seminar6 part1 ROI ts upd

seminar6/seminar_6_ROI_time_series.ipynb

@@ -24,8 +24,7 @@
       "provenance": [],
       "collapsed_sections": [
         "hqZksnfwt34o",
-        "D4_cLqPStvNy",
-        "_Pbss93ttqMP"
+        "D4_cLqPStvNy"
       ]
     },
     "accelerator": "GPU"
@@ -1158,7 +1157,9 @@
         "id": "Vlu1fvBts2fN"
       },
       "source": [
-        "## RNN for multivariate time series"
+        "## RNN for multivariate time series\n",
+        "\n",
+        "Since the data is essentially a multidimensional time series, it seems natural to try to analyze it with a recurrent neural network. "
       ]
     },
     {
@@ -1167,7 +1168,11 @@
         "id": "_Pbss93ttqMP"
       },
       "source": [
-        "### RNN on  ROI time series (with GRU units)"
+        "### RNN on  ROI time series (with GRU units)\n",
+        "\n",
+        "Here we define a general RNN architecture that consists of **1 or more** recurrent layers with GRU units followed by **2** fully connected layers. \n",
+        "\n",
+        "Data is sequentially processed by recurrent layers. Then we take the **last** computed hidden state, **mean** of all hidden states or a vector of **concatenated hidden states**, feed them into fully connected layer and predict the ASD probability. "
       ]
     },
     {
@@ -1238,7 +1243,11 @@
       "source": [
         "### Training\n",
         "\n",
-        "Split data into training and validation parts, and create **train and val dataloaders**. Then create **GRUModel** with parameters of your choice (don't make it too complex), optimizer and scheduler (if needed). Train the model to detect patients with ASD from healthy control and measure its **ROC AUC** on the validation set. "
+        "Split data into training and validation parts, and create **train and val dataloaders**. Then create **GRUModel** with parameters of your choice (don't make it too complex), optimizer and scheduler (if needed). Train the model to detect patients with ASD from healthy control and measure its **ROC AUC** on the validation set. \n",
+        "\n",
+        "You can define a model with arguments of your choice. For example, try to vary `hidden_size`, `n_layers`, `use_states` and `n_fc_units` arguments. However, remember that the training sample is still quite small for DL, and the recurrent network has many parameters, so don't make it too large. \n",
+        "\n",
+        "Also, since ovefitting is very probable, you may want to properly choose `dropout`, `weight_decay` and `n_epochs` values. "
       ]
     },
     {
@@ -1252,10 +1261,7 @@
         }
       },
       "source": [
-        "# training example\n",
-        "\n",
-        "# dataset includes data from several different sources (acqusition sites)\n",
-        "# we will use only 2 with largest contribution - NYU and UM\n",
+        "# use data from all sources\n",
         "dataset.use_sources = []\n",
         "\n",
         "# dataset have target classes 1 and 2\n",