def write(self):
if( self.inputName is None or self.outputNames is None ):
raise ValueError("inputName or outputName not set")
graph = self.loadGraph()
image = self.getImage(False)
if(image is None):
raise ValueError("Null image")
print("Input name: ", self.inputName)
print("Output names: ", self.outputNames)
print("Input shape: ", image.shape)
with tf.Session(graph=graph) as sess:
outputs = sess.run(
# self.outputName,
self.outputNames,
feed_dict={self.inputName: [image]})
print(outputs)
print("Outputs: ",outputs)
toSave = {}
toSave_dtype_dict = {}
for i in range(len(outputs)):
toSave[self.outputNames[i]] = outputs[i]
print("Output: ", self.outputNames[i])
print("Dtype: ", outputs[i].dtype)
toSave_dtype_dict[self.outputNames[i]] = str(outputs[i].dtype)
#print("Values to save: ", toSave)
tfp = TensorFlowPersistor(base_dir=self.baseDir, save_dir=self.name, verbose=False)
tfp._save_input(self.getImage(True), self.inputName)
dtype_dict = {}
dtype_dict[self.inputName] = str(image.dtype)
tfp._save_node_dtypes(dtype_dict)
# tfp._save_predictions({self.outputName:outputs})
tfp._save_predictions(toSave)
tfp._save_node_dtypes(toSave_dtype_dict)
in_arg = args.in_name
out_arg = args.out_name
shape = parse_shape(str_shape)
save_dir = get_save_dir(model_file)
model = load_model(model_file)
layers = model.layers
first = layers[0]
last = layers[-1]
graph = K.get_session().graph
if args.verbose:
print_nodes()
in_name = '{}{}'.format(first.name, in_arg)
in_node = graph.get_tensor_by_name(in_name + ':0')
out_name = '{}{}'.format(last.name, out_arg)
out_node = graph.get_tensor_by_name(out_name + ':0')
with tf.Session(graph=graph) as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
inValue = np.random.rand(*shape)
out = sess.run(out_node, feed_dict={in_node: inValue})
tfp = TensorFlowPersistor(base_dir=base_dir, save_dir=save_dir)
tfp._save_input(inValue, in_name)
tfp._save_predictions({out_name: out})
INPUT_SIZE = 513
INPUT_TENSOR_NAME = 'graph/ImageTensor:0'
OUTPUT_TENSOR_NAME = 'graph/SemanticPredictions:0'
width, height = image.size
resize_ratio = 1.0 * INPUT_SIZE / max(width, height)
target_size = (int(resize_ratio * width), int(resize_ratio * height))
resized_image = image.convert('RGB').resize(target_size, Image.ANTIALIAS)
input = np.asarray(resized_image)
with tf.Session(graph=graph) as sess:
batch_seg_map = sess.run(
OUTPUT_TENSOR_NAME,
feed_dict={INPUT_TENSOR_NAME: [input]})
seg_map = batch_seg_map[0]
tfp = TensorFlowPersistor(base_dir=base_dir, save_dir="deeplab_mobilenetv2_dm05_coco_voc_trainaug")
input4d = np.reshape(input, [1, input.shape[0], input.shape[1], input.shape[2]])
tfp._save_input(input4d, "graph/ImageTensor") #TF is weird here: placeholder is [1, -1, -1, 3] but it adds extra dimension if you pass 4d in :/
tfp._save_predictions({"graph/SemanticPredictions":seg_map})
#Save type info
dtype_dict = {}
dtype_dict[INPUT_TENSOR_NAME] = str(input.dtype)
dtype_dict["graph/SemanticPredictions"] = str(seg_map.dtype)
tfp._save_node_dtypes(dtype_dict)
print(seg_map)
