From b761325faaefcffb96d6f99bcf8b13005e98d19b Mon Sep 17 00:00:00 2001
From: gaotingquan <gaotingquan@baidu.com>
Date: Sat, 2 Apr 2022 07:35:01 +0000
Subject: [PATCH] fix: fp32 eval by default when enable amp

If you want to eval by fp16 when enable amp, please set Amp.use_fp16_test=True, False by default.
---
 ppcls/engine/evaluation/classification.py | 2 +-
 ppcls/static/train.py                     | 7 ++++---
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/ppcls/engine/evaluation/classification.py b/ppcls/engine/evaluation/classification.py
index 994eeb5e..e9836fcb 100644
--- a/ppcls/engine/evaluation/classification.py
+++ b/ppcls/engine/evaluation/classification.py
@@ -58,7 +58,7 @@ def classification_eval(engine, epoch_id=0):
             batch[1] = batch[1].reshape([-1, 1]).astype("int64")
 
         # image input
-        if engine.amp:
+        if engine.amp and engine.config["AMP"].get("use_fp16_test", False):
             amp_level = engine.config['AMP'].get("level", "O1").upper()
             with paddle.amp.auto_cast(
                     custom_black_list={
diff --git a/ppcls/static/train.py b/ppcls/static/train.py
index dd16cdb4..45c93762 100644
--- a/ppcls/static/train.py
+++ b/ppcls/static/train.py
@@ -161,12 +161,13 @@ def main(args):
     # load pretrained models or checkpoints
     init_model(global_config, train_prog, exe)
 
-    if 'AMP' in config and config.AMP.get("level", "O1") == "O2":
+    if 'AMP' in config:
         optimizer.amp_init(
             device,
             scope=paddle.static.global_scope(),
             test_program=eval_prog
-            if global_config["eval_during_train"] else None)
+            if global_config["eval_during_train"] else None,
+            use_fp16_test=config["AMP"].get("use_fp16_test", False))
 
     if not global_config.get("is_distributed", True):
         compiled_train_prog = program.compile(
@@ -182,7 +183,7 @@ def main(args):
         program.run(train_dataloader, exe, compiled_train_prog, train_feeds,
                     train_fetchs, epoch_id, 'train', config, vdl_writer,
                     lr_scheduler, args.profiler_options)
-        # 2. evaate with eval dataset
+        # 2. evaluate with eval dataset
         if global_config["eval_during_train"] and epoch_id % global_config[
                 "eval_interval"] == 0:
             top1_acc = program.run(eval_dataloader, exe, compiled_eval_prog,
-- 
GitLab