add d2t train support of svtr (#9479)

* add d2s train for slanet and v3

* fix bug

* udpate tipc to_static

* update db

* remove_print

* update benchmark_train.sh

* update maybe_download_params

* add d2t train support of svtr

configs/rec/rec_svtrnet.yml

@@ -20,6 +20,7 @@ Global:
   infer_mode: False
   use_space_char: False
   save_res_path: ./output/rec/predicts_svtr_tiny.txt
+  d2s_train_image_shape: [3, 64, 256]
 
 
 Optimizer:

configs/rec/rec_svtrnet_ch.yml

@@ -19,6 +19,7 @@ Global:
   infer_mode: false
   use_space_char: true
   save_res_path: ./output/rec/predicts_svtr_tiny_ch_all.txt
+  d2s_train_image_shape: [3, 32, 320]
 Optimizer:
   name: AdamW
   beta1: 0.9

ppocr/modeling/architectures/__init__.py

@@ -40,7 +40,9 @@ def apply_to_static(model, config, logger):
         return model
     assert "d2s_train_image_shape" in config[
         "Global"], "d2s_train_image_shape must be assigned for static training mode..."
-    supported_list = ["DB", "SVTR_LCNet", "TableMaster", "LayoutXLM", "SLANet"]
+    supported_list = [
+        "DB", "SVTR_LCNet", "TableMaster", "LayoutXLM", "SLANet", "SVTR"
+    ]
     if config["Architecture"]["algorithm"] in ["Distillation"]:
         algo = list(config["Architecture"]["Models"].values())[0]["algorithm"]
     else:
@@ -104,6 +106,13 @@ def apply_to_static(model, config, logger):
             InputSpec(
                 [None, 6], dtype='float64'),
         ])
+    elif algo == "SVTR":
+        specs.append([
+            InputSpec(
+                [None, config["Global"]["max_text_length"]], dtype='int64'),
+            InputSpec(
+                [None], dtype='int64')
+        ])
     model = to_static(model, input_spec=specs)
     logger.info("Successfully to apply @to_static with specs: {}".format(specs))
     return model

ppocr/modeling/transforms/tps_spatial_transformer.py

@@ -140,7 +140,9 @@ class TPSSpatialTransformer(nn.Layer):
 
         padding_matrix = paddle.expand(
             self.padding_matrix, shape=[batch_size, 3, 2])
-        Y = paddle.concat([source_control_points, padding_matrix], 1)
+        Y = paddle.concat([
+            source_control_points.astype(padding_matrix.dtype), padding_matrix
+        ], 1)
         mapping_matrix = paddle.matmul(self.inverse_kernel, Y)
         source_coordinate = paddle.matmul(self.target_coordinate_repr,
                                           mapping_matrix)
@@ -153,4 +155,4 @@ class TPSSpatialTransformer(nn.Layer):
         # the input to grid_sample is normalized [-1, 1], but what we get is [0, 1]
         grid = 2.0 * grid - 1.0
         output_maps = grid_sample(input, grid, canvas=None)
-        return output_maps, source_coordinate
\ No newline at end of file
+        return output_maps, source_coordinate

test_tipc/configs/rec_svtrnet/rec_svtrnet.yml

@@ -20,6 +20,7 @@ Global:
   infer_mode: False
   use_space_char: False
   save_res_path: ./output/rec/predicts_svtr_tiny.txt
+  d2s_train_image_shape: [3, 64, 256]
 
 
 Optimizer:

test_tipc/configs/rec_svtrnet/train_infer_python.txt

@@ -51,3 +51,11 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
 null:null
 ===========================infer_benchmark_params==========================
 random_infer_input:[{float32,[3,64,256]}]
+===========================train_benchmark_params==========================
+batch_size:512
+fp_items:fp32|fp16
+epoch:2
+--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
+flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096
+===========================to_static_train_benchmark_params===========================
+to_static_train:Global.to_static=true
\ No newline at end of file

test_tipc/prepare.sh

@@ -145,6 +145,19 @@ if [ ${MODE} = "benchmark_train" ];then
         cd ./train_data/ && tar xf StructureLabel_val_500.tar
         cd ../
     fi
+    if [ ${model_name} == "rec_svtrnet" ]; then
+        wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/dataset/ic15_data_benckmark.tar --no-check-certificate
+        cd ./train_data/ && tar xf ic15_data_benckmark.tar
+        ln -s ./ic15_data_benckmark ./ic15_data
+        cd ic15_data
+        mv rec_gt_train4w.txt rec_gt_train.txt
+
+        for i in `seq 10`;do cp rec_gt_train.txt dup$i.txt;done
+        cat dup* > rec_gt_train.txt && rm -rf dup*
+
+        cd ../
+        cd ../
+    fi
 fi
 
 if [ ${MODE} = "lite_train_lite_infer" ];then