Testing compat Tensorflow 2 upgrade

data_prepare/detect_3D_landmark.py

@@ -346,15 +346,15 @@ def detect_lmk86(origin_images_dir, mtcnn_dir, out_dir, names_list, pb_path):
     fopen = open(out_dir, "w")
 
     with tf.Graph().as_default():
-        graph_def = tf.GraphDef()
+        graph_def = tf.compat.v1.GraphDef()
         graph_file = pb_path
         with open(graph_file, "rb") as f:
             print("hello")
             graph_def.ParseFromString(f.read())
             tf.import_graph_def(graph_def, name="")
 
-        with tf.Session() as sess:
-            tf.global_variables_initializer().run()
+        with tf.compat.v1.Session() as sess:
+            tf.compat.v1.global_variables_initializer().run()
             image = sess.graph.get_tensor_by_name("lmk86pt_input:0")
             predict_lanmark = sess.graph.get_tensor_by_name("lmk86pt_output:0")
 

data_prepare/detect_face_with_mtcnn.py

@@ -346,16 +346,16 @@ def detect_with_MTCNN(origin_images_dir, out_dir, pb_path, mode="no_depth"):
     factor = 0.709  # scale factor
 
     with tf.Graph().as_default():
-        graph_def = tf.GraphDef()
+        graph_def = tf.compat.v1.GraphDef()
         graph_file = pb_path
         with open(graph_file, "rb") as f:
             print("hello")
             graph_def.ParseFromString(f.read())
             tf.import_graph_def(graph_def, name="")
 
-        sess = tf.Session()
+        sess = tf.compat.v1.Session()
         with sess.as_default():
-            tf.global_variables_initializer().run()
+            tf.compat.v1.global_variables_initializer().run()
             pnet, rnet, onet = create_mtcnn_pb(sess)
 
     # find files

data_prepare/run_data_preparation.py

@@ -60,17 +60,17 @@
 
 
 def detect_2Dlmk_all_imgs(graph_file, img_dir, lmk3D_txt_path, lmk2D_txt_path):
-    with tf.Graph().as_default():
-        graph_def = tf.GraphDef()
+    with tf.Graph().as_default(), tf.device('/device:XLA_CPU:0'):
+        graph_def = tf.compat.v1.GraphDef()
         graph_file = graph_file
 
         with open(graph_file, "rb") as f:
             print("hello")
             graph_def.ParseFromString(f.read())
             tf.import_graph_def(graph_def, name="")
 
-        with tf.Session() as sess:
-            tf.initialize_all_variables().run()
+        with tf.compat.v1.Session() as sess:
+            tf.compat.v1.initialize_all_variables().run()
 
             fopen = open(lmk2D_txt_path, "w")
 
@@ -155,17 +155,17 @@ def face_seg(graph_file, lmk3D_crop_txt_path, out_crop_dir, seg_dir):
 
     landmarks3D, images_name = load_landmark(lmk3D_crop_txt_path, 86)
 
-    with tf.Graph().as_default():
-        graph_def = tf.GraphDef()
+    with tf.Graph().as_default(), tf.device('/device:XLA_CPU:0'):
+        graph_def = tf.compat.v1.GraphDef()
         graph_file = graph_file
 
         with open(graph_file, "rb") as f:
             print("hello")
             graph_def.ParseFromString(f.read())
             tf.import_graph_def(graph_def, name="")
 
-        with tf.Session() as sess:
-            tf.initialize_all_variables().run()
+        with tf.compat.v1.Session() as sess:
+            tf.compat.v1.initialize_all_variables().run()
             count = 0
             for i in range(0, len(images_name)):
                 img_name = images_name[i]
@@ -204,14 +204,14 @@ def prepare_test_data_RGB(img_dir, out_dir):
     names_list = detect_face_with_mtcnn.detect_with_MTCNN(img_dir, mtcnn_dir, pb_path)
 
     print("start detect 86pt 3D lmk")
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "lmk3D_86_model.pb")
     detect_3D_landmark.detect_lmk86(
         img_dir, mtcnn_dir, lmk3D_ori_txt_path, names_list, pb_path
     )
 
     print("start detect 68pt 2D lmk")  # need to transfer RGB in the function
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "lmk2D_68_model.pb")
     detect_2Dlmk_all_imgs(
         pb_path, img_dir, lmk3D_ori_txt_path, lmk2D_ori_txt_path
@@ -229,7 +229,7 @@ def prepare_test_data_RGB(img_dir, out_dir):
     )
 
     print("start face seg")  # need to transfer RGB in the function
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "faceseg_model.pb")
     face_seg(pb_path, lmk3D_crop_txt_path, crop_dir, seg_dir)
 
@@ -277,14 +277,14 @@ def prepare_test_data_RGBD(img_dir, out_dir):
     )
 
     print("start detect 86pt 3D lmk")
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "lmk3D_86_model.pb")
     detect_3D_landmark.detect_lmk86(
         img_dir, mtcnn_dir, lmk3D_ori_txt_path, names_list, pb_path
     )
 
     print("start detect 68pt 2D lmk")  # need to transfer RGB in the function
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "lmk2D_68_model.pb")
     detect_2Dlmk_all_imgs(
         pb_path, img_dir, lmk3D_ori_txt_path, lmk2D_ori_txt_path
@@ -302,7 +302,7 @@ def prepare_test_data_RGBD(img_dir, out_dir):
     )
 
     print("start face seg")  # need to transfer RGB in the function
-    tf.reset_default_graph()
+    tf.compat.v1.reset_default_graph()
     pb_path = os.path.join(FLAGS.pb_path, "faceseg_model.pb")
     face_seg(pb_path, lmk3D_crop_txt_path, crop_dir, seg_dir)
 

optimization/rgb/run_RGB_opt.py

@@ -32,10 +32,13 @@
 import scipy.io as scio
 import cv2
 import tensorflow as tf
-import tensorflow.contrib.opt as tf_opt
+#import tensorflow.contrib.opt as tf_opt
 import os
 from absl import app, flags
 import sys
+tf.compat.v1.disable_eager_execution()
+tf.compat.v1.disable_v2_behavior()
+
 
 sys.path.append("../..")
 
@@ -51,67 +54,67 @@
 def define_variable(num_of_img, imageH, imageW, para_shape_shape, para_tex_shape, info):
 
     # variable-trainable=False
-    image_batch = tf.get_variable(
+    image_batch = tf.compat.v1.get_variable(
         shape=[num_of_img, imageH, imageW, 3],
         dtype=tf.float32,
         name="ori_img",
         trainable=False,
-        initializer=tf.constant_initializer(info["img_list"] / 255.0),
+        initializer=tf.compat.v1.constant_initializer(info["img_list"] / 255.0),
     )
 
-    segmentation = tf.get_variable(
+    segmentation = tf.compat.v1.get_variable(
         shape=[num_of_img, imageH, imageW, 19],
         dtype=tf.float32,
         name="face_segmentation",
         trainable=False,
-        initializer=tf.constant_initializer(info["seg_list"]),
+        initializer=tf.compat.v1.constant_initializer(info["seg_list"]),
     )
 
-    lmk_86_3d_batch = tf.get_variable(
+    lmk_86_3d_batch = tf.compat.v1.get_variable(
         shape=[num_of_img, 86, 2],
         dtype=tf.float32,
         name="lmk_86_3d_batch",
         trainable=False,
-        initializer=tf.constant_initializer(info["lmk_list3D"]),
+        initializer=tf.compat.v1.constant_initializer(info["lmk_list3D"]),
     )
 
-    lmk_68_2d_batch = tf.get_variable(
+    lmk_68_2d_batch = tf.compat.v1.get_variable(
         shape=[num_of_img, 68, 2],
         dtype=tf.float32,
         name="lmk_68_2d_batch",
         trainable=False,
-        initializer=tf.constant_initializer(info["lmk_list2D"]),
+        initializer=tf.compat.v1.constant_initializer(info["lmk_list2D"]),
     )
 
-    K = tf.get_variable(
+    K = tf.compat.v1.get_variable(
         shape=[1, 3, 3],
         dtype=tf.float32,
         name="K",
         trainable=False,
-        initializer=tf.constant_initializer(info["K"]),
+        initializer=tf.compat.v1.constant_initializer(info["K"]),
     )
 
     # variable-trainable=True
-    pose6 = tf.get_variable(
+    pose6 = tf.compat.v1.get_variable(
         shape=[num_of_img, 6, 1],
         dtype=tf.float32,
         name="para_pose6",
         trainable=True,
-        initializer=tf.constant_initializer(info["se3_list"]),
+        initializer=tf.compat.v1.constant_initializer(info["se3_list"]),
     )
 
-    para_shape = tf.get_variable(
+    para_shape = tf.compat.v1.get_variable(
         shape=[1, para_shape_shape], dtype=tf.float32, name="para_shape", trainable=True
     )
 
-    para_tex = tf.get_variable(
+    para_tex = tf.compat.v1.get_variable(
         shape=[1, para_tex_shape], dtype=tf.float32, name="para_tex", trainable=True
     )
 
-    para_illum = tf.get_variable(
+    para_illum = tf.compat.v1.get_variable(
         shape=[num_of_img, 27],
         dtype=tf.float32,
-        initializer=tf.zeros_initializer(),
+        initializer=tf.compat.v1.zeros_initializer(),
         name="para_illum",
         trainable=True,
     )
@@ -164,12 +167,12 @@ def build_RGB_opt_graph(var_list, basis3dmm, imageH, imageW):
     # optimizer
     global_step = tf.Variable(0, name="global_step_train", trainable=False)
     learning_rate = tf.maximum(
-        tf.train.exponential_decay(
+        tf.compat.v1.train.exponential_decay(
             FLAGS.learning_rate, global_step, FLAGS.lr_decay_step, FLAGS.lr_decay_rate
         ),
         FLAGS.min_learning_rate,
     )
-    optim = tf.train.AdamOptimizer(learning_rate=learning_rate)
+    optim = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)
 
     gvs_illum = optim.compute_gradients(tot_loss_illum)
     gvs = optim.compute_gradients(tot_loss)
@@ -241,20 +244,20 @@ def RGB_opt(_):
     out_list = build_RGB_opt_graph(var_list, basis3dmm, imageH, imageW)
 
     # summary_op
-    summary_op = tf.summary.merge_all()
-    summary_writer = tf.summary.FileWriter(FLAGS.summary_dir)
+    summary_op = tf.compat.v1.summary.merge_all()
+    summary_writer = tf.compat.v1.summary.FileWriter(FLAGS.summary_dir)
 
     if os.path.exists(FLAGS.summary_dir) is False:
         os.makedirs(FLAGS.summary_dir)
     if os.path.exists(FLAGS.out_dir) is False:
         os.makedirs(FLAGS.out_dir)
 
     # start opt
-    config = tf.ConfigProto()
+    config = tf.compat.v1.ConfigProto()
     # config.gpu_options.per_process_gpu_memory_fraction=0.5
     config.gpu_options.allow_growth = True
-    with tf.Session(config=config) as sess:
-        sess.run(tf.global_variables_initializer())
+    with tf.compat.v1.Session(config=config) as sess:
+        sess.run(tf.compat.v1.global_variables_initializer())
 
         import time
 

optimization/rgbd/RGBD_utils/PoseTools_TF.py

@@ -106,13 +106,13 @@ def tf_Rodrigues(a, b, c, theta):
 
             rr0 = (
                 cos_theta * I
-                + (1 - cos_theta) * (n * tf.transpose(n))
+                + (1 - cos_theta) * (n * tf.transpose(a=n))
                 + sin_theta * n_hat
             )
             return rr0
 
         rr = tf.cond(
-            tf.squeeze(theta) > 1e-3,
+            pred=tf.squeeze(theta) > 1e-3,
             true_fn=lambda: tf_Rodrigues(a, b, c, theta),
             false_fn=lambda: tf.eye(3, 3),
         )
@@ -125,7 +125,7 @@ def tf_trans_inverse(rr, tt):
         # in/out : n* 3 * 4
         # rr = TWC[0:3, 0:3]
         # tt = TWC[0:3, 3:]
-        rr_1 = tf.transpose(rr, [1, 0])
+        rr_1 = tf.transpose(a=rr, perm=[1, 0])
         tt_1 = -1 * tf.matmul(rr_1, tt)
         return rr_1, tt_1
 

optimization/rgbd/step4A_prefit_Albedo_Global.py

@@ -37,6 +37,8 @@
 from PIL import Image
 import sys
 sys.path.append('../..')
+tf.compat.v1.disable_eager_execution()
+tf.compat.v1.disable_v2_behavior()
 
 from utils.basis import load_3dmm_basis, get_region_uv_texture, construct
 from utils.misc import tf_blend_uv, blend_uv
@@ -124,14 +126,14 @@ def main(_):
     imageH = img_list[0].shape[0]
     imageW = img_list[0].shape[1]
     assert(img_list[0].shape[0] == img_list[1].shape[0] and img_list[0].shape[1] == img_list[2].shape[1])
-    image_mid_batch = tf.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_mid')
-    image_left_batch = tf.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_left')
-    image_right_batch = tf.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_right')
+    image_mid_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_mid')
+    image_left_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_left')
+    image_right_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, imageH, imageW, 3], name='image_right')
 
     NV = basis3dmm['basis_shape'].shape[1] // 3
-    proj_xyz_mid_batch = tf.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_mid')
-    proj_xyz_left_batch = tf.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_left')
-    proj_xyz_right_batch = tf.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_right')
+    proj_xyz_mid_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_mid')
+    proj_xyz_left_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_left')
+    proj_xyz_right_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, NV, 3], name='proj_xyz_right')
 
     ver_normals_mid_batch, _ = Projector.get_ver_norm(proj_xyz_mid_batch, basis3dmm['tri'], 'normal_mid')
     ver_normals_left_batch, _ = Projector.get_ver_norm(proj_xyz_left_batch, basis3dmm['tri'], 'normal_left')
@@ -179,17 +181,17 @@ def main(_):
     #------------------------------------------------------------------------------------------
     # build fitting graph
     uv_bases = basis3dmm['uv']
-    para_tex = tf.get_variable(
+    para_tex = tf.compat.v1.get_variable(
             shape=[1,uv_bases['basis'].shape[0]],
-            initializer=tf.zeros_initializer(),
+            initializer=tf.compat.v1.zeros_initializer(),
             name='para_tex'
             )
 
     uv_rgb, uv_mask = get_region_uv_texture(uv_bases, para_tex)
     print("uv_rgb: ", uv_rgb.shape )
 
     # build fitting loss
-    input_uv512_batch = tf.placeholder(dtype=tf.float32, shape=[1, 512, 512, 3], name='gt_uv')
+    input_uv512_batch = tf.compat.v1.placeholder(dtype=tf.float32, shape=[1, 512, 512, 3], name='gt_uv')
     tot_loss = 0.
     loss_str = 'total:{}'
     if FLAGS.photo_weight > 0:
@@ -206,16 +208,16 @@ def main(_):
         uv_reg_tex_loss = Losses.reg_loss(para_tex)
         tot_loss = tot_loss + uv_reg_tex_loss * FLAGS.uv_reg_tex_weight
         loss_str = loss_str + '; reg:{}'
-    optim = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)
+    optim = tf.compat.v1.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)
     train_op = optim.minimize(tot_loss)
 
 
-    with tf.Session() as sess:
+    with tf.compat.v1.Session() as sess:
 
         if FLAGS.write_graph:
-            tf.train.write_graph(sess.graph_def, '', FLAGS.pb_path, as_text=True)
+            tf.io.write_graph(sess.graph_def, '', FLAGS.pb_path, as_text=True)
             exit()
-        sess.run(tf.global_variables_initializer())
+        sess.run(tf.compat.v1.global_variables_initializer())
 
         start_time = time.time()
         uv_extract,   o_uv_left_batch,  o_uv_mid_batch, o_uv_right_batch,   o_mask_right_batch= \
@@ -305,4 +307,4 @@ def main(_):
     flags.DEFINE_float('learning_rate', 0.1, 'string : path for 3dmm')
     flags.DEFINE_string('GPU_NO', '7', 'which GPU')
 
-    tf.app.run(main)
+    tf.compat.v1.app.run(main)