CyberShake interpolation - Revision history

Akratzer: /* ML: Intensity Measures */

2024-08-15T00:09:56Z

ML: Intensity Measures

Akratzer: /* ML: Intensity Measures */

2024-08-08T17:45:30Z

ML: Intensity Measures

Akratzer: /* ML: Intensity Measures */

2024-08-08T17:31:32Z

ML: Intensity Measures

Akratzer: /* ML: Intensity Measures */

2024-08-02T19:18:30Z

ML: Intensity Measures

Akratzer: /* ML: Intensity Measures */

2024-08-02T19:18:03Z

ML: Intensity Measures

Akratzer: /* ML: Intensity Measures */

2024-08-02T19:06:05Z

ML: Intensity Measures

Akratzer: /* ML: Hazard Curves */

2024-07-31T01:00:00Z

ML: Hazard Curves

Akratzer: /* ML: Hazard Curves */

2024-07-31T00:58:58Z

ML: Hazard Curves

Akratzer: /* ML: Hazard Curves */

2024-07-31T00:54:44Z

ML: Hazard Curves

Akratzer: /* ML: Hazard Curves */

2024-07-31T00:52:20Z

ML: Hazard Curves

@@ Line 355: / Line 355: @@
 inputs: the IM values of the 4 sites for an event, their distances to the interpolated site, input and interp sites' velocity metrics (Vs30, Z1_0, Z2_5) <br>
 output: the interpolated IM value for that event <br>
 hidden layers with softplus activation function <br>
 I normalized the X inputs and y labels. <br>
 b) Results <br>
 i) Initial run <br>

@@ Line 326: / Line 326: @@
 Next steps: I next need to figure out how to update my model so it can handle all 600,000 events and ultimately perform this interpolation for many site groups as samples. <br>
-<br>
+<br> I realized that my model was having issues with the test loss being NaN, since my input file was missing 400,000 lines of data and the last line had only one value instead of the 8 inputs. Hence, I regenerated my input file, and this seemed to work. <br>
 <br>2) Using the previous model in (1) as a starting point, I expanded my inputs so it includes all events for all sites in my non-10km list. <br>

← Older revision		Revision as of 17:31, 8 August 2024
Line 326:		Line 326:
	Next steps: I next need to figure out how to update my model so it can handle all 600,000 events and ultimately perform this interpolation for many site groups as samples. <br>		Next steps: I next need to figure out how to update my model so it can handle all 600,000 events and ultimately perform this interpolation for many site groups as samples. <br>

−	<br>2)	+	<br>
		+
		+
		+
		+	<br>2) Using the previous model in (1) as a starting point, I expanded my inputs so it includes all events for all sites in my non-10km list. <br>
		+	This involved creating a csv file for each of the events in my list, and combining these files, excluding the files for sites COO, s505, STG, USC, and PLS - my inference sites, into one large file to serve as input to my model. <br>
		+
		+	a) Neural network architecture <br>
		+	23 inputs: the IM values of the 4 sites for an event, their distances to the interpolated site, input and interp sites' velocity metrics (Vs30, Z1_0, Z2_5) <br>
		+	1 output: the interpolated IM value for that event <br>
		+	5 hidden layers with softplus activation function <br>
		+	I normalized the X inputs and y labels. <br>
		+
		+	b) Results <br>
		+	i) Initial run <br>

@@ Line 288: / Line 288: @@
 a) Neural network architecture <br>
-inputs: the IM values of the 4 sites for an event and there distances to the interpolated site <br>
+inputs: the IM values of the 4 sites for an event and their distances to the interpolated site <br>
 output: the interpolated IM value for that event <br>
 hidden layers with softplus activation function <br>

@@ Line 288: / Line 288: @@
 a) Neural network architecture <br>
-inputs: the IM values of the 4 sites for an event and there distances to the interpolated site
+inputs: the IM values of the 4 sites for an event and there distances to the interpolated site <br>
-output: the interpolated IM value for that event
+output: the interpolated IM value for that event <br>
-hidden layers with softplus activation function
+hidden layers with softplus activation function <br>
-I normalized the X inputs and y labels.
+I normalized the X inputs and y labels. <br>
 b) Results <br>
 i) Initial run <br>

← Older revision		Revision as of 19:06, 2 August 2024
Line 281:		Line 281:

	== ML: Intensity Measures ==		== ML: Intensity Measures ==
		+
		+	I created an input csv file with columns: 'd1', 'd2', 'd3', 'd4','IMLB', 'IMRB', 'IMRT', 'IMLT', 'IMInterp'. <br>
		+
		+	1) I wanted to start simple, so I created a network to interpolate the IM values for a single event for this site group ('s385','s429','s431','s387','COO'). I could only train it well to deal with a portion of the 600,000 shared events for the site group. <br>
		+
		+	a) Neural network architecture <br>
		+
		+	8 inputs: the IM values of the 4 sites for an event and there distances to the interpolated site
		+	1 output: the interpolated IM value for that event
		+	3 hidden layers with softplus activation function
		+	I normalized the X inputs and y labels.
		+
		+	b) Results <br>
		+	i) Initial run <br>

@@ Line 272: / Line 272: @@
 [[File:abc6.png|500px]] <br>
 [[File:abc7.png|500px]] <br>
 I ended up sticking with this regularization. <br>
 Insight: one pattern I noticed was my interpolations tended to be better for smaller x values compared to larger ones, as the interpolated hazard curves tended to more closely follow the simulated hazard curves at lower x values, and then diverged. This may be because my data is denser at lower x values, and less dense at higher x values, so my model struggles more to infer probabilities at these higher values.

@@ Line 255: / Line 255: @@
 ) Since the large model was not very accurate, I decided to create 51 smaller models, one for each of the 51 x values. Each model is similar to the x = 0.50119 g model in (1). <br>
 a) The main difference between these smaller models and the model in (1) is I included extra inputs: the probability values of the 4 input sites and interp site at the x value below and above the x value I want to interpolate at. For instance, for my smaller model that interpolated x = 0.50119 g, I included 10 additional inputs: the LB, RB, LT, RT, interpsite probs at x = 0.39811 g and 0.63096 g. <br>
 Note that since including the velocity metrics increases the complexity of my network without much gain, I excluded them as inputs for these smaller models. <br>
@@ Line 274: / Line 272: @@
 [[File:abc6.png|500px]] <br>
 [[File:abc7.png|500px]] <br>
 I ended up sticking with this regularization. <br>
 Insight: one pattern I noticed was my interpolations tended to be better for smaller x values compared to larger ones, as the interpolated hazard curves tended to more closely follow the simulated hazard curves at lower x values, and then diverged. This may be because my data is denser at lower x values, and less dense at higher x values, so my model struggles more to infer probabilities at these higher values.

@@ Line 264: / Line 264: @@
 i) Initial run <br>
 Example of error plot <br>
-[[File:7301.png|400px]] <br>
+[[File:abc1.png|500px]] <br>
-[[File:7302.png|400px]] <br>
+[[File:abc2.png|500px]] <br>
-[[File:7303.png|400px]] <br>
+[[File:abc3.png|500px]] <br>
 ii) Tweaking parameters since the test loss curves did not closely resemble the training loss curves. <br>
 Increased batch size to 32 which made loss curve worse and created a hazard curve with unrealistic variations in probability.<br>
-[[File:7304.png|400px]] <br>
+[[File:abc4.png|500px]] <br>
 Reset batch size to 16, and added l2 regularizer with 0.0035 strength since this worked well for my smaller single model.<br>
-[[File:7305.png|400px]] <br>
+[[File:abc5.png|500px]] <br>
-[[File:7306.png|400px]] <br>
+[[File:abc6.png|500px]] <br>
-[[File:7307.png|400px]] <br>
+[[File:abc7.png|500px]] <br>
 I ended up sticking with this regularization. <br>
 Insight: one pattern I noticed was my interpolations tended to be better for smaller x values compared to larger ones, as the interpolated hazard curves tended to more closely follow the simulated hazard curves at lower x values, and then diverged. This may be because my data is denser at lower x values, and less dense at higher x values, so my model struggles more to infer probabilities at these higher values.

@@ Line 157: / Line 157: @@
 My model appeared to be overfitted since the training loss was significantly better than the testing loss, so I reduced the number of hidden layers from 5 to 3 and removed 3 repetitive data points from input csv. <br>
-[[File:a.png|300px]]
+[[File:a.png|500px]]
-[[File:b.png|300px]]
+[[File:b.png|500px]]
 Focusing on improving the loss curves, I decreased the batch size from 32 to 16 and reduced the number of epochs from 50 to 20. This resulted in my model converging too quickly (below). <br>
-[[File:batch1.png|300px]]
+[[File:batch1.png|500px]]
 <br> Keeping the batch size at 16, I increased the epochs from 20 to 35, allowing the model to train over a longer period. <br>
-[[File:batch2.png|300px]] <br>
+[[File:batch2.png|500px]] <br>
 ii) L2 regularizer to reduce overfitting: As my model was still overfitting, I used the L2 regularizer with 0.0035 strength. This increased my total test loss to 0.11 from 0.07, since L2 has a larger penalty for higher weights, but helped my test loss curve to more closely resemble training loss. <br>
-[[File:batch3.png|300px]] <br>
+[[File:batch3.png|500px]] <br>
 I also tried to implement dropout after each hidden layer to ensure my model was not dependent on any single neuron, but this resulted in my model losing too much data causing high variance in loss (not a smooth curve). For half the epochs, the testing loss was better than the training loss, which it shouldn't be. <br>
-[[File:batch4.png|300px]] <br>
+[[File:batch4.png|500px]] <br>
 iii) Changing learning rate of Adam optimizer: to further improve my model, I increased the learning rate from the default of 0.001 to 0.0016, which helped to decrease the test loss to 0.07. <br>
-[[File:batch5.png|300px]] <br>
+[[File:batch5.png|500px]] <br>
@@ Line 264: / Line 264: @@
 i) Initial run <br>
 Example of error plot <br>
-[[File:7301.png|300px]] <br>
+[[File:7301.png|400px]] <br>
-[[File:7302.png|300px]] <br>
+[[File:7302.png|400px]] <br>
-[[File:7303.png|300px]] <br>
+[[File:7303.png|400px]] <br>
 ii) Tweaking parameters since the test loss curves did not closely resemble the training loss curves. <br>
 Increased batch size to 32 which made loss curve worse and created a hazard curve with unrealistic variations in probability.<br>
-[[File:7304.png|300px]] <br>
+[[File:7304.png|400px]] <br>
 Reset batch size to 16, and added l2 regularizer with 0.0035 strength since this worked well for my smaller single model.<br>
-[[File:7305.png|300px]] <br>
+[[File:7305.png|400px]] <br>
-[[File:7306.png|300px]] <br>
+[[File:7306.png|400px]] <br>
-[[File:7307.png|300px]] <br>
+[[File:7307.png|400px]] <br>
 I ended up sticking with this regularization. <br>
 Insight: one pattern I noticed was my interpolations tended to be better for smaller x values compared to larger ones, as the interpolated hazard curves tended to more closely follow the simulated hazard curves at lower x values, and then diverged. This may be because my data is denser at lower x values, and less dense at higher x values, so my model struggles more to infer probabilities at these higher values.