[UnitTest] mmocr unittest (open-mmlab#130)

* WIP test_mmocr 8 out of 20

* test_mmocr_export

* test mmocr apis

* add test data

* add mmocr model unittest 5 passed 1 failed

* finish mmocr unittest

* fix lint

* fix yapf

* fix isort

* fix flake8

* fix docformatter

* fix docformatter

* try to fix unittest after merge master

* Change test.py for backend.DEFAULT

* fix flake8

* fix ut

* fix yapf

* fix ut build

* fix yapf

* fix mmocr_export ut

* fix mmocr_apis ort not cuda

* remove explicit .forward

* remove backendwrapper

* simplify the crnn and dbnet config

* simplify instance_test.json

* add another case of decoder

* increase coverage of test_mmocr_models base_recognizer

* improve coverage

* improve encode_decoder coverage

* reply for grimoire codereview

* what if not check cuda?

* remove image data

* reply to runningleon code review

* fix fpnc

* fix lint

* try to fix CI UT error

* fix fpnc with and wo custom ops

* fix yapf

* skip fpnc when cuda is not ready in ci

* reply for code review

* reply for code review

* fix yapf

* reply for code review

* fix yapf

* fix conflict

* remove unmatched data path

* remove unnecessary comments

Author: hanrui1sensetime

mmdeploy/mmocr/apis/inference.py

@@ -1,6 +1,7 @@
 from typing import Iterable, Sequence, Union
 
 import mmcv
+import numpy as np
 import torch
 from mmdet.models.builder import DETECTORS
 from mmocr.datasets import DATASETS
@@ -60,6 +61,8 @@ def simple_test(self,
             list: A list of predictions.
         """
         pred = self.forward_of_backend(img, img_metas, *args, **kwargs)
+        if isinstance(pred, np.ndarray):
+            pred = torch.from_numpy(pred[0])
         if len(img_metas) > 1:
             boundaries = [
                 self.bbox_head.get_boundary(
@@ -187,7 +190,6 @@ def forward_of_backend(self, img: torch.Tensor, img_metas: Iterable, *args,
             np.ndarray: Prediction of input model.
         """
         onnx_pred = self.model({'input': img})
-        onnx_pred = torch.from_numpy(onnx_pred[0])
         return onnx_pred
 
 
@@ -223,7 +225,6 @@ def forward_of_backend(self, img: torch.Tensor, img_metas: Sequence[dict],
             np.ndarray: Prediction of input model.
         """
         onnx_pred = self.model({'input': img})
-        onnx_pred = torch.from_numpy(onnx_pred[0])
         return onnx_pred
 
 
@@ -403,8 +404,8 @@ def forward_of_backend(self, img: torch.Tensor, img_metas: Sequence[dict],
         """
         with torch.cuda.device(self.device_id), torch.no_grad():
             ppl_pred = self.model({'input': img})
-
-        ppl_pred = torch.from_numpy(ppl_pred[0])
+        if isinstance(ppl_pred[0], np.ndarray):
+            ppl_pred = torch.from_numpy(ppl_pred[0])
         return ppl_pred
 
 
@@ -442,7 +443,8 @@ def forward_of_backend(self, img: torch.Tensor, img_metas: Sequence[dict],
         """
         with torch.cuda.device(self.device_id), torch.no_grad():
             ppl_pred = self.model({'input': img})[0]
-        ppl_pred = torch.from_numpy(ppl_pred[0])
+        if isinstance(ppl_pred[0], np.ndarray):
+            ppl_pred = torch.from_numpy(ppl_pred[0])
         return ppl_pred
 
 

mmdeploy/utils/test.py

@@ -295,6 +295,8 @@ def get_rewrite_outputs(wrapped_model: nn.Module,
         backend_model = openvino_apis.OpenVINOWrapper(openvino_file_path)
 
         backend_feats = flatten_model_inputs
+    elif backend == Backend.DEFAULT:
+        return ctx_outputs, False
     else:
         raise NotImplementedError(
             f'Unimplemented backend type: {backend.value}')

tests/test_mmocr/data/config/crnn.py

@@ -0,0 +1,15 @@
+_base_ = []
+
+# model
+label_convertor = dict(
+    type='CTCConvertor', dict_type='DICT36', with_unknown=False, lower=True)
+
+model = dict(
+    type='CRNNNet',
+    preprocessor=None,
+    backbone=dict(type='VeryDeepVgg', leaky_relu=False, input_channels=1),
+    encoder=None,
+    decoder=dict(type='CRNNDecoder', in_channels=512, rnn_flag=True),
+    loss=dict(type='CTCLoss'),
+    label_convertor=label_convertor,
+    pretrained=None)

tests/test_mmocr/data/config/dbnet.py

@@ -0,0 +1,49 @@
+model = dict(
+    type='DBNet',
+    backbone=dict(
+        type='mmdet.ResNet',
+        depth=18,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        frozen_stages=-1,
+        norm_cfg=dict(type='BN', requires_grad=True),
+        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet18'),
+        norm_eval=False,
+        style='caffe'),
+    neck=dict(type='FPNC', in_channels=[2, 4, 8, 16], lateral_channels=8),
+    bbox_head=dict(
+        type='DBHead',
+        text_repr_type='quad',
+        in_channels=8,
+        loss=dict(type='DBLoss', alpha=5.0, beta=10.0, bbce_loss=True)),
+    train_cfg=None,
+    test_cfg=None)
+
+dataset_type = 'IcdarDataset'
+data_root = 'data/icdar2015'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+
+test_pipeline = [
+    dict(type='LoadImageFromFile', color_type='color_ignore_orientation'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(128, 64),
+        flip=False,
+        transforms=[
+            dict(type='Resize', img_scale=(256, 128), keep_ratio=True),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=16,
+    test_dataloader=dict(samples_per_gpu=1),
+    test=dict(
+        type=dataset_type,
+        ann_file=data_root + '/instances_test.json',
+        img_prefix=data_root + '/imgs',
+        pipeline=test_pipeline))
+evaluation = dict(interval=100, metric='hmean-iou')

tests/test_mmocr/data/icdar2015/instances_test.json

@@ -0,0 +1 @@
+{"images": [], "categories": [], "annotations": []}