From ad8b98b476b4387d2fa2a6f2ff11c02c8ac819c1 Mon Sep 17 00:00:00 2001 From: linqingke Date: Mon, 24 Aug 2020 17:36:19 +0800 Subject: [PATCH] fix fasterrcnn and mass bug. --- model_zoo/official/cv/faster_rcnn/README.md | 6 ++++-- model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py | 2 +- model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py | 2 +- model_zoo/official/nlp/mass/train.py | 5 +++-- 4 files changed, 9 insertions(+), 6 deletions(-) diff --git a/model_zoo/official/cv/faster_rcnn/README.md b/model_zoo/official/cv/faster_rcnn/README.md index 7270c0abf8..dfe8c74624 100644 --- a/model_zoo/official/cv/faster_rcnn/README.md +++ b/model_zoo/official/cv/faster_rcnn/README.md @@ -27,7 +27,7 @@ FasterRcnn proposed that convolution feature maps based on region detectors (suc [Paper](https://arxiv.org/abs/1506.01497): Ren S , He K , Girshick R , et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39(6). -#Model Architecture +# Model Architecture FasterRcnn is a two-stage target detection network,This network uses a region proposal network (RPN), which can share the convolution features of the whole image with the detection network, so that the calculation of region proposal is almost cost free. The whole network further combines RPN and FastRcnn into a network by sharing the convolution features. @@ -42,7 +42,7 @@ Dataset used: [COCO2017]() - Data format：image and json files - Note：Data will be processed in dataset.py -#Environment Requirements +# Environment Requirements - Install [MindSpore](https://www.mindspore.cn/install/en). @@ -87,6 +87,8 @@ Dataset used: [COCO2017]() After installing MindSpore via the official website, you can start training and evaluation as follows: +Note: 1.the first run will generate the mindeocrd file, which will take a long time. 2. pretrained model is a resnet50 checkpoint that trained over ImageNet2012. 3. VALIDATION_JSON_FILE is label file. CHECKPOINT_PATH is a checkpoint file after trained. + ``` # standalone training sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py index cc58f92af6..152d37f7b8 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py @@ -97,7 +97,7 @@ class Rcnn(nn.Cell): self.relu = P.ReLU() self.logicaland = P.LogicalAnd() self.loss_cls = P.SoftmaxCrossEntropyWithLogits() - self.loss_bbox = P.SmoothL1Loss(sigma=1.0) + self.loss_bbox = P.SmoothL1Loss(beta=1.0) self.reshape = P.Reshape() self.onehot = P.OneHot() self.greater = P.Greater() diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py index 5d8335a709..ff36578db7 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py @@ -137,7 +137,7 @@ class RPN(nn.Cell): self.CheckValid = P.CheckValid() self.sum_loss = P.ReduceSum() self.loss_cls = P.SigmoidCrossEntropyWithLogits() - self.loss_bbox = P.SmoothL1Loss(sigma=1.0/9.0) + self.loss_bbox = P.SmoothL1Loss(beta=1.0/9.0) self.squeeze = P.Squeeze() self.cast = P.Cast() self.tile = P.Tile() diff --git a/model_zoo/official/nlp/mass/train.py b/model_zoo/official/nlp/mass/train.py index 81854bcb8c..14ab3c1b82 100644 --- a/model_zoo/official/nlp/mass/train.py +++ b/model_zoo/official/nlp/mass/train.py @@ -151,7 +151,7 @@ def _build_training_pipeline(config: TransformerConfig, if dataset is None: raise ValueError("pre-training dataset or fine-tuning dataset must be provided one.") - update_steps = dataset.get_repeat_count() * dataset.get_dataset_size() + update_steps = config.epochs * dataset.get_dataset_size() if config.lr_scheduler == "isr": lr = Tensor(square_root_schedule(lr=config.lr, update_num=update_steps, @@ -331,7 +331,8 @@ if __name__ == '__main__': mode=context.GRAPH_MODE, device_target=args.platform, reserve_class_name_in_scope=False, - device_id=device_id) + device_id=device_id, + max_call_depth=2000) _rank_size = os.getenv('RANK_SIZE')