def raw_model(n_classes):
model_dir = './models'
model_weight_filename = os.path.join(model_dir, 'sports1M_weights_tf.h5')
model_json_filename = os.path.join(model_dir, 'sports1M_weights_tf.json')
print("[Info] Reading model architecture...")
# model = model_from_json(open(model_json_filename, 'r').read())
model = c3d_model.get_model(backend=backend)
print("[Info] Loading model weights...")
model.load_weights(model_weight_filename)
model.add(Dense(n_classes, activation='softmax'))
print("[Info] Loading model weights -- DONE!")
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
return model
def main():
#dim_ordering = 'th'
#dim_ordering = 'th'
import keras.backend as K
dim_ordering = K.image_dim_ordering()
print("[Info] image_dim_order (from default ~/.keras/keras.json)={}",
dim_ordering)
# get C3D model placeholder
model = c3d_model.get_model(summary=True, backend=dim_ordering)
# input caffe model
caffe_model_filename = './models/conv3d_deepnetA_sport1m_iter_1900000'
# output dir/files
model_dir = './models'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
output_model_filename = os.path.join(
model_dir, 'sports1M_weights_{}.h5'.format(dim_ordering))
output_json_filename = os.path.join(
model_dir, 'sports1M_weights_{}.json'.format(dim_ordering))
# read caffe model
print("-" * 19)
print("Reading model file={}...", caffe_model_filename)
p = caffe.NetParameter()
p.ParseFromString(open(caffe_model_filename, 'rb').read())
params = []
print("-" * 19)
print("Converting model...")
# read every conv/fc layer and append to "params" list
for i in range(len(p.layers)):
layer = p.layers[i]
# skip non-conv/fc layers
if 'conv' not in layer.name and 'fc' not in layer.name:
continue
print("[Info] Massaging \"{}\" layer...", layer.name)
weights_b = np.array(layer.blobs[1].data, dtype=np.float32)
weights_p = np.array(layer.blobs[0].data, dtype=np.float32).reshape(
layer.blobs[0].num,
layer.blobs[0].channels,
layer.blobs[0].length,
layer.blobs[0].height,
layer.blobs[0].width,
)
if 'conv' in layer.name:
# theano vs tensorflow: https://github.com/fchollet/keras/blob/master/keras/utils/np_utils.py#L90-L115
if dim_ordering == 'th':
weights_p = reindex(weights_p)
else:
weights_p = np.transpose(weights_p, (2, 3, 4, 1, 0))
elif 'fc' in layer.name:
weights_p = weights_p[0, 0, 0, :, :].T
if 'fc6' in layer.name:
print(
"[Info] First FC layer after flattening layer needs special care..."
)
weights_p = convert_dense(weights_p)
params.append([weights_p, weights_b])
valid_layer_count = 0
for layer_indx in range(len(model.layers)):
layer_name = model.layers[layer_indx].name
if 'conv' in layer_name or 'fc' in layer_name:
print("[Info] Transplanting \"{}\" layer...", layer_name)
model.layers[layer_indx].set_weights(params[valid_layer_count])
valid_layer_count += 1
print("-" * 19)
print("Saving pre-trained model weights as {}...", output_model_filename)
model.save_weights(output_model_filename, overwrite=True)
json_string = model.to_json()
with open(output_json_filename, 'w') as f:
f.write(json_string)
print("-" * 39)
print("Conversion done!")
print("-" * 39)
import c3d_model
import numpy as np
from keras.layers import Input, Dense
from keras.models import Sequential, Model
c3d_model = c3d_model.get_model(nb_classes=22)
weights_path = '/media/lq/C13E-1ED0/dataset/UCF_Crimes/results/gan_1/weights/c3d_TC_GAN_22_outputs_it1000.hdf5'
save_path = weights_path.replace('_22_outputs_', '_21_outputs_')
n_classes =21
c3d_model.load_weights(weights_path)
w = c3d_model.get_layer('fc8').get_weights()
print(w)
w[0] = np.delete(w[0], np.s_[-1], axis=1)
w[1] = np.delete(w[1], np.s_[-1])
#Deleting the old output layer
c3d_model.layers.pop()
last_layer = c3d_model.get_layer('dropout_2').output
#New output layer
out = Dense(n_classes, activation='softmax', name='fc8')(last_layer)
c3d_model = Model(inputs=c3d_model.input, outputs=out)
c3d_model.get_layer('fc8').set_weights(w)
print(c3d_model.get_layer('fc8').get_weights())
c3d_model.summary()
c3d_model.save_weights(save_path)
def main():
# dim_ordering = 'th'
# dim_ordering = 'th'
import tensorflow.keras.backend as K
dim_ordering = "tf"
print(
"[Info] image_dim_order (from default ~/.keras/keras.json)={}".format(
dim_ordering))
# get C3D model placeholder
print(c3d_model.get_model.__code__.co_varnames)
gpus = tf.config.list_physical_devices('GPU')
if gpus:
try:
# Currently, memory growth needs to be the same across GPUs
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(gpus), "Physical GPUs,", len(logical_gpus),
"Logical GPUs")
except RuntimeError as e:
# Memory growth must be set before GPUs have been initialized
print(e)
print("test1")
model = c3d_model.get_model(summary=True, backend=dim_ordering)
print("test2")
# input caffe model
caffe_model_filename = './models/conv3d_deepnetA_sport1m_iter_1900000'
# output dir/files
model_dir = './models'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
output_model_filename = os.path.join(
model_dir, 'sports1M_weights_{}.h5'.format(dim_ordering))
output_json_filename = os.path.join(
model_dir, 'sports1M_weights_{}.json'.format(dim_ordering))
# read caffe model
print("-" * 19)
print("Reading model file={}...".format(caffe_model_filename))
p = caffe.NetParameter()
p.ParseFromString(open(caffe_model_filename, 'rb').read())
params = []
print("-" * 19)
print("Converting model...")
# read every conv/fc layer and append to "params" list
for i in range(len(p.layers)):
layer = p.layers[i]
# skip non-conv/fc layers
if 'conv' not in layer.name and 'fc' not in layer.name:
continue
print("[Info] Massaging \"{}\" layer...".format(layer.name))
weights_b = np.array(layer.blobs[1].data, dtype=np.float32)
weights_p = np.array(layer.blobs[0].data, dtype=np.float32).reshape(
layer.blobs[0].num,
layer.blobs[0].channels,
layer.blobs[0].length,
layer.blobs[0].height,
layer.blobs[0].width,
)
if 'conv' in layer.name:
# theano vs tensorflow: https://github.com/fchollet/keras/blob/master/keras/utils/np_utils.py#L90-L115
if dim_ordering == 'th':
weights_p = reindex(weights_p)
else:
weights_p = np.transpose(weights_p, (2, 3, 4, 1, 0))
elif 'fc' in layer.name:
weights_p = weights_p[0, 0, 0, :, :].T
if 'fc6' in layer.name:
print("[Info] First FC layer after flattening layer needs "
"special care...")
weights_p = convert_dense(weights_p)
params.append([weights_p, weights_b])
valid_layer_count = 0
for layer_indx in range(len(model.layers)):
layer_name = model.layers[layer_indx].name
if 'conv' in layer_name or 'fc' in layer_name:
print("[Info] Transplanting \"{}\" layer...".format(layer_name))
model.layers[layer_indx].set_weights(params[valid_layer_count])
valid_layer_count += 1
print("-" * 19)
print("Saving pre-trained model weights as {}...".format(
output_model_filename))
model.save_weights(output_model_filename, overwrite=True)
json_string = model.to_json()
with open(output_json_filename, 'w') as f:
f.write(json_string)
print("-" * 39)
print("Conversion done!")
print("-" * 39)