Remove tests with save_quant_model.py

python/paddle/static/quantization/tests/CMakeLists.txt

@@ -200,32 +200,6 @@ function(download_quant_model install_dir data_file check_sum)
   endif()
 endfunction()
 
-function(save_quant_ic_model_test target quant_model_dir int8_model_save_path)
-  py_test(
-    ${target}
-    SRCS ${CMAKE_CURRENT_SOURCE_DIR}/save_quant_model.py
-         ARGS
-         --quant_model_path
-         ${quant_model_dir}
-         --int8_model_save_path
-         ${int8_model_save_path}
-         --debug)
-endfunction()
-
-function(save_quant_nlp_model_test target quant_model_dir int8_model_save_path
-         ops_to_quantize)
-  py_test(
-    ${target}
-    SRCS ${CMAKE_CURRENT_SOURCE_DIR}/save_quant_model.py
-         ARGS
-         --quant_model_path
-         ${quant_model_dir}
-         --int8_model_save_path
-         ${int8_model_save_path}
-         --ops_to_quantize
-         ${ops_to_quantize})
-endfunction()
-
 function(convert_model2dot_test target model_path save_graph_dir
          save_graph_name)
   py_test(
@@ -438,31 +412,6 @@ if(LINUX AND WITH_MKLDNN)
   set(QUANT2_LSTM_MODEL_DIR "${QUANT_INSTALL_DIR}/lstm_quant_test")
   download_quant_model(${QUANT2_LSTM_MODEL_DIR} ${QUANT2_LSTM_MODEL_ARCHIVE}
                        40a693803b12ee9e251258f32559abcb)
-  set(QUANT2_LSTM_OPS_TO_QUANTIZE "fusion_lstm")
-
-  ### Save FP32 model or INT8 model from Quant model
-
-  set(QUANT2_INT8_RESNET50_SAVE_PATH
-      "${QUANT_INSTALL_DIR}/ResNet50_quant2_int8")
-  save_quant_ic_model_test(
-    save_quant2_model_resnet50
-    ${QUANT2_RESNET50_MODEL_DIR}/ResNet50_qat_perf/float
-    ${QUANT2_INT8_RESNET50_SAVE_PATH})
-
-  set(QUANT2_INT8_ERNIE_SAVE_PATH "${QUANT_INSTALL_DIR}/Ernie_quant2_int8")
-  save_quant_nlp_model_test(
-    save_quant2_model_ernie ${QUANT2_ERNIE_MODEL_DIR}/Ernie_qat/float
-    ${QUANT2_INT8_ERNIE_SAVE_PATH} ${QUANT2_ERNIE_OPS_TO_QUANTIZE})
-
-  set(QUANT2_INT8_GRU_SAVE_PATH "${QUANT_INSTALL_DIR}/GRU_quant2_int8")
-  save_quant_nlp_model_test(
-    save_quant2_model_gru ${QUANT2_GRU_MODEL_DIR}/GRU_quant_acc
-    ${QUANT2_INT8_GRU_SAVE_PATH} ${QUANT2_GRU_OPS_TO_QUANTIZE})
-
-  set(QUANT2_INT8_LSTM_SAVE_PATH "${QUANT_INSTALL_DIR}/lstm_quant2_int8")
-  save_quant_nlp_model_test(
-    save_quant2_model_lstm ${QUANT2_LSTM_MODEL_DIR}/lstm_quant
-    ${QUANT2_INT8_LSTM_SAVE_PATH} ${QUANT2_LSTM_OPS_TO_QUANTIZE})
 
   # Convert Quant2 model to dot and pdf files
   set(QUANT2_INT8_ERNIE_DOT_SAVE_PATH
@@ -474,6 +423,7 @@ if(LINUX AND WITH_MKLDNN)
   ### PTQ INT8
 
   # PTQ int8 lstm model
+  set(QUANT2_INT8_LSTM_SAVE_PATH "${QUANT_INSTALL_DIR}/lstm_quant2_int8")
   set(LSTM_DATA_FILE "quant_lstm_input_data.tar.gz")
   set(LSTM_URL "${INFERENCE_URL}/int8/unittest_model_data")
   download_data(${QUANT2_INT8_LSTM_SAVE_PATH} ${LSTM_URL} ${LSTM_DATA_FILE}
@@ -561,7 +511,6 @@ if(LINUX AND WITH_MKLDNN)
                                                                      120)
   set_tests_properties(test_quant2_int8_resnet50_range_mkldnn PROPERTIES TIMEOUT
                                                                          120)
-  set_tests_properties(save_quant2_model_resnet50 PROPERTIES TIMEOUT 120)
   set_tests_properties(test_quant_int8_resnet50_mkldnn PROPERTIES TIMEOUT 120)
   set_tests_properties(test_quant_int8_mobilenetv1_mkldnn PROPERTIES TIMEOUT
                                                                      120)

python/paddle/static/quantization/tests/quant2_int8_lstm_model.py

@@ -19,7 +19,6 @@
 import unittest
 
 import numpy as np
-from save_quant_model import transform_and_save_int8_model
 
 import paddle
 from paddle.framework import core
@@ -107,7 +106,7 @@ def set_config(
         mkldnn_cache_capacity,
         warmup_data=None,
         use_analysis=False,
-        enable_ptq=False,
+        mode="fp32",
     ):
         config = core.AnalysisConfig(model_path)
         config.set_cpu_math_library_num_threads(num_threads)
@@ -118,12 +117,15 @@ def set_config(
             config.enable_mkldnn()
             config.disable_mkldnn_fc_passes()  # fc passes caused dnnl error
             config.set_mkldnn_cache_capacity(mkldnn_cache_capacity)
-            if enable_ptq:
+            if mode == "ptq":
                 # This pass to work properly, must be added before fc_fuse_pass
                 config.pass_builder().insert_pass(5, "fc_lstm_fuse_pass")
                 config.enable_quantizer()
                 config.quantizer_config().set_quant_data(warmup_data)
                 config.quantizer_config().set_quant_batch_size(1)
+            elif mode == "qat":
+                config.enable_mkldnn_int8()
+
         return config
 
     def run_program(
@@ -134,7 +136,7 @@ def run_program(
         mkldnn_cache_capacity,
         warmup_iter,
         use_analysis=False,
-        enable_ptq=False,
+        mode="fp32",
     ):
         place = paddle.CPUPlace()
         warmup_data, inputs = self.get_warmup_tensor(data_path, place)
@@ -145,7 +147,7 @@ def run_program(
             mkldnn_cache_capacity,
             warmup_data,
             use_analysis,
-            enable_ptq,
+            mode,
         )
 
         predictor = core.create_paddle_predictor(config)
@@ -228,7 +230,7 @@ def test_lstm_model(self):
             mkldnn_cache_capacity,
             warmup_iter,
             False,
-            False,
+            mode="fp32",
         )
 
         (int8_hx_acc, int8_ctc_acc, int8_fps) = self.run_program(
@@ -238,23 +240,17 @@ def test_lstm_model(self):
             mkldnn_cache_capacity,
             warmup_iter,
             True,
-            True,
-        )
-
-        quant_model_save_path = quant_model + "_int8"
-        # transform model to quant2
-        transform_and_save_int8_model(
-            quant_model, quant_model_save_path, "fusion_lstm,concat"
+            mode="ptq",
         )
 
         (quant_hx_acc, quant_ctc_acc, quant_fps) = self.run_program(
-            quant_model_save_path,
+            quant_model + "_int8",
             infer_data,
             num_threads,
             mkldnn_cache_capacity,
             warmup_iter,
             True,
-            False,
+            mode="qat",
         )
 
         print(
@@ -270,7 +266,7 @@ def test_lstm_model(self):
         )
 
         print(
-            "QUANT2_INT8: fps {0}, hx_acc {1}, ctc_acc {2}".format(
+            "QAT: fps {0}, hx_acc {1}, ctc_acc {2}".format(
                 quant_fps, quant_hx_acc, quant_ctc_acc
             )
         )