fix grad as list (#2493)

* fix grad and scipy version

PaddleCV/gan/data_reader.py

@@ -296,7 +296,7 @@ class celeba_reader_creator(reader_creator):
                         label = np.array(label).astype("float32")
                         label = (label + 1) // 2
                         img = CentorCrop(img, args.crop_size, args.crop_size)
-                        img = img.resize((args.load_size, args.load_size),
+                        img = img.resize((args.image_size, args.image_size),
                                          Image.BILINEAR)
                     img = (np.array(img).astype('float32') / 255.0 - 0.5) / 0.5
                     img = img.transpose([2, 0, 1])

PaddleCV/gan/infer.py

@@ -23,9 +23,10 @@ from PIL import Image
 import paddle.fluid as fluid
 import paddle
 import numpy as np
-from scipy.misc import imsave
+import imageio
 import glob
 from util.config import add_arguments, print_arguments
+import copy
 
 parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
@@ -153,7 +154,7 @@ def infer(args):
                 images.append(fake_temp)
             images_concat = np.concatenate(images, 1)
             images_concat = np.concatenate(images_concat, 1)
-            imsave(args.output + "/fake_img_" + name[0], (
+            imageio.imwrite(args.output + "/fake_img_" + name[0], (
                 (images_concat + 1) * 127.5).astype(np.uint8))
     elif args.model_net == 'StarGAN':
         test_reader = celeba_reader_creator(
@@ -184,8 +185,8 @@ def infer(args):
                 fake_temp = np.squeeze(out[0]).transpose([1, 2, 0])
                 images.append(fake_temp)
             images_concat = np.concatenate(images, 1)
-            imsave(out_path + "/fake_img" + str(epoch) + "_" + name[0], (
-                (images_concat + 1) * 127.5).astype(np.uint8))
+            imageio.imwrite(out_path + "/fake_img" + str(epoch) + "_" + name[0],
+                            ((images_concat + 1) * 127.5).astype(np.uint8))
 
     elif args.model_net == 'Pix2pix' or args.model_net == 'cyclegan':
         for file in glob.glob(args.input):
@@ -204,7 +205,7 @@ def infer(args):
             fake_temp = np.squeeze(fake_temp[0]).transpose([1, 2, 0])
             input_temp = np.squeeze(data).transpose([1, 2, 0])
 
-            imsave(args.output + "/fake_" + image_name, (
+            imageio.imwrite(args.output + "/fake_" + image_name, (
                 (fake_temp + 1) * 127.5).astype(np.uint8))
     else:
         raise NotImplementedError("model_net {} is not support".format(

PaddleCV/gan/trainer/AttGAN.py

@@ -161,20 +161,21 @@ class DTrainer():
             if fluid.io.is_parameter(var) and var.name.startswith(
                     "discriminator"):
                 vars.append(var.name)
-        grad = fluid.gradients(pred, x, no_grad_set=vars)
+        grad = fluid.gradients(pred, x, no_grad_set=vars)[0]
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])
-        epsilon = 1e-5
         norm = fluid.layers.sqrt(
             fluid.layers.reduce_sum(
-                fluid.layers.square(grad), dim=1) + epsilon)
+                fluid.layers.square(grad), dim=1))
         gp = fluid.layers.reduce_mean(fluid.layers.square(norm - 1.0))
         return gp
 
 
 class AttGAN(object):
     def add_special_args(self, parser):
+        parser.add_argument(
+            '--image_size', type=int, default=256, help="image size")
         parser.add_argument(
             '--g_lr',
             type=float,
@@ -241,7 +242,7 @@ class AttGAN(object):
         self.batch_num = batch_num
 
     def build_model(self):
-        data_shape = [-1, 3, self.cfg.load_size, self.cfg.load_size]
+        data_shape = [-1, 3, self.cfg.image_size, self.cfg.image_size]
 
         image_real = fluid.layers.data(
             name='image_real', shape=data_shape, dtype='float32')

PaddleCV/gan/trainer/STGAN.py

@@ -159,20 +159,21 @@ class DTrainer():
             if fluid.io.is_parameter(var) and var.name.startswith(
                     "discriminator"):
                 vars.append(var.name)
-        grad = fluid.gradients(pred, x, no_grad_set=vars)
+        grad = fluid.gradients(pred, x, no_grad_set=vars)[0]
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])
-        epsilon = 1e-5
         norm = fluid.layers.sqrt(
             fluid.layers.reduce_sum(
-                fluid.layers.square(grad), dim=1) + epsilon)
+                fluid.layers.square(grad), dim=1))
         gp = fluid.layers.reduce_mean(fluid.layers.square(norm - 1.0))
         return gp
 
 
 class STGAN(object):
     def add_special_args(self, parser):
+        parser.add_argument(
+            '--image_size', type=int, default=256, help="image size")
         parser.add_argument(
             '--g_lr',
             type=float,
@@ -246,7 +247,7 @@ class STGAN(object):
         self.batch_num = batch_num
 
     def build_model(self):
-        data_shape = [-1, 3, self.cfg.load_size, self.cfg.load_size]
+        data_shape = [-1, 3, self.cfg.image_size, self.cfg.image_size]
 
         image_real = fluid.layers.data(
             name='image_real', shape=data_shape, dtype='float32')

PaddleCV/gan/trainer/StarGAN.py

@@ -185,7 +185,7 @@ class DTrainer():
         for var in fluid.default_main_program().list_vars():
             if fluid.io.is_parameter(var) and var.name.startswith('d_'):
                 vars.append(var.name)
-        grad = fluid.gradients(pred, x, no_grad_set=vars)
+        grad = fluid.gradients(pred, x, no_grad_set=vars)[0]
         grad_shape = grad.shape
         grad = fluid.layers.reshape(
             grad, [-1, grad_shape[1] * grad_shape[2] * grad_shape[3]])

PaddleCV/gan/util/utility.py

@@ -27,7 +27,8 @@ import six
 matplotlib.use('agg')
 import matplotlib.pyplot as plt
 import matplotlib.gridspec as gridspec
-from scipy.misc import imsave
+import imageio
+import copy
 
 img_dim = 28
 
@@ -94,11 +95,11 @@ def save_test_image(epoch,
             input_A_temp = np.squeeze(data_A[0]).transpose([1, 2, 0])
             input_B_temp = np.squeeze(data_A[0]).transpose([1, 2, 0])
 
-            imsave(out_path + "/fakeB_" + str(epoch) + "_" + name, (
+            imageio.imwrite(out_path + "/fakeB_" + str(epoch) + "_" + name, (
                 (fake_B_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/inputA_" + str(epoch) + "_" + name, (
+            imageio.imwrite(out_path + "/inputA_" + str(epoch) + "_" + name, (
                 (input_A_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/inputB_" + str(epoch) + "_" + name, (
+            imageio.imwrite(out_path + "/inputB_" + str(epoch) + "_" + name, (
                 (input_B_temp + 1) * 127.5).astype(np.uint8))
     elif cfg.model_net == "StarGAN":
         for data in zip(A_test_reader()):
@@ -127,8 +128,8 @@ def save_test_image(epoch,
                 images.append(fake_temp)
                 images.append(rec_temp)
             images_concat = np.concatenate(images, 1)
-            imsave(out_path + "/fake_img" + str(epoch) + "_" + name[0], (
-                (images_concat + 1) * 127.5).astype(np.uint8))
+            imageio.imwrite(out_path + "/fake_img" + str(epoch) + "_" + name[0],
+                            ((images_concat + 1) * 127.5).astype(np.uint8))
     elif cfg.model_net == 'AttGAN' or cfg.model_net == 'STGAN':
         for data in zip(A_test_reader()):
             real_img, label_org, name = data[0]
@@ -168,8 +169,8 @@ def save_test_image(epoch,
                 images.append(fake_temp)
             images_concat = np.concatenate(images, 1)
             images_concat = np.concatenate(images_concat, 1)
-            imsave(out_path + "/fake_img" + str(epoch) + '_' + name[0], (
-                (images_concat + 1) * 127.5).astype(np.uint8))
+            imageio.imwrite(out_path + "/fake_img" + str(epoch) + '_' + name[0],
+                            ((images_concat + 1) * 127.5).astype(np.uint8))
 
     else:
         for data_A, data_B in zip(A_test_reader(), B_test_reader()):
@@ -194,17 +195,17 @@ def save_test_image(epoch,
             input_A_temp = np.squeeze(data_A[0][0]).transpose([1, 2, 0])
             input_B_temp = np.squeeze(data_B[0][0]).transpose([1, 2, 0])
 
-            imsave(out_path + "/fakeB_" + str(epoch) + "_" + A_name, (
+            imageio.imwrite(out_path + "/fakeB_" + str(epoch) + "_" + A_name, (
                 (fake_B_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/fakeA_" + str(epoch) + "_" + B_name, (
+            imageio.imwrite(out_path + "/fakeA_" + str(epoch) + "_" + B_name, (
                 (fake_A_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/cycA_" + str(epoch) + "_" + A_name, (
+            imageio.imwrite(out_path + "/cycA_" + str(epoch) + "_" + A_name, (
                 (cyc_A_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/cycB_" + str(epoch) + "_" + B_name, (
+            imageio.imwrite(out_path + "/cycB_" + str(epoch) + "_" + B_name, (
                 (cyc_B_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/inputA_" + str(epoch) + "_" + A_name, (
+            imageio.imwrite(out_path + "/inputA_" + str(epoch) + "_" + A_name, (
                 (input_A_temp + 1) * 127.5).astype(np.uint8))
-            imsave(out_path + "/inputB_" + str(epoch) + "_" + B_name, (
+            imageio.imwrite(out_path + "/inputB_" + str(epoch) + "_" + B_name, (
                 (input_B_temp + 1) * 127.5).astype(np.uint8))