EventNet Creation#
EventNet consists of 2 models: a masking model and a binary classification model. In order to mask potential event-related deformation in interferograms, a masking model is required. An event presence model is then trained in order automatically provide a probabilistic answer on whether there is an event in the masked image.
Thus, the four main steps are: 1. Create a Dataset with Simulated Wrapped Interferograms and their Masks 2. Train a masking model in the Simulated Masking Dataset 3. Create a Binary Classification Dataset of Masks and their Presence 4. Train a Binary Classification Model from the Mask/Presence Dataset
Imports#
[1]:
if "NOTEBOOK_INITIATED_FLAG" not in globals():
NOTEBOOK_INITIATED_FLAG = True
%cd ..
%pwd
import numpy as np
import matplotlib.pyplot as plt
from os import environ
environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
from insar_eventnet.config import SYNTHETIC_DIR
from insar_eventnet.io import make_simulated_dataset, split_dataset
from insar_eventnet.training import train
from insar_eventnet.inference import plot_results
from insar_eventnet.sarsim import gen_simulated_deformation, gen_sim_noise
plt.rcParams["figure.figsize"] = (20, 10)
/home/andrew/repos/AI-Event-Monitoring
INFO:tensorflow:Mixed precision compatibility check (mixed_float16): OK
Your GPU will likely run quickly with dtype policy mixed_float16 as it has compute capability of at least 7.0. Your GPU: NVIDIA GeForce RTX 2080 SUPER, compute capability 7.5
Simulated Events#
By using the gen_simulated_deformation and gen_sim_noise functions, we can create simulated interferogram datasets containing “Positives” which contain events and “Negatives” which contain noise. gen_simulated_deformation generates our “Positive” training examples and supports ‘quake’, ‘dyke’, and ‘sill’ event types. gen_sim_noise generates our “Negative” training images with either atmospheric and topographic noise or gaussian noise. The gaussian noise is primarily used as a negative representing unmasked incoherence.
Each of these functions returns 4 things: an unwrapped phase image, a deformation mask, a wrapped image, and a presence value. These are used as follows: - Unwrapped: Only used as a visual aid. Since gen_simulated_deformation contaminates this image, you must use the Okada class manually if you want an accurate displacement image. - Wrapped: This is the primary input for training the masking model. Wrapped images are turned into deformation masks. - Deformation Mask: This is used as a ground truth for training. Pixels are set to 1 where there is an event and its phase is greater than 2π. - Presence: [1] if the image contains a deformation event (it was from gen_simulated_deformation), otherwise its [0] (it was from gen_sim_noise).
[2]:
seed = 232323
tile_size = 512
event_type = "quake"
unwrapped_def, masked_def, wrapped_def, presence_def = gen_simulated_deformation(
seed=seed, tile_size=tile_size, event_type=event_type
)
unwrapped_mix, masked_mix, wrapped_mix, presence_mix = gen_sim_noise(
seed=seed, tile_size=tile_size
)
unwrapped_gsn, masked_gsn, wrapped_gsn, presence_gsn = gen_sim_noise(
seed=seed, tile_size=tile_size, gaussian_only=True
)
print(f"Deformation Presence: {presence_def}")
print(f"Mixed Noise Presence: {presence_mix}")
print(f"Gauss Noise Presence: {presence_gsn}")
_, [axs_unwrapped_def, axs_wrapped_def, axs_mask_def] = plt.subplots(
1, 3, sharex=True, sharey=True, tight_layout=True
)
_, [axs_unwrapped_mix, axs_wrapped_mix, axs_mask_mix] = plt.subplots(
1, 3, sharex=True, sharey=True, tight_layout=True
)
_, [axs_unwrapped_gsn, axs_wrapped_gsn, axs_mask_gsn] = plt.subplots(
1, 3, sharex=True, sharey=True, tight_layout=True
)
axs_unwrapped_def.set_title("Deformation Event")
axs_unwrapped_mix.set_title("Atmospheric/Topographic Noise")
axs_unwrapped_gsn.set_title("Gaussian Noise")
axs_unwrapped_def.imshow(unwrapped_def, origin="lower", cmap="jet")
axs_unwrapped_mix.imshow(unwrapped_mix, origin="lower", cmap="jet")
axs_unwrapped_gsn.imshow(unwrapped_gsn, origin="lower", cmap="jet")
axs_wrapped_def.imshow(wrapped_def, origin="lower", cmap="jet")
axs_wrapped_mix.imshow(wrapped_mix, origin="lower", cmap="jet")
axs_wrapped_gsn.imshow(wrapped_gsn, origin="lower", cmap="jet")
axs_mask_def.imshow(masked_def, origin="lower", cmap="jet", vmin=0.0, vmax=1.0)
axs_mask_mix.imshow(masked_mix, origin="lower", cmap="jet", vmin=0.0, vmax=1.0)
axs_mask_gsn.imshow(masked_gsn, origin="lower", cmap="jet", vmin=0.0, vmax=1.0)
plt.show()
Deformation Presence: [1]
Mixed Noise Presence: [0]
Gauss Noise Presence: [0]
Create Dataset of Simulated Events#
The make_simulated_dataset function is used to generate the datasets for training the masking model. It creates a dataset of specified size with a positive-to-negative ratio of 60/40, where 30% of the negatives are atmospheric noise and 10% are gaussian noise. The ratios can be modified easily in the source code, but the default ratios are what has worked best so far.
Note: while you can train a binary model with the same dataset that you train the masking model with, it is not recommended.
[3]:
name = "mask_model_dataset"
amount = (
2000 # The number of images to generate. ~1500 is sufficient for a decent model.
)
seed = 0 # The seed for the random number generators
tile_size = 512 # The size of the tiles to generate
crop_size = 512 # Crop size for model (same as tile_size for our U-Net currently)
split = 0.2 # The percentage of the dataset to use for validation
name, count, dir_name, distribution, dataset_info = make_simulated_dataset(
name, SYNTHETIC_DIR, amount, seed, tile_size, crop_size
)
dataset_path = SYNTHETIC_DIR.__str__() + "/" + dir_name
num_train, num_validation = split_dataset(dataset_path, split)
print("")
print(f"Dataset Path: {dataset_path}")
print(f"Train Amount: {num_train}")
print(f"Validation Amount: {num_validation}")
Generated 10 of 2000 simulated interferogram pairs.
Generated 20 of 2000 simulated interferogram pairs.
Generated 30 of 2000 simulated interferogram pairs.
Generated 40 of 2000 simulated interferogram pairs.
Generated 50 of 2000 simulated interferogram pairs.
Generated 60 of 2000 simulated interferogram pairs.
Generated 70 of 2000 simulated interferogram pairs.
Generated 80 of 2000 simulated interferogram pairs.
Generated 90 of 2000 simulated interferogram pairs.
Generated 100 of 2000 simulated interferogram pairs.
Generated 110 of 2000 simulated interferogram pairs.
Generated 120 of 2000 simulated interferogram pairs.
Generated 130 of 2000 simulated interferogram pairs.
Generated 140 of 2000 simulated interferogram pairs.
Generated 150 of 2000 simulated interferogram pairs.
Generated 160 of 2000 simulated interferogram pairs.
Generated 170 of 2000 simulated interferogram pairs.
Generated 180 of 2000 simulated interferogram pairs.
Generated 190 of 2000 simulated interferogram pairs.
Generated 200 of 2000 simulated interferogram pairs.
Generated 210 of 2000 simulated interferogram pairs.
Generated 220 of 2000 simulated interferogram pairs.
Generated 230 of 2000 simulated interferogram pairs.
Generated 240 of 2000 simulated interferogram pairs.
Generated 250 of 2000 simulated interferogram pairs.
Generated 260 of 2000 simulated interferogram pairs.
Generated 270 of 2000 simulated interferogram pairs.
Generated 280 of 2000 simulated interferogram pairs.
Generated 290 of 2000 simulated interferogram pairs.
Generated 300 of 2000 simulated interferogram pairs.
Generated 310 of 2000 simulated interferogram pairs.
Generated 320 of 2000 simulated interferogram pairs.
Generated 330 of 2000 simulated interferogram pairs.
Generated 340 of 2000 simulated interferogram pairs.
Generated 350 of 2000 simulated interferogram pairs.
Generated 360 of 2000 simulated interferogram pairs.
Generated 370 of 2000 simulated interferogram pairs.
Generated 380 of 2000 simulated interferogram pairs.
Generated 390 of 2000 simulated interferogram pairs.
Generated 400 of 2000 simulated interferogram pairs.
Generated 410 of 2000 simulated interferogram pairs.
Generated 420 of 2000 simulated interferogram pairs.
Generated 430 of 2000 simulated interferogram pairs.
Generated 440 of 2000 simulated interferogram pairs.
Generated 450 of 2000 simulated interferogram pairs.
Generated 460 of 2000 simulated interferogram pairs.
Generated 470 of 2000 simulated interferogram pairs.
Generated 480 of 2000 simulated interferogram pairs.
Generated 490 of 2000 simulated interferogram pairs.
Generated 500 of 2000 simulated interferogram pairs.
Generated 510 of 2000 simulated interferogram pairs.
Generated 520 of 2000 simulated interferogram pairs.
Generated 530 of 2000 simulated interferogram pairs.
Generated 540 of 2000 simulated interferogram pairs.
Generated 550 of 2000 simulated interferogram pairs.
Generated 560 of 2000 simulated interferogram pairs.
Generated 570 of 2000 simulated interferogram pairs.
Generated 580 of 2000 simulated interferogram pairs.
Generated 590 of 2000 simulated interferogram pairs.
Generated 600 of 2000 simulated interferogram pairs.
Generated 610 of 2000 simulated interferogram pairs.
Generated 620 of 2000 simulated interferogram pairs.
Generated 630 of 2000 simulated interferogram pairs.
Generated 640 of 2000 simulated interferogram pairs.
Generated 650 of 2000 simulated interferogram pairs.
Generated 660 of 2000 simulated interferogram pairs.
Generated 670 of 2000 simulated interferogram pairs.
Generated 680 of 2000 simulated interferogram pairs.
Generated 690 of 2000 simulated interferogram pairs.
Generated 700 of 2000 simulated interferogram pairs.
Generated 710 of 2000 simulated interferogram pairs.
Generated 720 of 2000 simulated interferogram pairs.
Generated 730 of 2000 simulated interferogram pairs.
Generated 740 of 2000 simulated interferogram pairs.
Generated 750 of 2000 simulated interferogram pairs.
Generated 760 of 2000 simulated interferogram pairs.
Generated 770 of 2000 simulated interferogram pairs.
Generated 780 of 2000 simulated interferogram pairs.
Generated 790 of 2000 simulated interferogram pairs.
Generated 800 of 2000 simulated interferogram pairs.
Generated 810 of 2000 simulated interferogram pairs.
Generated 820 of 2000 simulated interferogram pairs.
Generated 830 of 2000 simulated interferogram pairs.
Generated 840 of 2000 simulated interferogram pairs.
Generated 850 of 2000 simulated interferogram pairs.
Generated 860 of 2000 simulated interferogram pairs.
Generated 870 of 2000 simulated interferogram pairs.
Generated 880 of 2000 simulated interferogram pairs.
Generated 890 of 2000 simulated interferogram pairs.
Generated 900 of 2000 simulated interferogram pairs.
Generated 910 of 2000 simulated interferogram pairs.
Generated 920 of 2000 simulated interferogram pairs.
Generated 930 of 2000 simulated interferogram pairs.
Generated 940 of 2000 simulated interferogram pairs.
Generated 950 of 2000 simulated interferogram pairs.
Generated 960 of 2000 simulated interferogram pairs.
Generated 970 of 2000 simulated interferogram pairs.
Generated 980 of 2000 simulated interferogram pairs.
Generated 990 of 2000 simulated interferogram pairs.
Generated 1000 of 2000 simulated interferogram pairs.
Generated 1010 of 2000 simulated interferogram pairs.
Generated 1020 of 2000 simulated interferogram pairs.
Generated 1030 of 2000 simulated interferogram pairs.
Generated 1040 of 2000 simulated interferogram pairs.
Generated 1050 of 2000 simulated interferogram pairs.
Generated 1060 of 2000 simulated interferogram pairs.
Generated 1070 of 2000 simulated interferogram pairs.
Generated 1080 of 2000 simulated interferogram pairs.
Generated 1090 of 2000 simulated interferogram pairs.
Generated 1100 of 2000 simulated interferogram pairs.
Generated 1110 of 2000 simulated interferogram pairs.
Generated 1120 of 2000 simulated interferogram pairs.
Generated 1130 of 2000 simulated interferogram pairs.
Generated 1140 of 2000 simulated interferogram pairs.
Generated 1150 of 2000 simulated interferogram pairs.
Generated 1160 of 2000 simulated interferogram pairs.
Generated 1170 of 2000 simulated interferogram pairs.
Generated 1180 of 2000 simulated interferogram pairs.
Generated 1190 of 2000 simulated interferogram pairs.
Generated 1200 of 2000 simulated interferogram pairs.
Generated 1210 of 2000 simulated interferogram pairs.
Generated 1220 of 2000 simulated interferogram pairs.
Generated 1230 of 2000 simulated interferogram pairs.
Generated 1240 of 2000 simulated interferogram pairs.
Generated 1250 of 2000 simulated interferogram pairs.
Generated 1260 of 2000 simulated interferogram pairs.
Generated 1270 of 2000 simulated interferogram pairs.
Generated 1280 of 2000 simulated interferogram pairs.
Generated 1290 of 2000 simulated interferogram pairs.
Generated 1300 of 2000 simulated interferogram pairs.
Generated 1310 of 2000 simulated interferogram pairs.
Generated 1320 of 2000 simulated interferogram pairs.
Generated 1330 of 2000 simulated interferogram pairs.
Generated 1340 of 2000 simulated interferogram pairs.
Generated 1350 of 2000 simulated interferogram pairs.
Generated 1360 of 2000 simulated interferogram pairs.
Generated 1370 of 2000 simulated interferogram pairs.
Generated 1380 of 2000 simulated interferogram pairs.
Generated 1390 of 2000 simulated interferogram pairs.
Generated 1400 of 2000 simulated interferogram pairs.
Generated 1410 of 2000 simulated interferogram pairs.
Generated 1420 of 2000 simulated interferogram pairs.
Generated 1430 of 2000 simulated interferogram pairs.
Generated 1440 of 2000 simulated interferogram pairs.
Generated 1450 of 2000 simulated interferogram pairs.
Generated 1460 of 2000 simulated interferogram pairs.
Generated 1470 of 2000 simulated interferogram pairs.
Generated 1480 of 2000 simulated interferogram pairs.
Generated 1490 of 2000 simulated interferogram pairs.
Generated 1500 of 2000 simulated interferogram pairs.
Generated 1510 of 2000 simulated interferogram pairs.
Generated 1520 of 2000 simulated interferogram pairs.
Generated 1530 of 2000 simulated interferogram pairs.
Generated 1540 of 2000 simulated interferogram pairs.
Generated 1550 of 2000 simulated interferogram pairs.
Generated 1560 of 2000 simulated interferogram pairs.
Generated 1570 of 2000 simulated interferogram pairs.
Generated 1580 of 2000 simulated interferogram pairs.
Generated 1590 of 2000 simulated interferogram pairs.
Generated 1600 of 2000 simulated interferogram pairs.
Generated 1610 of 2000 simulated interferogram pairs.
Generated 1620 of 2000 simulated interferogram pairs.
Generated 1630 of 2000 simulated interferogram pairs.
Generated 1640 of 2000 simulated interferogram pairs.
Generated 1650 of 2000 simulated interferogram pairs.
Generated 1660 of 2000 simulated interferogram pairs.
Generated 1670 of 2000 simulated interferogram pairs.
Generated 1680 of 2000 simulated interferogram pairs.
Generated 1690 of 2000 simulated interferogram pairs.
Generated 1700 of 2000 simulated interferogram pairs.
Generated 1710 of 2000 simulated interferogram pairs.
Generated 1720 of 2000 simulated interferogram pairs.
Generated 1730 of 2000 simulated interferogram pairs.
Generated 1740 of 2000 simulated interferogram pairs.
Generated 1750 of 2000 simulated interferogram pairs.
Generated 1760 of 2000 simulated interferogram pairs.
Generated 1770 of 2000 simulated interferogram pairs.
Generated 1780 of 2000 simulated interferogram pairs.
Generated 1790 of 2000 simulated interferogram pairs.
Generated 1800 of 2000 simulated interferogram pairs.
Generated 1810 of 2000 simulated interferogram pairs.
Generated 1820 of 2000 simulated interferogram pairs.
Generated 1830 of 2000 simulated interferogram pairs.
Generated 1840 of 2000 simulated interferogram pairs.
Generated 1850 of 2000 simulated interferogram pairs.
Generated 1860 of 2000 simulated interferogram pairs.
Generated 1870 of 2000 simulated interferogram pairs.
Generated 1880 of 2000 simulated interferogram pairs.
Generated 1890 of 2000 simulated interferogram pairs.
Generated 1900 of 2000 simulated interferogram pairs.
Generated 1910 of 2000 simulated interferogram pairs.
Generated 1920 of 2000 simulated interferogram pairs.
Generated 1930 of 2000 simulated interferogram pairs.
Generated 1940 of 2000 simulated interferogram pairs.
Generated 1950 of 2000 simulated interferogram pairs.
Generated 1960 of 2000 simulated interferogram pairs.
Generated 1970 of 2000 simulated interferogram pairs.
Generated 1980 of 2000 simulated interferogram pairs.
Generated 1990 of 2000 simulated interferogram pairs.
Generated 2000 of 2000 simulated interferogram pairs.
Dataset Path: /home/andrew/repos/AI-Event-Monitoring/data/working/synthetic/mask_model_dataset_amount2000_seed512283
Train Amount: 1602
Validation Amount: 398
Train UNet Masking Model#
The main model we use for masking is a form of U-Net. It is also possible to use a ResNet-like model by switching “unet” to “resnet”, but the “unet” tends to produce better results in less time. The default hyperparameters here should be sufficient for getting good results.
[4]:
model_name = "mask_modelV36"
model_type = "unet"
dataset_path = dataset_path
input_shape = 512
epochs = 15
filters = 64
batch_size = 1
learning_rate = 1e-4
use_wandb = False
using_aws = False
using_jupyter = True
logs_dir = ""
mask_model, mask_history = train(
model_name,
dataset_path,
model_type,
input_shape,
epochs,
filters,
batch_size,
learning_rate,
use_wandb,
using_aws,
using_jupyter,
logs_dir,
)
Epoch 1/15
1602/1602 [==============================] - ETA: 0s - loss: 0.0757 - mean_squared_error: 0.1515 - mean_absolute_error: 0.3034
Epoch 1: val_loss improved from inf to 0.03571, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 65s 34ms/step - loss: 0.0757 - mean_squared_error: 0.1515 - mean_absolute_error: 0.3034 - val_loss: 0.0357 - val_mean_squared_error: 0.0714 - val_mean_absolute_error: 0.1696
Epoch 2/15
1602/1602 [==============================] - ETA: 0s - loss: 0.0303 - mean_squared_error: 0.0606 - mean_absolute_error: 0.1406
Epoch 2: val_loss improved from 0.03571 to 0.02510, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 53s 33ms/step - loss: 0.0303 - mean_squared_error: 0.0606 - mean_absolute_error: 0.1406 - val_loss: 0.0251 - val_mean_squared_error: 0.0502 - val_mean_absolute_error: 0.1331
Epoch 3/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0297 - mean_squared_error: 0.0594 - mean_absolute_error: 0.1354
Epoch 3: val_loss improved from 0.02510 to 0.02078, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 52s 33ms/step - loss: 0.0297 - mean_squared_error: 0.0594 - mean_absolute_error: 0.1353 - val_loss: 0.0208 - val_mean_squared_error: 0.0416 - val_mean_absolute_error: 0.0879
Epoch 4/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0213 - mean_squared_error: 0.0426 - mean_absolute_error: 0.1024
Epoch 4: val_loss did not improve from 0.02078
1602/1602 [==============================] - 53s 33ms/step - loss: 0.0213 - mean_squared_error: 0.0425 - mean_absolute_error: 0.1024 - val_loss: 0.0209 - val_mean_squared_error: 0.0418 - val_mean_absolute_error: 0.0930
Epoch 5/15
1602/1602 [==============================] - ETA: 0s - loss: 0.0173 - mean_squared_error: 0.0347 - mean_absolute_error: 0.0848
Epoch 5: val_loss improved from 0.02078 to 0.01769, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 50s 31ms/step - loss: 0.0173 - mean_squared_error: 0.0347 - mean_absolute_error: 0.0848 - val_loss: 0.0177 - val_mean_squared_error: 0.0354 - val_mean_absolute_error: 0.0825
Epoch 6/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0160 - mean_squared_error: 0.0320 - mean_absolute_error: 0.0778
Epoch 6: val_loss did not improve from 0.01769
1602/1602 [==============================] - 48s 30ms/step - loss: 0.0160 - mean_squared_error: 0.0320 - mean_absolute_error: 0.0777 - val_loss: 0.0350 - val_mean_squared_error: 0.0701 - val_mean_absolute_error: 0.1694
Epoch 7/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0125 - mean_squared_error: 0.0251 - mean_absolute_error: 0.0639
Epoch 7: val_loss improved from 0.01769 to 0.01671, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 50s 31ms/step - loss: 0.0125 - mean_squared_error: 0.0251 - mean_absolute_error: 0.0639 - val_loss: 0.0167 - val_mean_squared_error: 0.0334 - val_mean_absolute_error: 0.0699
Epoch 8/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0094 - mean_squared_error: 0.0187 - mean_absolute_error: 0.0487
Epoch 8: val_loss did not improve from 0.01671
1602/1602 [==============================] - 49s 30ms/step - loss: 0.0094 - mean_squared_error: 0.0187 - mean_absolute_error: 0.0487 - val_loss: 0.0169 - val_mean_squared_error: 0.0339 - val_mean_absolute_error: 0.0588
Epoch 9/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0078 - mean_squared_error: 0.0156 - mean_absolute_error: 0.0414
Epoch 9: val_loss improved from 0.01671 to 0.01618, saving model to ./data/output/models/checkpoints/mask_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/mask_modelV36/assets
1602/1602 [==============================] - 50s 31ms/step - loss: 0.0078 - mean_squared_error: 0.0156 - mean_absolute_error: 0.0414 - val_loss: 0.0162 - val_mean_squared_error: 0.0324 - val_mean_absolute_error: 0.0597
Epoch 10/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0063 - mean_squared_error: 0.0125 - mean_absolute_error: 0.0335
Epoch 10: val_loss did not improve from 0.01618
1602/1602 [==============================] - 49s 30ms/step - loss: 0.0063 - mean_squared_error: 0.0125 - mean_absolute_error: 0.0335 - val_loss: 0.0184 - val_mean_squared_error: 0.0367 - val_mean_absolute_error: 0.0868
Epoch 11/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0053 - mean_squared_error: 0.0106 - mean_absolute_error: 0.0292
Epoch 11: val_loss did not improve from 0.01618
1602/1602 [==============================] - 62s 39ms/step - loss: 0.0053 - mean_squared_error: 0.0106 - mean_absolute_error: 0.0292 - val_loss: 0.0171 - val_mean_squared_error: 0.0341 - val_mean_absolute_error: 0.0534
Epoch 12/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0048 - mean_squared_error: 0.0095 - mean_absolute_error: 0.0268
Epoch 12: val_loss did not improve from 0.01618
1602/1602 [==============================] - 55s 34ms/step - loss: 0.0048 - mean_squared_error: 0.0095 - mean_absolute_error: 0.0268 - val_loss: 0.0173 - val_mean_squared_error: 0.0346 - val_mean_absolute_error: 0.0518
Epoch 13/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0044 - mean_squared_error: 0.0088 - mean_absolute_error: 0.0253
Epoch 13: val_loss did not improve from 0.01618
1602/1602 [==============================] - 58s 36ms/step - loss: 0.0044 - mean_squared_error: 0.0088 - mean_absolute_error: 0.0253 - val_loss: 0.0169 - val_mean_squared_error: 0.0339 - val_mean_absolute_error: 0.0638
Epoch 14/15
1602/1602 [==============================] - ETA: 0s - loss: 0.0059 - mean_squared_error: 0.0118 - mean_absolute_error: 0.0297
Epoch 14: val_loss did not improve from 0.01618
1602/1602 [==============================] - 55s 34ms/step - loss: 0.0059 - mean_squared_error: 0.0118 - mean_absolute_error: 0.0297 - val_loss: 0.0175 - val_mean_squared_error: 0.0350 - val_mean_absolute_error: 0.0563
Epoch 15/15
1601/1602 [============================>.] - ETA: 0s - loss: 0.0034 - mean_squared_error: 0.0068 - mean_absolute_error: 0.0208
Epoch 15: val_loss did not improve from 0.01618
1602/1602 [==============================] - 55s 35ms/step - loss: 0.0034 - mean_squared_error: 0.0068 - mean_absolute_error: 0.0208 - val_loss: 0.0172 - val_mean_squared_error: 0.0345 - val_mean_absolute_error: 0.0489
INFO:tensorflow:Assets written to: ./data/output/models/mask_modelV36/assets
[5]:
def print_model_info(model, history):
losses = history.history["loss"]
val_losses = history.history["val_loss"]
min_loss = min(losses)
min_val_loss = min(val_losses)
min_loss_at = losses.index(min_loss)
min_val_loss_at = val_losses.index(min_val_loss)
summary_list = []
model.summary(print_fn=lambda x: summary_list.append(" " + x))
summary_string = "\n".join(summary_list[1:])
print(
f"\n",
f"Model Name: {model_name}\n",
f"Model Type: {model_type}\n",
f"\n",
f"Minimum Training Loss: {min_loss: 0.15f} at epoch {min_loss_at}.\n",
f"Minimum Validation Loss: {min_val_loss: 0.15f} at epoch {min_val_loss_at}.\n"
f"\n",
f"All Training Losses:\n {losses}\n",
f"\n",
f"All Validation Losses:\n {val_losses}\n",
f"\n",
f"Model Parameters from History:\n {history.params}\n",
f"\n",
f"Model Summary:\n{summary_string}\n",
)
print_model_info(mask_model, mask_history)
Model Name: mask_modelV36
Model Type: unet
Minimum Training Loss: 0.003421352710575 at epoch 14.
Minimum Validation Loss: 0.016181165352464 at epoch 8.
All Training Losses:
[0.07569071650505066, 0.030308237299323082, 0.029678167775273323, 0.021265355870127678, 0.01732829585671425, 0.015977075323462486, 0.012531604617834091, 0.009357718750834465, 0.007819187827408314, 0.006264141295105219, 0.005277558695524931, 0.004769966006278992, 0.004400080535560846, 0.005891966633498669, 0.0034213527105748653]
All Validation Losses:
[0.03570585325360298, 0.025097331032156944, 0.0207799281924963, 0.020895613357424736, 0.017689848318696022, 0.035025373101234436, 0.016706326976418495, 0.016934551298618317, 0.016181165352463722, 0.018362348899245262, 0.01706208847463131, 0.01730268821120262, 0.016937458887696266, 0.017519645392894745, 0.017228467389941216]
Model Parameters from History:
{'verbose': 1, 'epochs': 15, 'steps': 1602}
Model Summary:
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 512, 512, 1 0 []
)]
conv2d (Conv2D) (None, 256, 256, 64 640 ['input_1[0][0]']
)
conv2d_1 (Conv2D) (None, 256, 256, 64 36928 ['conv2d[0][0]']
)
dropout (Dropout) (None, 256, 256, 64 0 ['conv2d_1[0][0]']
)
conv2d_2 (Conv2D) (None, 128, 128, 12 73856 ['dropout[0][0]']
8)
conv2d_3 (Conv2D) (None, 128, 128, 12 147584 ['conv2d_2[0][0]']
8)
conv2d_4 (Conv2D) (None, 64, 64, 256) 295168 ['conv2d_3[0][0]']
conv2d_5 (Conv2D) (None, 64, 64, 256) 590080 ['conv2d_4[0][0]']
conv2d_6 (Conv2D) (None, 32, 32, 512) 1180160 ['conv2d_5[0][0]']
conv2d_7 (Conv2D) (None, 32, 32, 512) 2359808 ['conv2d_6[0][0]']
conv2d_8 (Conv2D) (None, 16, 16, 1024 4719616 ['conv2d_7[0][0]']
)
conv2d_9 (Conv2D) (None, 16, 16, 1024 9438208 ['conv2d_8[0][0]']
)
dropout_1 (Dropout) (None, 16, 16, 1024 0 ['conv2d_9[0][0]']
)
conv2d_transpose (Conv2DTransp (None, 32, 32, 512) 2097664 ['dropout_1[0][0]']
ose)
conv2d_10 (Conv2D) (None, 32, 32, 512) 2359808 ['conv2d_transpose[0][0]']
conv2d_11 (Conv2D) (None, 32, 32, 512) 2359808 ['conv2d_10[0][0]']
concatenate (Concatenate) (None, 32, 32, 1024 0 ['conv2d_11[0][0]',
) 'conv2d_7[0][0]']
conv2d_transpose_1 (Conv2DTran (None, 64, 64, 256) 1048832 ['concatenate[0][0]']
spose)
conv2d_12 (Conv2D) (None, 64, 64, 256) 590080 ['conv2d_transpose_1[0][0]']
conv2d_13 (Conv2D) (None, 64, 64, 256) 590080 ['conv2d_12[0][0]']
concatenate_1 (Concatenate) (None, 64, 64, 512) 0 ['conv2d_13[0][0]',
'conv2d_5[0][0]']
conv2d_transpose_2 (Conv2DTran (None, 128, 128, 12 262272 ['concatenate_1[0][0]']
spose) 8)
conv2d_14 (Conv2D) (None, 128, 128, 12 147584 ['conv2d_transpose_2[0][0]']
8)
conv2d_15 (Conv2D) (None, 128, 128, 12 147584 ['conv2d_14[0][0]']
8)
concatenate_2 (Concatenate) (None, 128, 128, 25 0 ['conv2d_15[0][0]',
6) 'conv2d_3[0][0]']
conv2d_transpose_3 (Conv2DTran (None, 256, 256, 64 65600 ['concatenate_2[0][0]']
spose) )
conv2d_16 (Conv2D) (None, 256, 256, 64 36928 ['conv2d_transpose_3[0][0]']
)
conv2d_17 (Conv2D) (None, 256, 256, 64 36928 ['conv2d_16[0][0]']
)
concatenate_3 (Concatenate) (None, 256, 256, 12 0 ['conv2d_17[0][0]',
8) 'dropout[0][0]']
conv2d_transpose_4 (Conv2DTran (None, 512, 512, 64 32832 ['concatenate_3[0][0]']
spose) )
last_layer (Conv2D) (None, 512, 512, 1) 65 ['conv2d_transpose_4[0][0]']
activation (Activation) (None, 512, 512, 1) 0 ['last_layer[0][0]']
==================================================================================================
Total params: 28,618,113
Trainable params: 28,618,113
Non-trainable params: 0
__________________________________________________________________________________________________
Run Inference on a Simulated Event with the Masking Model#
Now that we have a masking model, we can do a basic evaluation of our model using the simulated data. As you will see, the U-Net outputs pixel values that can range from slightly below 0 to slightly above 1. These values can be roughly thought of as a confidence in whether the pixel represents deformation from an event.
[6]:
def eval_mask_model(
model, event_type, seed, noise_only, round_mask, zero_output, tile_size, crop_size
):
if not noise_only:
unwrapped, mask, wrapped, presence = gen_simulated_deformation(
seed=seed, tile_size=tile_size, event_type=event_type
)
else:
unwrapped, mask, wrapped, presence = gen_sim_noise(
seed=seed, tile_size=tile_size
)
wrapped = wrapped.reshape((1, tile_size, tile_size, 1))
mask_pred = np.float32(model.predict(wrapped).reshape((crop_size, crop_size)))
wrapped = wrapped.reshape((tile_size, tile_size))
mask_pred_rounded = np.copy(mask_pred)
if round_mask:
tolerance = 0.7
round_up = mask_pred >= tolerance
round_down = mask_pred < tolerance
mask_pred_rounded[round_up] = 1
mask_pred_rounded[round_down] = 0
if zero_output:
zeros = x == 0
mask[zeros] = 0
mask_pred_rounded[zeros] = 0
return wrapped, mask, mask_pred, mask_pred_rounded
event_type = "quake"
seed = 0
noise_only = False
round_mask = True
zero_output = False
wrapped, mask, mask_pred, mask_pred_rounded = eval_mask_model(
mask_model,
event_type,
seed,
noise_only,
round_mask,
zero_output,
tile_size,
crop_size,
)
plot_results(wrapped, mask, mask_pred, mask_pred_rounded)
Create Dataset of Simulated Events with Masks from the UNet#
As mentioned before, it is possible to train the binary classification model on any dataset generated by make_simulated_dataset; however, each masking model may have slight quirks and imperfections which can result in worse performance when a binary model is expecting perfection. Thus, it is better to use the masking model to generate the masks for training the binary classification model. This is done by including the masking model’s path in the function arguments.
[7]:
name = "classification_model_dataset"
model_path = "data/output/models/checkpoints/" + model_name
amount = 1000 # A size of ~1000 usually produces good results.
split = 0.1
name, count, dir_name, _, _ = make_simulated_dataset(
name, SYNTHETIC_DIR, amount, seed, tile_size, crop_size, model_path=model_path
)
dataset_path = SYNTHETIC_DIR.__str__() + "/" + dir_name
num_train, num_validation = split_dataset(dataset_path, split)
print("")
print(f"Dataset Path: {dataset_path}")
print(f"Train Amount: {num_train}")
print(f"Validation Amount: {num_validation}")
Generated 10 of 1000 simulated interferogram pairs.
Generated 20 of 1000 simulated interferogram pairs.
Generated 30 of 1000 simulated interferogram pairs.
Generated 40 of 1000 simulated interferogram pairs.
Generated 50 of 1000 simulated interferogram pairs.
Generated 60 of 1000 simulated interferogram pairs.
Generated 70 of 1000 simulated interferogram pairs.
Generated 80 of 1000 simulated interferogram pairs.
Generated 90 of 1000 simulated interferogram pairs.
Generated 100 of 1000 simulated interferogram pairs.
Generated 110 of 1000 simulated interferogram pairs.
Generated 120 of 1000 simulated interferogram pairs.
Generated 130 of 1000 simulated interferogram pairs.
Generated 140 of 1000 simulated interferogram pairs.
Generated 150 of 1000 simulated interferogram pairs.
Generated 160 of 1000 simulated interferogram pairs.
Generated 170 of 1000 simulated interferogram pairs.
Generated 180 of 1000 simulated interferogram pairs.
Generated 190 of 1000 simulated interferogram pairs.
Generated 200 of 1000 simulated interferogram pairs.
Generated 210 of 1000 simulated interferogram pairs.
Generated 220 of 1000 simulated interferogram pairs.
Generated 230 of 1000 simulated interferogram pairs.
Generated 240 of 1000 simulated interferogram pairs.
Generated 250 of 1000 simulated interferogram pairs.
Generated 260 of 1000 simulated interferogram pairs.
Generated 270 of 1000 simulated interferogram pairs.
Generated 280 of 1000 simulated interferogram pairs.
Generated 290 of 1000 simulated interferogram pairs.
Generated 300 of 1000 simulated interferogram pairs.
Generated 310 of 1000 simulated interferogram pairs.
Generated 320 of 1000 simulated interferogram pairs.
Generated 330 of 1000 simulated interferogram pairs.
Generated 340 of 1000 simulated interferogram pairs.
Generated 350 of 1000 simulated interferogram pairs.
Generated 360 of 1000 simulated interferogram pairs.
Generated 370 of 1000 simulated interferogram pairs.
Generated 380 of 1000 simulated interferogram pairs.
Generated 390 of 1000 simulated interferogram pairs.
Generated 400 of 1000 simulated interferogram pairs.
Generated 410 of 1000 simulated interferogram pairs.
Generated 420 of 1000 simulated interferogram pairs.
Generated 430 of 1000 simulated interferogram pairs.
Generated 440 of 1000 simulated interferogram pairs.
Generated 450 of 1000 simulated interferogram pairs.
Generated 460 of 1000 simulated interferogram pairs.
Generated 470 of 1000 simulated interferogram pairs.
Generated 480 of 1000 simulated interferogram pairs.
Generated 490 of 1000 simulated interferogram pairs.
Generated 500 of 1000 simulated interferogram pairs.
Generated 510 of 1000 simulated interferogram pairs.
Generated 520 of 1000 simulated interferogram pairs.
Generated 530 of 1000 simulated interferogram pairs.
Generated 540 of 1000 simulated interferogram pairs.
Generated 550 of 1000 simulated interferogram pairs.
Generated 560 of 1000 simulated interferogram pairs.
Generated 570 of 1000 simulated interferogram pairs.
Generated 580 of 1000 simulated interferogram pairs.
Generated 590 of 1000 simulated interferogram pairs.
Generated 600 of 1000 simulated interferogram pairs.
Generated 610 of 1000 simulated interferogram pairs.
Generated 620 of 1000 simulated interferogram pairs.
Generated 630 of 1000 simulated interferogram pairs.
Generated 640 of 1000 simulated interferogram pairs.
Generated 650 of 1000 simulated interferogram pairs.
Generated 660 of 1000 simulated interferogram pairs.
Generated 670 of 1000 simulated interferogram pairs.
Generated 680 of 1000 simulated interferogram pairs.
Generated 690 of 1000 simulated interferogram pairs.
Generated 700 of 1000 simulated interferogram pairs.
Generated 710 of 1000 simulated interferogram pairs.
Generated 720 of 1000 simulated interferogram pairs.
Generated 730 of 1000 simulated interferogram pairs.
Generated 740 of 1000 simulated interferogram pairs.
Generated 750 of 1000 simulated interferogram pairs.
Generated 760 of 1000 simulated interferogram pairs.
Generated 770 of 1000 simulated interferogram pairs.
Generated 780 of 1000 simulated interferogram pairs.
Generated 790 of 1000 simulated interferogram pairs.
Generated 800 of 1000 simulated interferogram pairs.
Generated 810 of 1000 simulated interferogram pairs.
Generated 820 of 1000 simulated interferogram pairs.
Generated 830 of 1000 simulated interferogram pairs.
Generated 840 of 1000 simulated interferogram pairs.
Generated 850 of 1000 simulated interferogram pairs.
Generated 860 of 1000 simulated interferogram pairs.
Generated 870 of 1000 simulated interferogram pairs.
Generated 880 of 1000 simulated interferogram pairs.
Generated 890 of 1000 simulated interferogram pairs.
Generated 900 of 1000 simulated interferogram pairs.
Generated 910 of 1000 simulated interferogram pairs.
Generated 920 of 1000 simulated interferogram pairs.
Generated 930 of 1000 simulated interferogram pairs.
Generated 940 of 1000 simulated interferogram pairs.
Generated 950 of 1000 simulated interferogram pairs.
Generated 960 of 1000 simulated interferogram pairs.
Generated 970 of 1000 simulated interferogram pairs.
Generated 980 of 1000 simulated interferogram pairs.
Generated 990 of 1000 simulated interferogram pairs.
Generated 1000 of 1000 simulated interferogram pairs.
Dataset Path: /home/andrew/repos/AI-Event-Monitoring/data/working/synthetic/classification_model_dataset_amount1000_seed890963
Train Amount: 896
Validation Amount: 104
Train the Binary Classification Model#
Finally, we can train the binary classification model. This model is simple, being made up of only a couple convolutional layers and a dense layer, so it trains very fast. The default hyperparameters here should be sufficient for achieving satisfactory results.
[8]:
model_name_bin = "classification_modelV36"
model_type = "eventnet"
dataset_path = dataset_path
input_shape = 512
epochs = 5
filters = 64
batch_size = 1
learning_rate = 5e-3
use_wandb = False
using_aws = False
using_jupyter = True
logs_dir = ""
binary_model, binary_history = train(
model_name_bin,
dataset_path,
model_type,
input_shape,
epochs,
filters,
batch_size,
learning_rate,
use_wandb,
using_aws,
using_jupyter,
logs_dir,
)
Epoch 1/5
893/896 [============================>.] - ETA: 0s - loss: 0.1261 - acc: 0.8365 - mean_absolute_error: 0.2982
Epoch 1: val_loss improved from inf to 0.08135, saving model to ./data/output/models/checkpoints/classification_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/classification_modelV36/assets
896/896 [==============================] - 9s 9ms/step - loss: 0.1257 - acc: 0.8371 - mean_absolute_error: 0.2973 - val_loss: 0.0814 - val_acc: 1.0000 - val_mean_absolute_error: 0.2092
Epoch 2/5
890/896 [============================>.] - ETA: 0s - loss: 0.0603 - acc: 1.0000 - mean_absolute_error: 0.1729
Epoch 2: val_loss improved from 0.08135 to 0.05094, saving model to ./data/output/models/checkpoints/classification_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/classification_modelV36/assets
896/896 [==============================] - 7s 8ms/step - loss: 0.0600 - acc: 1.0000 - mean_absolute_error: 0.1723 - val_loss: 0.0509 - val_acc: 1.0000 - val_mean_absolute_error: 0.1589
Epoch 3/5
894/896 [============================>.] - ETA: 0s - loss: 0.0388 - acc: 1.0000 - mean_absolute_error: 0.1351
Epoch 3: val_loss improved from 0.05094 to 0.03424, saving model to ./data/output/models/checkpoints/classification_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/classification_modelV36/assets
896/896 [==============================] - 7s 8ms/step - loss: 0.0388 - acc: 1.0000 - mean_absolute_error: 0.1351 - val_loss: 0.0342 - val_acc: 1.0000 - val_mean_absolute_error: 0.1279
Epoch 4/5
895/896 [============================>.] - ETA: 0s - loss: 0.0270 - acc: 1.0000 - mean_absolute_error: 0.1107
Epoch 4: val_loss improved from 0.03424 to 0.02464, saving model to ./data/output/models/checkpoints/classification_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/classification_modelV36/assets
896/896 [==============================] - 7s 8ms/step - loss: 0.0270 - acc: 1.0000 - mean_absolute_error: 0.1109 - val_loss: 0.0246 - val_acc: 1.0000 - val_mean_absolute_error: 0.1075
Epoch 5/5
893/896 [============================>.] - ETA: 0s - loss: 0.0199 - acc: 1.0000 - mean_absolute_error: 0.0945
Epoch 5: val_loss improved from 0.02464 to 0.01866, saving model to ./data/output/models/checkpoints/classification_modelV36
INFO:tensorflow:Assets written to: ./data/output/models/checkpoints/classification_modelV36/assets
896/896 [==============================] - 6s 7ms/step - loss: 0.0199 - acc: 1.0000 - mean_absolute_error: 0.0944 - val_loss: 0.0187 - val_acc: 1.0000 - val_mean_absolute_error: 0.0928
INFO:tensorflow:Assets written to: ./data/output/models/classification_modelV36/assets
[9]:
print_model_info(binary_model, binary_history)
Model Name: mask_modelV36
Model Type: eventnet
Minimum Training Loss: 0.019914656877518 at epoch 4.
Minimum Validation Loss: 0.018656941130757 at epoch 4.
All Training Losses:
[0.1256965696811676, 0.05998332425951958, 0.038810163736343384, 0.026984436437487602, 0.0199146568775177]
All Validation Losses:
[0.08135028183460236, 0.050937019288539886, 0.03424043580889702, 0.024643391370773315, 0.018656941130757332]
Model Parameters from History:
{'verbose': 1, 'epochs': 5, 'steps': 896}
Model Summary:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_2 (InputLayer) [(None, 512, 512, 1)] 0
conv2d_18 (Conv2D) (None, 256, 256, 64) 3200
leaky_re_lu (LeakyReLU) (None, 256, 256, 64) 0
conv2d_19 (Conv2D) (None, 128, 128, 64) 36928
leaky_re_lu_1 (LeakyReLU) (None, 128, 128, 64) 0
conv2d_20 (Conv2D) (None, 128, 128, 1) 65
leaky_re_lu_2 (LeakyReLU) (None, 128, 128, 1) 0
global_average_pooling2d (G (None, 1, 1, 1) 0
lobalAveragePooling2D)
dense (Dense) (None, 1, 1, 1) 2
=================================================================
Total params: 40,195
Trainable params: 40,195
Non-trainable params: 0
_________________________________________________________________
Run Inference on a Simulated Event with the Binary Classification Model#
Just like before, a great way to get a basic evaulation of this model is to use simulated events. However, this time we need to remember to get the mask from the masking model, rather than from gen_simulated_deformation or gen_sim_noise.
[10]:
def eval_binary_model(
mask_model,
binary_model,
event_type,
seed,
noise_only,
round_mask,
zero_output,
tile_size,
crop_size,
):
wrapped, unwrapped, mask_pred, mask_pred_rounded = eval_mask_model(
mask_model,
event_type,
seed,
noise_only,
round_mask,
zero_output,
tile_size,
crop_size,
)
presence_pred = binary_model.predict(
mask_pred_rounded.reshape(1, tile_size, tile_size, 1)
)
return wrapped, unwrapped, mask_pred, mask_pred_rounded, presence_pred
event_type = "quake"
seed = 0
noise_only = False
round_mask = True
zero_output = False
wrapped, unwrapped, mask_pred, mask_pred_rounded, presence_pred = eval_binary_model(
mask_model,
binary_model,
event_type,
seed,
noise_only,
round_mask,
zero_output,
tile_size,
crop_size,
)
plot_results(wrapped, unwrapped, mask_pred, mask_pred_rounded)
print(f"Event Prediction: {presence_pred[0]}")
Event Prediction: [[[0.9688]]]
Run Inference on a Real Interferogram#
Finally, now that we have working masking and classification models, we can generate a mask and a presence prediction for a real interferogram.
[11]:
from insar_eventnet.inference import mask_with_model
from insar_eventnet.io import get_image_array
image_path = "tests/test_image.tif"
wrapped, _ = get_image_array(image_path)
mask, presence_mask, presence_values = mask_with_model(
mask_model=mask_model,
pres_model=binary_model,
arr_w=wrapped,
tile_size=tile_size,
crop_size=crop_size,
)
[12]:
_, [axs_wrapped, axs_mask, axs_presence_mask] = plt.subplots(1, 3)
axs_wrapped.imshow(wrapped, origin="lower", cmap="jet")
axs_mask.imshow(mask, origin="lower", cmap="jet")
axs_presence_mask.imshow(presence_mask, origin="lower", cmap="jet")
[12]:
<matplotlib.image.AxesImage at 0x7f47007cf9d0>
[ ]: