def tower_loss(data_tensor, label_tensor, num_classes, train_mode):
with slim.arg_scope(my_vgg_relu6_fc.vgg_arg_scope(weight_decay=args.weight_decay)):
logits, endpoints_dict = my_vgg_relu6_fc.vgg_16_tcomb(data_tensor, num_classes=num_classes, is_training=train_mode,dropout_keep_prob=args.dropout_keep_prob)
loss=tf.reduce_mean(tf.losses.sparse_softmax_cross_entropy(labels=label_tensor, logits=logits))
return loss, logits
def main(args):
pdb.set_trace()
#modelDir = '/trn_dir/models/model1'
model_path = '/trn_dir/models/model3_relu6/model-000039'
num_classes=20
im0 = Image.open('/trn_dir/imgnet/data/trn/whippet/n02091134_12142.JPEG')
im0=im0.resize((224,224))
image_np = load_image_into_numpy_array(im0)
image_np_expanded = np.expand_dims(image_np, axis=0)
vgg16_variable_names, vgg16_t_variable_names = my_get_variable_names()
#builder = tf.saved_model.builder.SavedModelBuilder(modelDir)
# --------------------------------------------------------------------------
# In TensorFlow, you first want to define the computation graph with all the
# necessary operations: loss, training op, accuracy...
# Any tensor created in the `graph.as_default()` scope will be part of `graph`
graph = tf.Graph()
with graph.as_default():
# Indicates whether we are in training or in test mode
is_training = tf.placeholder(tf.bool)
images = tf.placeholder(dtype=tf.float32, shape=[None,224,224,3])
res = layers.conv2d(images, 64, [3,3], scope='tmp')
ninp_ch = tf.shape(res)[3]
chidx = tf.range(0,tf.cast(ninp_ch/2,tf.int32))
rndidx = tf.random.shuffle(chidx)
res2 = tf.gather(res, chidx, axis=3) #res[:,:,:,rndidx,:]
res2.set_shape([None,res.shape[1],res.shape[2], res.shape[3].__floordiv__(2)] )
res1 = layers.conv2d(res2, 128, [3,3], scope='tmp_2c')
chidx1 = tf.range(ninp_ch)
chidx1_shift = tf.mod(chidx1+1,ninp_ch)
#vgg = tf.contrib.slim.nets.vgg
with slim.arg_scope(my_vgg_relu6_fc.vgg_arg_scope(weight_decay=args.weight_decay)):
logits, endpoints_dict = my_vgg_relu6_fc.vgg_16(images, num_classes=num_classes, is_training=is_training,dropout_keep_prob=args.dropout_keep_prob)
with slim.arg_scope(my_vgg_relu6_fc.vgg_arg_scope(weight_decay=args.weight_decay)):
logits_t, endpoints_dict_t = my_vgg_relu6_fc.vgg_16_t(images, num_classes=num_classes, is_training=is_training,dropout_keep_prob=args.dropout_keep_prob)
# Specify where the model checkpoint is (pretrained weights).
#model_path = args.model_path
#assert(os.path.isfile(model_path))
# Restore only the layers up to fc7 (included)
# Calling function `init_fn(sess)` will load all the pretrained weights.
main_variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='vgg_16')
#target_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='vgg_16_t')
#variables_to_restore = tf.contrib.framework.get_variables_to_restore()#exclude=['vgg_16/fc8'])
vgg_trainable_names = tf.trainable_variables(scope='vgg_16')
vgg_t_trainable_names = tf.trainable_variables('vgg_16_t')
variables_to_restore = tf.contrib.framework.get_variables_to_restore(vgg16_variable_names)
var_load_fn = tf.contrib.framework.assign_from_checkpoint_fn(model_path, variables_to_restore)
#new_variables = tf.contrib.framework.get_variables('vgg_16_ext')
#new_var_init = tf.variables_initializer(var_list=fc8_variables)
copy_op = copy_values_from_vgg16(vgg16_variable_names, vgg16_t_variable_names)
conv_weights = tf.contrib.framework.get_variables('vgg_16/conv2/conv2_1/weights:0')
zero_weights = tf.zeros_like(conv_weights, dtype=conv_weights[0].dtype)
ext_weights = tf.concat([conv_weights, zero_weights], axis=3)
c1_1w=tf.contrib.framework.get_variables('vgg_16/conv1/conv1_1/biases:0')
tc1_1w=tf.contrib.framework.get_variables('vgg_16_t/conv1/conv1_1/biases:0')
assign_op = tf.assign(tc1_1w[0],tf.identity(c1_1w[0]))
idx = tf.range(0,15)
randidx = tf.random.shuffle(idx)
'''
conv_weights = tf.contrib.framework.get_variables('vgg_16_ext/conv2/conv2_1/weights:0')
conv_input_dim = tf.shape(conv_weights)[3]
new_weights = tf.contrib.framework.get_variables('vgg_16_ext/conv2/conv2_1/weights:0')
ext_weights = tf.concat([conv_weights, new_weights[:,:,:,conv_input_dim:,:]], axis=3)
'''
'''
'''
# Initialization operation from scratch for the new "fc8" layers
# `get_variables` will only return the variables whose name starts with the given pattern
global_init=tf.global_variables_initializer()
with tf.Session(graph=graph) as sess:
pdb.set_trace()
#sess.run(new_var_init)
sess.run(global_init)
var_load_fn(sess) # load the pretrained weights
# print name, shape of each trainable variable
trainable_values = sess.run(vgg_trainable_names)
for nm, v in zip(vgg_trainable_names, trainable_values):
print('name = ',nm.name, ' shape = ',v.shape)
conv_val = sess.run(conv_weights, {images:image_np_expanded, is_training:True})
ext_val = sess.run(ext_weights, {images:image_np_expanded, is_training:True})
randidx = sess.run(randidx, {images:image_np_expanded, is_training:True})
pp=1
def main(args):
pdb.set_trace()
#modelDir = '/trn_dir/models/model1'
#model_path = '/trn_dir/models/model_multi_cl20_relu6/model-000075'
model_path = '/trn_dir/models/model_multi_cl20_preprocess/model-000060'
num_classes = 20
im0 = Image.open('/trn_dir/imgnet/data/trn/whippet/n02091134_12142.JPEG')
im0 = im0.resize((224, 224))
image_np = load_image_into_numpy_array(im0)
image_np_expanded = np.expand_dims(image_np, axis=0)
vgg16_variable_names, vgg16_t_variable_names = my_get_variable_names()
#builder = tf.saved_model.builder.SavedModelBuilder(modelDir)
# --------------------------------------------------------------------------
# In TensorFlow, you first want to define the computation graph with all the
# necessary operations: loss, training op, accuracy...
# Any tensor created in the `graph.as_default()` scope will be part of `graph`
graph = tf.Graph()
with graph.as_default():
# Indicates whether we are in training or in test mode
is_training = tf.placeholder(tf.bool)
images = tf.placeholder(dtype=tf.float32, shape=[None, 224, 224, 3])
#vgg = tf.contrib.slim.nets.vgg
with slim.arg_scope(
my_vgg_relu6_fc.vgg_arg_scope(weight_decay=args.weight_decay)):
logits, endpoints_dict = my_vgg_relu6_fc.vgg_16(
images,
num_classes=num_classes,
is_training=is_training,
dropout_keep_prob=args.dropout_keep_prob)
main_variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='vgg_16')
#target_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='vgg_16_t')
#variables_to_restore = tf.contrib.framework.get_variables_to_restore()#exclude=['vgg_16/fc8'])
vgg_trainable_names = tf.trainable_variables(scope='vgg_16')
variables_to_restore = tf.contrib.framework.get_variables_to_restore(
vgg16_variable_names)
var_load_fn = tf.contrib.framework.assign_from_checkpoint_fn(
model_path, variables_to_restore)
global_init = tf.global_variables_initializer()
with tf.Session(graph=graph) as sess:
pdb.set_trace()
#sess.run(new_var_init)
sess.run(global_init)
var_load_fn(sess) # load the pretrained weights
# print name, shape of each trainable variable
savename = os.path.basename(model_path) + '_featcorr.json'
vgg_variable_values = {}
feat_corr = {}
trainable_values = sess.run(vgg_trainable_names)
for nm, v in zip(vgg_trainable_names[:44], trainable_values[:44]):
print('name = ', nm.name, ' shape = ', v.shape)
if 'vgg_16_tcorr' not in nm.name:
vgg_variable_values[nm.name] = np.array(v)
for vname in vgg16_variable_names:
print(vname)
wname = vname + '/weights:0'
bname = vname + '/biases:0'
feat_corr[vname] = compute_feature_correlation(
vgg_variable_values[wname],
vgg_variable_values[bname]).tolist()
with open(savename, 'w') as fid:
json.dump(feat_corr, fid)
fid.close()
#feat_corr[nm.name] = compute_feature_correlation(vgg_variable_values[nm.name])
#feat_corr = compute_feature_correlation(vgg_variable_values['vgg_16/conv3/conv3_2/weights:0'])
pp = 1