dataset(name='bias:0', data=new_bias)
# Make sure all data is written to our output file before this sub-routine exits.
weights_destination_file.flush()
conv4_3_norm_mbox_conf_kernel = weights_destination_file[classifier_names[0]][classifier_names[0]]['kernel:0']
conv4_3_norm_mbox_conf_bias = weights_destination_file[classifier_names[0]][classifier_names[0]]['bias:0']
print("Shape of the '{}' weights:".format(classifier_names[0]))
print()
print("kernel:\t", conv4_3_norm_mbox_conf_kernel.shape)
print("bias:\t", conv4_3_norm_mbox_conf_bias.shape)
def init_weights(self, classes=''):
# saves coco trained weights in destination # that can be loaded by model
# if passing classes, original coco classes may be passed, e.g 1=person,9=boat
#looky here https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoDemo.ipynb
print('\nloading weights')
#pdb.set_trace()
if classes == '':
classes = self.classes
weights_source_path = f'{json_conf["ssd_base_weights"]}' # /home/gidish/models/VGG_coco_SSD_300x300_iter_400000.h5'
weights_destination_path = f'{json_conf["ssd_base_weights"]}_subsampled_{len(classes)}_classes.h5'
shutil.copy(weights_source_path, weights_destination_path)
weights_source_file = h5py.File(weights_source_path, 'r')
weights_destination_file = h5py.File(weights_destination_path)
classifier_names = [
'conv4_3_norm_mbox_conf', 'fc7_mbox_conf', 'conv6_2_mbox_conf',
'conv7_2_mbox_conf', 'conv8_2_mbox_conf', 'conv9_2_mbox_conf'
]
n_classes_source = 81
classes_of_interest = list(classes)
print('classes are:', classes_of_interest)
subsampling_indices = []
for i in range(int(324 / n_classes_source)):
indices = np.array(classes_of_interest) + i * n_classes_source
subsampling_indices.append(indices)
subsampling_indices = list(np.concatenate(subsampling_indices))
print('last layer indicies', subsampling_indices)
for name in classifier_names:
# Get the trained weights for this layer from the source HDF5 weights file.
kernel = weights_source_file[name][name]['kernel:0'].value
bias = weights_source_file[name][name]['bias:0'].value
# Get the shape of the kernel. We're interested in sub-sampling
# the last dimension, 'o'.
height, width, in_channels, out_channels = kernel.shape
# Compute the indices of the elements we want to sub-sample.
# Keep in mind that each classification predictor layer predicts multiple
# bounding boxes for every spatial location, so we want to sub-sample
# the relevant classes for each of these boxes.
if isinstance(classes_of_interest, (list, tuple)):
subsampling_indices = []
for i in range(int(out_channels / n_classes_source)):
indices = np.array(
classes_of_interest) + i * n_classes_source
subsampling_indices.append(indices)
subsampling_indices = list(np.concatenate(subsampling_indices))
elif isinstance(classes_of_interest, int):
subsampling_indices = int(classes_of_interest *
(out_channels / n_classes_source))
else:
raise ValueError(
"`classes_of_interest` must be either an integer or a list/tuple."
)
# Sub-sample the kernel and bias.
# The `sample_tensors()` function used below provides extensive
# documentation, so don't hesitate to read it if you want to know
# what exactly is going on here.
new_kernel, new_bias = sample_tensors(
weights_list=[kernel, bias],
sampling_instructions=[
height, width, in_channels, subsampling_indices
],
axes=[
[3]
], # The one bias dimension corresponds to the last kernel dimension.
init=['gaussian', 'zeros'],
mean=0.0,
stddev=0.005)
# Delete the old weights from the destination file.
del weights_destination_file[name][name]['kernel:0']
del weights_destination_file[name][name]['bias:0']
# Create new datasets for the sub-sampled weights.
weights_destination_file[name][name].create_dataset(
name='kernel:0', data=new_kernel)
weights_destination_file[name][name].create_dataset(name='bias:0',
data=new_bias)
# Make sure all data is written to our output file before this sub-routine exits.
weights_destination_file.flush()
conv4_3_norm_mbox_conf_kernel = weights_destination_file[
classifier_names[0]][classifier_names[0]]['kernel:0']
conv4_3_norm_mbox_conf_bias = weights_destination_file[
classifier_names[0]][classifier_names[0]]['bias:0']
print("Shape of the '{}' weights:".format(classifier_names[0]))
print()
print("kernel:\t", conv4_3_norm_mbox_conf_kernel.shape)
print("bias:\t", conv4_3_norm_mbox_conf_bias.shape)
return weights_destination_path
