Merge branch 'main' of github.com:Project-MONAI/tutorials into update-mil

Author: bhashemian

README.md

@@ -175,7 +175,7 @@ And compares the training speed and memory usage with/without AMP.
 This notebook compares the performance of `Dataset`, `CacheDataset` and `PersistentDataset`. These classes differ in how data is stored (in memory or on disk), and at which moment transforms are applied.
 #### [fast_training_tutorial](./acceleration/fast_training_tutorial.ipynb)
 This tutorial compares the training performance of pure PyTorch program and optimized program in MONAI based on NVIDIA GPU device and latest CUDA library.
-The optimization methods mainly include: `AMP`, `CacheDataset` and `Novograd`.
+The optimization methods mainly include: `AMP`, `CacheDataset`, `GPU transforms`, `ThreadDataLoader`, `DiceCELoss` and `SGD`.
 #### [multi_gpu_test](./acceleration/multi_gpu_test.ipynb)
 This notebook is a quick demo for devices, run the Ignite trainer engine on CPU, GPU and multiple GPUs.
 #### [threadbuffer_performance](./acceleration/threadbuffer_performance.ipynb)

acceleration/fast_model_training_guide.md

@@ -305,14 +305,13 @@ With all the above strategies, in this section, we introduce how to apply them t
 ### 1. Spleen segmentation
 
 - Select the algorithms based on the experiments.
-  1. As a binary segmentation task, we replaced the baseline `Dice` loss with a `DiceCE` loss, it can help improve the convergence. To achieve the target metric (mean Dice = 0.95) it reduces the number of training epochs from 200 to 50.
-  2. We tried several numerical optimizers, and finally replaced the baseline `Adam` optimizer with `Novograd`, which consistently reduce the number of training epochs from 50 to 30.
+  As a binary segmentation task, we replaced the baseline `Dice` loss with a `DiceCE` loss, it can help improve the convergence. And we tried to analyze the training curve and tuned different parameters of the network and tested several numerical optimizers, finally replaced the baseline `Adam` optimizer with `SGD`. To achieve the target metric (`mean Dice = 0.94` of the `foreground` channel only) it reduces the number of training epochs from 280 to 60.
 - Optimize GPU utilization.
   1. With `AMP`, the training speed is significantly improved and can achieve almost the same validation metric as without `AMP`.
   2. The deterministic transform results of all the spleen dataset is around 8 GB, which can be cached in a V100 GPU memory. So, we cached all the data in GPU memory and executed the following transforms in GPU directly.
 - Replace `DataLoader` with `ThreadDataLoader`. As all the data are cached in GPU, the computation of randomized transforms is on GPU and light-weighted, `ThreadDataLoader` help avoid the IPC cost of multi-processing in `DataLoader` and increase the GPU utilization.
 
-In summary, with a V100 GPU, we can achieve the training converges at a target validation mean Dice of `0.95` within one minute (`52s` on a V100 GPU, `41s` on an A100 GPU), it is approximately `200x` faster compared with the native PyTorch implementation when achieving the target metric. And each epoch is `20x` faster than the regular training.
+In summary, with a V100 GPU and the target validation `mean dice = 0.94` of the `forground` channel only,  it's more than `100x` speedup compared with the Pytorch regular implementation when achieving the same metric (validation accuracies). And every epoch is `20x` faster than regular training.
 ![spleen fast training](../figures/fast_training.png)
 
 More details are available at [Spleen fast training tutorial](https://github.com/Project-MONAI/tutorials/blob/main/acceleration/fast_training_tutorial.ipynb).

acceleration/fast_training_tutorial.ipynb

@@ -16,7 +16,7 @@
     "authors": "MONAI team",
     "copyright": "Copyright (c) MONAI Consortium",
     "data_source": "Task09_Spleen.tar from http://medicaldecathlon.com/",
-    "data_type": "dicom",
+    "data_type": "nibabel",
     "image_classes": "single channel data, intensity scaled to [0, 1]",
     "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",
@@ -32,19 +32,20 @@
         "inputs": {
             "image": {
                 "type": "image",
-                "format": "magnitude",
+                "format": "hounsfield",
+                "modality": "CT",
                 "num_channels": 1,
                 "spatial_shape": [
-                    160,
-                    160,
-                    160
+                    96,
+                    96,
+                    96
                 ],
                 "dtype": "float32",
                 "value_range": [
                     0,
                     1
                 ],
-                "is_patch_data": false,
+                "is_patch_data": true,
                 "channel_def": {
                     "0": "image"
                 }
@@ -56,16 +57,16 @@
                 "format": "segmentation",
                 "num_channels": 2,
                 "spatial_shape": [
-                    160,
-                    160,
-                    160
+                    96,
+                    96,
+                    96
                 ],
                 "dtype": "float32",
                 "value_range": [
                     0,
                     1
                 ],
-                "is_patch_data": false,
+                "is_patch_data": true,
                 "channel_def": {
                     "0": "background",
                     "1": "spleen"