From 358b76fa4ca71d797417f6456b2a4c2cba1959f5 Mon Sep 17 00:00:00 2001
From: ceci3 <ceci3@users.noreply.github.com>
Date: Fri, 8 Jul 2022 16:19:45 +0800
Subject: [PATCH] change threshold for ptq hpo (#1253)

---
 paddleslim/auto_compression/auto_strategy.py |  4 ++--
 paddleslim/quant/post_quant_hpo.py           | 18 ++++++++++--------
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/paddleslim/auto_compression/auto_strategy.py b/paddleslim/auto_compression/auto_strategy.py
index 2826601a..451f6007 100644
--- a/paddleslim/auto_compression/auto_strategy.py
+++ b/paddleslim/auto_compression/auto_strategy.py
@@ -77,8 +77,8 @@ EXPERIENCE_STRATEGY_WITHOUT_LOSS = [
 MAGIC_SPARSE_RATIO = 0.75
 ### TODO: 0.02 threshold maybe not suitable, need to check
 ### NOTE: reduce magic data to choose quantization aware training.
-MAGIC_MAX_EMD_DISTANCE = 0.0002  #0.02
-MAGIC_MIN_EMD_DISTANCE = 0.0001  #0.01
+MAGIC_MAX_EMD_DISTANCE = 0.00002  #0.02
+MAGIC_MIN_EMD_DISTANCE = 0.00001  #0.01
 
 DEFAULT_TRANSFORMER_STRATEGY = 'prune_0.25_int8'
 DEFAULT_STRATEGY = 'origin_int8'
diff --git a/paddleslim/quant/post_quant_hpo.py b/paddleslim/quant/post_quant_hpo.py
index 9f9275f8..adb5ffa5 100755
--- a/paddleslim/quant/post_quant_hpo.py
+++ b/paddleslim/quant/post_quant_hpo.py
@@ -144,7 +144,12 @@ def standardization(data):
     """standardization numpy array"""
     mu = np.mean(data, axis=0)
     sigma = np.std(data, axis=0)
-    sigma = 1e-13 if sigma == 0. else sigma
+    if isinstance(sigma, list) or isinstance(sigma, np.ndarray):
+        for idx, sig in enumerate(sigma):
+            if sig == 0.:
+                sigma[idx] = 1e-13
+    else:
+        sigma = 1e-13 if sigma == 0. else sigma
     return (data - mu) / sigma
 
 
@@ -241,18 +246,15 @@ def eval_quant_model():
         if have_invalid_num(out_float) or have_invalid_num(out_quant):
             continue
 
-        try:
-            out_float = standardization(out_float)
-            out_quant = standardization(out_quant)
-        except:
-            continue
-        out_float_list.append(out_float)
-        out_quant_list.append(out_quant)
+        out_float_list.append(list(out_float))
+        out_quant_list.append(list(out_quant))
         valid_data_num += 1
 
         if valid_data_num >= max_eval_data_num:
             break
 
+    out_float_list = standardization(out_float_list)
+    out_quant_list = standardization(out_quant_list)
     emd_sum = cal_emd_lose(out_float_list, out_quant_list,
                            out_len_sum / float(valid_data_num))
     _logger.info("output diff: {}".format(emd_sum))
-- 
GitLab