Print time cost with more precision (PaddlePaddle#3032)

* update gan model * update gan model time display * update gan script * update gan script
nuaaceieyty · Aug 7, 2019 · 6db35e7 · 6db35e7
1 parent 64dc246
commit 6db35e7
Show file tree

Hide file tree

Showing 9 changed files with 66 additions and 46 deletions.
diff --git a/PaddleCV/PaddleGAN/cycle_gan/train.py b/PaddleCV/PaddleGAN/cycle_gan/train.py
@@ -9,6 +9,7 @@ def set_paddle_flags(flags):
         if os.environ.get(key, None) is None:
             os.environ[key] = str(value)
 
+
 use_cudnn_deterministic = os.environ.get('FLAGS_cudnn_deterministic', None)
 
 if use_cudnn_deterministic:
@@ -22,7 +23,7 @@ def set_paddle_flags(flags):
 set_paddle_flags({
     'FLAGS_cudnn_exhaustive_search': use_cudnn_exhaustive_search,
     'FLAGS_conv_workspace_size_limit': 256,
-    'FLAGS_eager_delete_tensor_gb': 0, # enable gc 
+    'FLAGS_eager_delete_tensor_gb': 0,  # enable gc 
     # You can omit the following settings, because the default
     # value of FLAGS_memory_fraction_of_eager_deletion is 1,
     # and default value of FLAGS_fast_eager_deletion_mode is 1 
@@ -247,7 +248,7 @@ def init_model():
             t_time += batch_time
             print(
                 "epoch{}; batch{}; g_A_loss: {}; d_B_loss: {}; g_B_loss: {}; d_A_loss: {}; "
-                "Batch_time_cost: {:.2f}".format(epoch, batch_id, g_A_loss[
+                "Batch_time_cost: {}".format(epoch, batch_id, g_A_loss[
                     0], d_B_loss[0], g_B_loss[0], d_A_loss[0], batch_time))
             losses[0].append(g_A_loss[0])
             losses[1].append(d_A_loss[0])

diff --git a/PaddleCV/PaddleGAN/trainer/AttGAN.py b/PaddleCV/PaddleGAN/trainer/AttGAN.py
@@ -337,10 +337,9 @@ def build_model(self):
                     t_time += batch_time
                     print("epoch{}: batch{}:  \n\
                          d_loss: {}; d_loss_real: {}; d_loss_fake: {}; d_loss_cls: {}; d_loss_gp: {} \n\
-                         Batch_time_cost: {:.2f}"
-                          .format(epoch_id, batch_id, d_loss[0], d_loss_real[
-                              0], d_loss_fake[0], d_loss_cls[0], d_loss_gp[0],
-                                  batch_time))
+                         Batch_time_cost: {}".format(epoch_id, batch_id, d_loss[
+                        0], d_loss_real[0], d_loss_fake[0], d_loss_cls[0],
+                                                     d_loss_gp[0], batch_time))
                 # optimize the generator network
                 else:
                     d_fetches = [

diff --git a/PaddleCV/PaddleGAN/trainer/CGAN.py b/PaddleCV/PaddleGAN/trainer/CGAN.py
@@ -188,15 +188,15 @@ def build_model(self):
                         g_trainer.infer_program,
                         feed={'noise': const_n,
                               'condition': condition_data},
-                        fetch_list={g_trainer.fake})[0]
+                        fetch_list=[g_trainer.fake])[0]
 
                     generate_image_reshape = np.reshape(generate_const_image, (
                         self.cfg.batch_size, -1))
                     total_images = np.concatenate(
                         [real_image, generate_image_reshape])
                     fig = utility.plot(total_images)
                     print(
-                        'Epoch ID={} Batch ID={} D_loss={} G_loss={} Batch_time_cost={:.2f}'.
+                        'Epoch ID: {} Batch ID: {} D_loss: {} G_loss: {} Batch_time_cost: {}'.
                         format(epoch_id, batch_id, d_loss[0], g_loss[0],
                                batch_time))
                     plt.title('Epoch ID={}, Batch ID={}'.format(epoch_id,

diff --git a/PaddleCV/PaddleGAN/trainer/CycleGAN.py b/PaddleCV/PaddleGAN/trainer/CycleGAN.py
@@ -319,7 +319,7 @@ def build_model(self):
                     print("epoch{}: batch{}: \n\
                          d_A_loss: {}; g_A_loss: {}; g_A_cyc_loss: {}; g_A_idt_loss: {}; \n\
                          d_B_loss: {}; g_B_loss: {}; g_B_cyc_loss: {}; g_B_idt_loss: {}; \n\
-                         Batch_time_cost: {:.2f}".format(
+                         Batch_time_cost: {}".format(
                         epoch_id, batch_id, d_A_loss[0], g_A_loss[0],
                         g_A_cyc_loss[0], g_A_idt_loss[0], d_B_loss[0], g_B_loss[
                             0], g_B_cyc_loss[0], g_B_idt_loss[0], batch_time))

diff --git a/PaddleCV/PaddleGAN/trainer/DCGAN.py b/PaddleCV/PaddleGAN/trainer/DCGAN.py
@@ -172,15 +172,16 @@ def build_model(self):
                     generate_const_image = exe.run(
                         g_trainer.infer_program,
                         feed={'noise': const_n},
-                        fetch_list={g_trainer.fake})[0]
+                        fetch_list=[g_trainer.fake])[0]
 
                     generate_image_reshape = np.reshape(generate_const_image, (
                         self.cfg.batch_size, -1))
                     total_images = np.concatenate(
                         [real_image, generate_image_reshape])
                     fig = utility.plot(total_images)
+
                     print(
-                        'Epoch ID={} Batch ID={} D_loss={} G_loss={} Batch_time_cost={:.2f}'.
+                        'Epoch ID: {} Batch ID: {} D_loss: {} G_loss: {} Batch_time_cost: {}'.
                         format(epoch_id, batch_id, d_loss[0], g_loss[0],
                                batch_time))
                     plt.title('Epoch ID={}, Batch ID={}'.format(epoch_id,

diff --git a/PaddleCV/PaddleGAN/trainer/Pix2pix.py b/PaddleCV/PaddleGAN/trainer/Pix2pix.py
@@ -277,7 +277,7 @@ def build_model(self):
                     print("epoch{}: batch{}: \n\
                          g_loss_gan: {}; g_loss_l1: {}; \n\
                          d_loss_real: {}; d_loss_fake: {}; \n\
-                         Batch_time_cost: {:.2f}"
+                         Batch_time_cost: {}"
                           .format(epoch_id, batch_id, g_loss_gan[0], g_loss_l1[
                               0], d_loss_real[0], d_loss_fake[0], batch_time))
 

diff --git a/PaddleCV/PaddleGAN/trainer/STGAN.py b/PaddleCV/PaddleGAN/trainer/STGAN.py
@@ -342,10 +342,9 @@ def build_model(self):
                     t_time += batch_time
                     print("epoch{}: batch{}:  \n\
                          d_loss: {}; d_loss_real: {}; d_loss_fake: {}; d_loss_cls: {}; d_loss_gp: {} \n\
-                         Batch_time_cost: {:.2f}"
-                          .format(epoch_id, batch_id, d_loss[0], d_loss_real[
-                              0], d_loss_fake[0], d_loss_cls[0], d_loss_gp[0],
-                                  batch_time))
+                         Batch_time_cost: {}".format(epoch_id, batch_id, d_loss[
+                        0], d_loss_real[0], d_loss_fake[0], d_loss_cls[0],
+                                                     d_loss_gp[0], batch_time))
                 # optimize the generator network
                 else:
                     d_fetches = [

diff --git a/PaddleCV/PaddleGAN/trainer/StarGAN.py b/PaddleCV/PaddleGAN/trainer/StarGAN.py
@@ -360,7 +360,7 @@ def build_model(self):
                 if batch_id % self.cfg.print_freq == 0:
                     print("epoch{}: batch{}: \n\
                          d_loss_real: {}; d_loss_fake: {}; d_loss_cls: {}; d_loss_gp: {} \n\
-                         Batch_time_cost: {:.2f}".format(
+                         Batch_time_cost: {}".format(
                         epoch_id, batch_id, d_loss_real[0], d_loss_fake[
                             0], d_loss_cls[0], d_loss_gp[0], batch_time))
 

diff --git a/dygraph/cycle_gan/train.py b/dygraph/cycle_gan/train.py
@@ -41,26 +41,25 @@
 add_arg('save_checkpoints',  bool,  True,       "Whether to save checkpoints.")
 # yapf: enable
 
-
 lambda_A = 10.0
 lambda_B = 10.0
 lambda_identity = 0.5
 tep_per_epoch = 2974
 
+
 def optimizer_setting():
-    lr=0.0002
+    lr = 0.0002
     optimizer = fluid.optimizer.Adam(
         learning_rate=fluid.layers.piecewise_decay(
             boundaries=[
                 100 * step_per_epoch, 120 * step_per_epoch,
-                140 * step_per_epoch, 160 * step_per_epoch,
-                180 * step_per_epoch
+                140 * step_per_epoch, 160 * step_per_epoch, 180 * step_per_epoch
             ],
-            values=[
-                lr , lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1
-            ]),
-        beta1=0.5)    
+            values=[lr, lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1]),
+        beta1=0.5)
     return optimizer
+
+
 def train(args):
     with fluid.dygraph.guard():
         max_images_num = data_reader.max_images_num()
@@ -85,7 +84,7 @@ def train(args):
         A_test_reader = data_reader.a_test_reader()
         B_test_reader = data_reader.b_test_reader()
 
-        cycle_gan = Cycle_Gan("cycle_gan",istrain=True)
+        cycle_gan = Cycle_Gan("cycle_gan", istrain=True)
 
         losses = [[], []]
         t_time = 0
@@ -101,75 +100,94 @@ def train(args):
                 data_B = next(B_reader)
 
                 s_time = time.time()
-                data_A = np.array([data_A[0].reshape(3,256,256)]).astype("float32")
-                data_B = np.array([data_B[0].reshape(3,256,256)]).astype("float32")
+                data_A = np.array(
+                    [data_A[0].reshape(3, 256, 256)]).astype("float32")
+                data_B = np.array(
+                    [data_B[0].reshape(3, 256, 256)]).astype("float32")
                 data_A = to_variable(data_A)
                 data_B = to_variable(data_B)
 
                 # optimize the g_A network
-                fake_A,fake_B,cyc_A,cyc_B,g_A_loss,g_B_loss,idt_loss_A,idt_loss_B,cyc_A_loss,cyc_B_loss,g_loss = cycle_gan(data_A,data_B,True,False,False)
+                fake_A, fake_B, cyc_A, cyc_B, g_A_loss, g_B_loss, idt_loss_A, idt_loss_B, cyc_A_loss, cyc_B_loss, g_loss = cycle_gan(
+                    data_A, data_B, True, False, False)
 
                 g_loss_out = g_loss.numpy()
 
                 g_loss.backward()
                 vars_G = []
                 for param in cycle_gan.parameters():
-                    if param.name[:52]=="cycle_gan/Cycle_Gan_0/build_generator_resnet_9blocks":    
+                    if param.name[:
+                                  52] == "cycle_gan/Cycle_Gan_0/build_generator_resnet_9blocks":
                         vars_G.append(param)
 
-                optimizer1.minimize(g_loss,parameter_list=vars_G)                
+                optimizer1.minimize(g_loss, parameter_list=vars_G)
                 cycle_gan.clear_gradients()
 
-
                 fake_pool_B = B_pool.pool_image(fake_B).numpy()
-                fake_pool_B = np.array([fake_pool_B[0].reshape(3,256,256)]).astype("float32")
+                fake_pool_B = np.array(
+                    [fake_pool_B[0].reshape(3, 256, 256)]).astype("float32")
                 fake_pool_B = to_variable(fake_pool_B)
 
                 fake_pool_A = A_pool.pool_image(fake_A).numpy()
-                fake_pool_A = np.array([fake_pool_A[0].reshape(3,256,256)]).astype("float32")
+                fake_pool_A = np.array(
+                    [fake_pool_A[0].reshape(3, 256, 256)]).astype("float32")
                 fake_pool_A = to_variable(fake_pool_A)
 
                 # optimize the d_A network
-                rec_B, fake_pool_rec_B = cycle_gan(data_B,fake_pool_B,False,True,False)
+                rec_B, fake_pool_rec_B = cycle_gan(data_B, fake_pool_B, False,
+                                                   True, False)
                 d_loss_A = (fluid.layers.square(fake_pool_rec_B) +
-                    fluid.layers.square(rec_B - 1)) / 2.0
+                            fluid.layers.square(rec_B - 1)) / 2.0
                 d_loss_A = fluid.layers.reduce_mean(d_loss_A)
 
                 d_loss_A.backward()
                 vars_da = []
                 for param in cycle_gan.parameters():
-                    if param.name[:47]=="cycle_gan/Cycle_Gan_0/build_gen_discriminator_0":
+                    if param.name[:
+                                  47] == "cycle_gan/Cycle_Gan_0/build_gen_discriminator_0":
                         vars_da.append(param)
-                optimizer2.minimize(d_loss_A,parameter_list=vars_da)
+                optimizer2.minimize(d_loss_A, parameter_list=vars_da)
                 cycle_gan.clear_gradients()
 
                 # optimize the d_B network
 
-                rec_A, fake_pool_rec_A = cycle_gan(data_A,fake_pool_A,False,False,True)
+                rec_A, fake_pool_rec_A = cycle_gan(data_A, fake_pool_A, False,
+                                                   False, True)
                 d_loss_B = (fluid.layers.square(fake_pool_rec_A) +
-                    fluid.layers.square(rec_A - 1)) / 2.0
+                            fluid.layers.square(rec_A - 1)) / 2.0
                 d_loss_B = fluid.layers.reduce_mean(d_loss_B)
 
                 d_loss_B.backward()
                 vars_db = []
                 for param in cycle_gan.parameters():
-                    if param.name[:47]=="cycle_gan/Cycle_Gan_0/build_gen_discriminator_1":
+                    if param.name[:
+                                  47] == "cycle_gan/Cycle_Gan_0/build_gen_discriminator_1":
                         vars_db.append(param)
-                optimizer3.minimize(d_loss_B,parameter_list=vars_db)
+                optimizer3.minimize(d_loss_B, parameter_list=vars_db)
 
                 cycle_gan.clear_gradients()
 
                 batch_time = time.time() - s_time
                 t_time += batch_time
                 print(
-                    "epoch{}; batch{}; g_loss:{}; d_A_loss: {}; d_B_loss:{} ; \n g_A_loss: {}; g_A_cyc_loss: {}; g_A_idt_loss: {}; g_B_loss: {}; g_B_cyc_loss:  {}; g_B_idt_loss: {};Batch_time_cost: {:.2f}".format(epoch, batch_id,g_loss_out[0],d_loss_A.numpy()[0], d_loss_B.numpy()[0],g_A_loss.numpy()[0],cyc_A_loss.numpy()[0], idt_loss_A.numpy()[0],  g_B_loss.numpy()[0],cyc_B_loss.numpy()[0],idt_loss_B.numpy()[0], batch_time))
+                    "epoch{}; batch{}; g_loss:{}; d_A_loss: {}; d_B_loss:{} ; \n g_A_loss: {}; g_A_cyc_loss: {}; g_A_idt_loss: {}; g_B_loss: {}; g_B_cyc_loss:  {}; g_B_idt_loss: {};Batch_time_cost: {}".
+                    format(epoch, batch_id, g_loss_out[0],
+                           d_loss_A.numpy()[0],
+                           d_loss_B.numpy()[0],
+                           g_A_loss.numpy()[0],
+                           cyc_A_loss.numpy()[0],
+                           idt_loss_A.numpy()[0],
+                           g_B_loss.numpy()[0],
+                           cyc_B_loss.numpy()[0],
+                           idt_loss_B.numpy()[0], batch_time))
                 with open('logging_train.txt', 'a') as log_file:
                     now = time.strftime("%c")
                     log_file.write(
                     "time: {}; epoch{}; batch{}; d_A_loss: {}; g_A_loss: {}; \
                     g_A_cyc_loss: {}; g_A_idt_loss: {}; d_B_loss: {}; \
                     g_B_loss: {}; g_B_cyc_loss: {}; g_B_idt_loss: {}; \
-                    Batch_time_cost: {:.2f}\n".format(now, epoch, \
+                    Batch_time_cost: {}\n"
+                                          .format(now, epoch, \
                         batch_id, d_loss_A[0], g_A_loss[ 0], cyc_A_loss[0], \
                         idt_loss_A[0], d_loss_B[0], g_A_loss[0], \
                         cyc_B_loss[0], idt_loss_B[0], batch_time))
@@ -184,10 +202,12 @@ def train(args):
                     print("kpis\td_A_loss\t%0.3f" % d_loss_A.numpy()[0])
                     print("kpis\td_B_loss\t%0.3f" % d_loss_B.numpy()[0])
                     break
-
 
             if args.save_checkpoints:
-                fluid.dygraph.save_persistables(cycle_gan.state_dict(),args.output+"/checkpoints/{}".format(epoch))
+                fluid.dygraph.save_persistables(
+                    cycle_gan.state_dict(),
+                    args.output + "/checkpoints/{}".format(epoch))
+
 
 if __name__ == "__main__":
     args = parser.parse_args()