fix figure and weights inconsistent error

README.md

@@ -29,8 +29,8 @@ The training was performed with the following:
 - GPU: at least 12GB of GPU memory
 - Actual Model Input: 96 x 96 x 96
 - AMP: True
-- Optimizer: Adam
-- Learning Rate: 1e-4
+- Optimizer: Novograd
+- Learning Rate: 0.002
 - Loss: DiceCELoss
 
 ### Input
@@ -43,13 +43,13 @@ Two channels
 - Label 0: everything else
 
 ## Performance
-Dice score is used for evaluating the performance of the model. This model achieves a mean dice score of 0.96.
+Dice score is used for evaluating the performance of the model. This model achieves a mean dice score of 0.959.
 
 #### Training Loss
-![A graph showing the training loss over 1260 epochs (10080 iterations).](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_train_2.png)
+![A graph showing the training loss over 1260 epochs (10080 iterations).](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_train_3.png)
 
 #### Validation Dice
-![A graph showing the validation mean Dice over 1260 epochs.](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_val_2.png)
+![A graph showing the validation mean Dice over 1260 epochs.](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_val_3.png)
 
 #### TensorRT speedup
 The `spleen_ct_segmentation` bundle supports the TensorRT acceleration. The table below shows the speedup ratios benchmarked on an A100 80G GPU. The `model computation` means the speedup ratio of model's inference with a random input without preprocessing and postprocessing. The `model computation(onnx)` basically means the same thing as the `model computation`, except that the model is converted through the onnx-torchscript way. We add this line in the table since it has a better performance than the model converted through Torch-TensorRT. The `end2end` means run the bundle end to end with the TensorRT based model converted through Torch-TensorRT. The `torch_fp32` and `torch_amp` is for the pytorch model with or without `amp` mode. The `trt_fp32` and `trt_fp16` is for the TensorRT based model converted in corresponding precision. The `speedup amp`, `speedup fp32` and `speedup fp16` is the speedup ratio of corresponding models versus the pytorch float32 model, while the `amp vs fp16` is between the pytorch amp model and the TensorRT float16 based model.

configs/metadata.json

@@ -1,7 +1,8 @@
 {
     "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20220324.json",
-    "version": "0.4.2",
+    "version": "0.4.3",
     "changelog": {
+        "0.4.3": "fix figure and weights inconsistent error",
         "0.4.2": "use torch 1.13.1",
         "0.4.1": "update the readme file with TensorRT convert",
         "0.4.0": "fix multi-gpu train config typo",
@@ -38,7 +39,7 @@
     "label_classes": "single channel data, 1 is spleen, 0 is everything else",
     "pred_classes": "2 channels OneHot data, channel 1 is spleen, channel 0 is background",
     "eval_metrics": {
-        "mean_dice": 0.96
+        "mean_dice": 0.959
     },
     "intended_use": "This is an example, not to be used for diagnostic purposes",
     "references": [

configs/train.json

@@ -10,7 +10,8 @@
     "dataset_dir": "/workspace/data/Task09_Spleen",
     "images": "$list(sorted(glob.glob(@dataset_dir + '/imagesTr/*.nii.gz')))",
     "labels": "$list(sorted(glob.glob(@dataset_dir + '/labelsTr/*.nii.gz')))",
-    "val_interval": 5,
+    "val_interval": 1,
+    "epochs": 800,
     "device": "$torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')",
     "network_def": {
         "_target_": "UNet",
@@ -36,15 +37,26 @@
     "network": "$@network_def.to(@device)",
     "loss": {
         "_target_": "DiceCELoss",
+        "include_background": true,
         "to_onehot_y": true,
         "softmax": true,
         "squared_pred": true,
-        "batch": true
+        "batch": true,
+        "smooth_nr": 1e-05,
+        "smooth_dr": 1e-05,
+        "lambda_dice": 0.5,
+        "lambda_ce": 0.5
     },
     "optimizer": {
-        "_target_": "torch.optim.Adam",
+        "_target_": "Novograd",
         "params": "$@network.parameters()",
-        "lr": 0.0001
+        "lr": 0.002
+    },
+    "lr_scheduler": {
+        "_target_": "torch.optim.lr_scheduler.StepLR",
+        "optimizer": "@optimizer",
+        "step_size": 5000,
+        "gamma": 0.1
     },
     "train": {
         "deterministic_transforms": [
@@ -167,6 +179,11 @@
             ]
         },
         "handlers": [
+            {
+                "_target_": "LrScheduleHandler",
+                "lr_scheduler": "@lr_scheduler",
+                "print_lr": true
+            },
             {
                 "_target_": "ValidationHandler",
                 "validator": "@validate#evaluator",
@@ -193,7 +210,7 @@
         },
         "trainer": {
             "_target_": "SupervisedTrainer",
-            "max_epochs": 100,
+            "max_epochs": "@epochs",
             "device": "@device",
             "train_data_loader": "@train#dataloader",
             "network": "@network",

docs/README.md

@@ -22,8 +22,8 @@ The training was performed with the following:
 - GPU: at least 12GB of GPU memory
 - Actual Model Input: 96 x 96 x 96
 - AMP: True
-- Optimizer: Adam
-- Learning Rate: 1e-4
+- Optimizer: Novograd
+- Learning Rate: 0.002
 - Loss: DiceCELoss
 
 ### Input
@@ -36,13 +36,13 @@ Two channels
 - Label 0: everything else
 
 ## Performance
-Dice score is used for evaluating the performance of the model. This model achieves a mean dice score of 0.96.
+Dice score is used for evaluating the performance of the model. This model achieves a mean dice score of 0.959.
 
 #### Training Loss
-![A graph showing the training loss over 1260 epochs (10080 iterations).](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_train_2.png)
+![A graph showing the training loss over 1260 epochs (10080 iterations).](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_train_3.png)
 
 #### Validation Dice
-![A graph showing the validation mean Dice over 1260 epochs.](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_val_2.png)
+![A graph showing the validation mean Dice over 1260 epochs.](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_spleen_ct_segmentation_val_3.png)
 
 #### TensorRT speedup
 The `spleen_ct_segmentation` bundle supports the TensorRT acceleration. The table below shows the speedup ratios benchmarked on an A100 80G GPU. The `model computation` means the speedup ratio of model's inference with a random input without preprocessing and postprocessing. The `model computation(onnx)` basically means the same thing as the `model computation`, except that the model is converted through the onnx-torchscript way. We add this line in the table since it has a better performance than the model converted through Torch-TensorRT. The `end2end` means run the bundle end to end with the TensorRT based model converted through Torch-TensorRT. The `torch_fp32` and `torch_amp` is for the pytorch model with or without `amp` mode. The `trt_fp32` and `trt_fp16` is for the TensorRT based model converted in corresponding precision. The `speedup amp`, `speedup fp32` and `speedup fp16` is the speedup ratio of corresponding models versus the pytorch float32 model, while the `amp vs fp16` is between the pytorch amp model and the TensorRT float16 based model.

models/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aeb453bda5be3653f3eec5795de5c5435c41e4b712e7d39e2d44f2461aab7ac8
-size 19303897
+oid sha256:57801867b520d353b6b8fa93a511ad4b3050659872255361fcfc5d5b77320692
+size 19297197

models/model.ts

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1bfeacbda35620f7a8edd7a5b75dc34255a234bb516dfd5c8df1408191c5159a
-size 19398019
+oid sha256:f325fbb60833b0946e234ab14590bde652223503d41f445da879f0396f08a21a
+size 19411907