add resnet export of onnx (#341)

* add checkpoint_sync_export for resnet config

configs/classification/imagenet/resnet/imagenet_resnet50_jpg.py

@@ -6,3 +6,6 @@ model = dict(
         depth=50,
         out_indices=[4],  # 0: conv-1, x: stage-x
         norm_cfg=dict(type='BN')))
+
+checkpoint_sync_export = True
+export = dict(export_type='raw', export_neck=True)

easycv/apis/export.py

@@ -157,6 +157,37 @@ def _export_jit_and_blade(model, cfg, filename, dummy_inputs, fp16=False):
         torch.jit.save(blade_model, ofile)
 
 
+def _export_onnx_cls(model, model_config, cfg, filename, meta):
+
+    if model_config['backbone'].get(
+            'type', None) == 'ResNet' and model_config['backbone'].get(
+                'depth', None) == 50:
+        # save json config for test_pipline and class
+        with io.open(
+                filename +
+                '.config.json' if filename.endswith('onnx') else filename +
+                '.onnx.config.json', 'w') as ofile:
+            json.dump(meta, ofile)
+
+        device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        model.eval()
+        model.to(device)
+        img_size = int(cfg.image_size2)
+        x_input = torch.randn((1, 3, img_size, img_size)).to(device)
+        torch.onnx.export(
+            model,
+            (x_input, 'onnx'),
+            filename if filename.endswith('onnx') else filename + '.onnx',
+            export_params=True,
+            opset_version=12,
+            do_constant_folding=True,
+            input_names=['input'],
+            output_names=['output'],
+        )
+    else:
+        raise ValueError('Only support export onnx model for ResNet now!')
+
+
 def _export_cls(model, cfg, filename):
     """ export cls (cls & metric learning)model and preprocess config
 
@@ -170,6 +201,7 @@ def _export_cls(model, cfg, filename):
     else:
         export_cfg = dict(export_neck=False)
 
+    export_type = export_cfg.get('export_type', 'raw')
     export_neck = export_cfg.get('export_neck', True)
     label_map_path = cfg.get('label_map_path', None)
     class_list = None
@@ -232,9 +264,14 @@ def _export_cls(model, cfg, filename):
         if export_neck and (k.startswith('neck') or k.startswith('head')):
             state_dict[k] = v
 
-    checkpoint = dict(state_dict=state_dict, meta=meta, author='EasyCV')
+    if export_type == 'raw':
-    with io.open(filename, 'wb') as ofile:
+        checkpoint = dict(state_dict=state_dict, meta=meta, author='EasyCV')
-        torch.save(checkpoint, ofile)
+        with io.open(filename, 'wb') as ofile:
+            torch.save(checkpoint, ofile)
+    elif export_type == 'onnx':
+        _export_onnx_cls(model, model_config, cfg, filename, config)
+    else:
+        raise ValueError('Only support export onnx/raw model!')
 
 
 def _export_yolox(model, cfg, filename):

easycv/models/classification/classification.py

@@ -151,6 +151,20 @@ class Classification(BaseModel):
         x = self.backbone(img)
         return x
 
+    def forward_onnx(self, img: torch.Tensor) -> Dict[str, torch.Tensor]:
+        """
+            forward_onnx means generate prob from image only support one neck  + one head
+        """
+        x = self.forward_backbone(img)  # tuple
+
+        # if self.neck_num > 0:
+        if hasattr(self, 'neck_0'):
+            x = self.neck_0([i for i in x])
+
+        out = self.head_0(x)[0].cpu()
+        out = self.activate_fn(out)
+        return out
+
     @torch.jit.unused
     def forward_train(self, img, gt_labels) -> Dict[str, torch.Tensor]:
         """
@@ -290,6 +304,9 @@ class Classification(BaseModel):
                 return self.forward_test_label(img, gt_labels)
             else:
                 return self.forward_test(img)
+        elif mode == 'onnx':
+            return self.forward_onnx(img)
+
         elif mode == 'extract':
             rd = self.forward_feature(img)
             rv = {}

easycv/predictors/classifier.py

@@ -1,5 +1,7 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
+import glob
 import math
+import os
 
 import numpy as np
 import torch
@@ -7,11 +9,18 @@ from PIL import Image
 
 from easycv.file import io
 from easycv.framework.errors import ValueError
+from easycv.utils.checkpoint import load_checkpoint
 from easycv.utils.misc import deprecated
 from .base import InputProcessor, OutputProcessor, Predictor, PredictorV2
 from .builder import PREDICTORS
 
 
+# onnx specific
+def onnx_to_numpy(tensor):
+    return tensor.detach().cpu().numpy(
+    ) if tensor.requires_grad else tensor.cpu().numpy()
+
+
 class ClsInputProcessor(InputProcessor):
     """Process inputs for classification models.
 
@@ -146,6 +155,20 @@ class ClassificationPredictor(PredictorV2):
         self.pil_input = pil_input
         self.label_map_path = label_map_path
 
+        if model_path.endswith('onnx'):
+            self.model_type = 'onnx'
+            pwd_model = os.path.dirname(model_path)
+            raw_model = glob.glob(
+                os.path.join(pwd_model, '*.onnx.config.json'))
+            if len(raw_model) != 0:
+                config_file = raw_model[0]
+            else:
+                assert len(
+                    raw_model
+                ) == 0, 'Please have a file with the .onnx.config.json extension in your directory'
+        else:
+            self.model_type = 'raw'
+
         if self.pil_input:
             mode = 'RGB'
         super(ClassificationPredictor, self).__init__(
@@ -186,6 +209,41 @@ class ClassificationPredictor(PredictorV2):
 
         return ClsOutputProcessor(topk=self.topk, label_map=self.label_map)
 
+    def prepare_model(self):
+        """Build model from config file by default.
+        If the model is not loaded from a configuration file, e.g. torch jit model, you need to reimplement it.
+        """
+        if self.model_type == 'raw':
+            model = self._build_model()
+            model.to(self.device)
+            model.eval()
+            load_checkpoint(model, self.model_path, map_location='cpu')
+            return model
+        else:
+            import onnxruntime
+            if onnxruntime.get_device() == 'GPU':
+                onnx_model = onnxruntime.InferenceSession(
+                    self.model_path, providers=['CUDAExecutionProvider'])
+            else:
+                onnx_model = onnxruntime.InferenceSession(self.model_path)
+
+            return onnx_model
+
+    def model_forward(self, inputs):
+        """Model forward.
+        If you need refactor model forward, you need to reimplement it.
+        """
+        with torch.no_grad():
+            if self.model_type == 'raw':
+                outputs = self.model(**inputs, mode='test')
+            else:
+                outputs = self.model.run(None, {
+                    self.model.get_inputs()[0].name:
+                    onnx_to_numpy(inputs['img'])
+                })[0]
+                outputs = dict(prob=torch.from_numpy(outputs))
+        return outputs
+
 
 try:
     from easy_vision.python.inference.predictor import PredictorInterface

easycv/version.py

@@ -2,5 +2,5 @@
 # GENERATED VERSION FILE
 # TIME: Thu Nov  5 14:17:50 2020
 
-__version__ = '0.11.6'
+__version__ = '0.11.7'
-short_version = '0.11.6'
+short_version = '0.11.7'

tests/test_apis/test_export.py

@@ -116,6 +116,7 @@ class ModelExportTest(unittest.TestCase):
         cfg = mmcv_config_fromfile(config_file)
         cfg_options = {
             'model.backbone.norm_cfg.type': 'SyncBN',
+            'export.export_type': 'raw'
         }
         if cfg_options is not None:
             cfg.merge_from_dict(cfg_options)
@@ -210,6 +211,27 @@ class ModelExportTest(unittest.TestCase):
 
             self.assertTrue(os.path.exists(filename + '.jit'))
 
+    def test_export_resnet_onnx(self):
+
+        ckpt_path = PRETRAINED_MODEL_RESNET50
+
+        easycv_dir = os.path.dirname(easycv.__file__)
+
+        if os.path.exists(os.path.join(easycv_dir, 'configs')):
+            config_dir = os.path.join(easycv_dir, 'configs')
+        else:
+            config_dir = os.path.join(os.path.dirname(easycv_dir), 'configs')
+        config_file = os.path.join(
+            config_dir,
+            'classification/imagenet/resnet/imagenet_resnet50_jpg.py')
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            cfg = mmcv_config_fromfile(config_file)
+            cfg.export.export_type = 'onnx'
+            filename = os.path.join(tmpdir, 'imagenet_resnet50')
+            export(cfg, ckpt_path, filename)
+            self.assertTrue(os.path.exists(filename + '.onnx'))
+
 
 if __name__ == '__main__':
     unittest.main()

tests/test_predictors/test_classifier.py

@@ -11,7 +11,8 @@ import torch
 from easycv.predictors.classifier import ClassificationPredictor
 from easycv.utils.test_util import clean_up, get_tmp_dir
 from tests.ut_config import (PRETRAINED_MODEL_RESNET50_WITHOUTHEAD,
-                             IMAGENET_LABEL_TXT, TEST_IMAGES_DIR)
+                             IMAGENET_LABEL_TXT, TEST_IMAGES_DIR,
+                             PRETRAINED_MODEL_RESNET50_ONNX_WITHOUTHEAD)
 
 
 class ClassificationPredictorTest(unittest.TestCase):
@@ -33,6 +34,17 @@ class ClassificationPredictorTest(unittest.TestCase):
         self.assertListEqual(results['class_name'], ['"Persian cat",'])
         self.assertEqual(len(results['class_probs']), 1000)
 
+    def test_onnx_single(self):
+        checkpoint = PRETRAINED_MODEL_RESNET50_ONNX_WITHOUTHEAD
+        predict_op = ClassificationPredictor(model_path=checkpoint)
+
+        img_path = os.path.join(TEST_IMAGES_DIR, 'catb.jpg')
+
+        results = predict_op([img_path])[0]
+        self.assertListEqual(results['class'], [578])
+        self.assertListEqual(results['class_name'], ['gown'])
+        self.assertEqual(len(results['class_probs']), 1000)
+
     def test_batch(self):
         checkpoint = PRETRAINED_MODEL_RESNET50_WITHOUTHEAD
         config_file = 'configs/classification/imagenet/resnet/imagenet_resnet50_jpg.py'

tests/test_predictors/test_pose_predictor.py

@@ -54,10 +54,10 @@ class PoseTopDownPredictorTest(unittest.TestCase):
 
         assert_array_almost_equal(
             result0['bbox'],
-            np.array([[352.3085, 59.00325, 691.4247, 511.15814, 1.],
+            np.array([[438.9, 59., 604.8, 511.2, 0.9],
-                      [10.511196, 177.74883, 101.824326, 299.49966, 1.],
+                      [10.5, 179.6, 101.8, 297.7, 0.9],
-                      [224.82036, 114.439865, 312.51306, 231.36348, 1.],
+                      [229.6, 114.4, 307.8, 231.4, 0.6],
-                      [200.71407, 114.716736, 337.17535, 296.6651, 1.]],
+                      [229.4, 114.7, 308.5, 296.7, 0.6]],
                      dtype=np.float32),
             decimal=1)
         vis_result = predictor.show_result(img1, result0)
@@ -92,10 +92,10 @@ class PoseTopDownPredictorTest(unittest.TestCase):
 
         assert_array_almost_equal(
             result1['bbox'][:4],
-            np.array([[436.23096, 214.72766, 584.26013, 412.09985, 1.],
+            np.array([[470.6, 214.7, 549.9, 412.1, 0.9],
-                      [43.990044, 91.04126, 164.28406, 251.43329, 1.],
+                      [71.6, 91., 136.7, 251.4, 0.9],
-                      [127.44148, 100.38604, 254.219, 269.42273, 1.],
+                      [159.7, 100.4, 221.9, 269.4, 0.9],
-                      [190.08075, 117.31801, 311.22394, 278.8423, 1.]],
+                      [219.4, 117.3, 281.9, 278.8, 0.9]],
                      dtype=np.float32),
             decimal=1)
         vis_result = predictor.show_result(img2, result1)

tests/ut_config.py

@@ -179,6 +179,9 @@ PRETRAINED_MODEL_RESNET50 = os.path.join(
 PRETRAINED_MODEL_RESNET50_WITHOUTHEAD = os.path.join(
     BASE_LOCAL_PATH,
     'pretrained_models/classification/resnet/resnet50_withhead.pth')
+PRETRAINED_MODEL_RESNET50_ONNX_WITHOUTHEAD = os.path.join(
+    BASE_LOCAL_PATH,
+    'pretrained_models/classification/resnet/imagenet_resnet50.onnx')
 PRETRAINED_MODEL_FACEID = os.path.join(BASE_LOCAL_PATH,
                                        'pretrained_models/faceid')
 PRETRAINED_MODEL_YOLOXS_EXPORT = os.path.join(