From bb1376db452be0692a5c526317dc48f586e0e3ae Mon Sep 17 00:00:00 2001
From: zhiqiu <chenqiuliang@baidu.com>
Date: Fri, 16 Jul 2021 13:26:03 +0000
Subject: [PATCH] add flags setting

---
 .../ImageNet/ResNet/ResNet50_fp16_dygraph.yaml  | 17 +++++++++++++++--
 ppcls/engine/trainer.py                         |  6 ++++++
 2 files changed, 21 insertions(+), 2 deletions(-)

diff --git a/ppcls/configs/ImageNet/ResNet/ResNet50_fp16_dygraph.yaml b/ppcls/configs/ImageNet/ResNet/ResNet50_fp16_dygraph.yaml
index 7f51d334..59d4bcec 100644
--- a/ppcls/configs/ImageNet/ResNet/ResNet50_fp16_dygraph.yaml
+++ b/ppcls/configs/ImageNet/ResNet/ResNet50_fp16_dygraph.yaml
@@ -10,19 +10,26 @@ Global:
   epochs: 120
   print_batch_step: 10
   use_visualdl: False
+  image_channel: &image_channel 4
   # used for static mode and model export
-  image_shape: [3, 224, 224]
+  image_shape: [*image_channel, 224, 224]
   save_inference_dir: ./inference
   # training model under @to_static
   to_static: False
   use_dali: True
 
+# mixed precision training
 AMP:
+  scale_loss: 128.0
+  use_dynamic_loss_scaling: True
+  use_pure_fp16: &use_pure_fp16 False
 
 # model architecture
 Arch:
   name: ResNet50
   class_num: 1000
+  input_image_channel: *image_channel
+  data_format: "NHWC"
  
 # loss function config for traing/eval process
 Loss:
@@ -67,10 +74,12 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
+            output_fp16: *use_pure_fp16
+            channel_num: *image_channel
 
     sampler:
       name: DistributedBatchSampler
-      batch_size: 64
+      batch_size: 256
       drop_last: False
       shuffle: True
     loader:
@@ -95,6 +104,8 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
+            output_fp16: *use_pure_fp16
+            channel_num: *image_channel
     sampler:
       name: DistributedBatchSampler
       batch_size: 64
@@ -120,6 +131,8 @@ Infer:
         mean: [0.485, 0.456, 0.406]
         std: [0.229, 0.224, 0.225]
         order: ''
+        output_fp16: *use_pure_fp16
+        channel_num: *image_channel
     - ToCHWImage:
   PostProcess:
     name: Topk
diff --git a/ppcls/engine/trainer.py b/ppcls/engine/trainer.py
index af091659..451531c1 100644
--- a/ppcls/engine/trainer.py
+++ b/ppcls/engine/trainer.py
@@ -112,6 +112,12 @@ class Trainer(object):
         else:
             self.scale_loss = 1.0
             self.use_dynamic_loss_scaling = False
+        if self.amp:
+            AMP_RELATED_FLAGS_SETTING = {
+                'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
+                'FLAGS_max_inplace_grad_add': 8,
+            }
+            paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
         self.train_loss_func = None
         self.eval_loss_func = None
         self.train_metric_func = None
-- 
GitLab