-
-
Notifications
You must be signed in to change notification settings - Fork 3
/
experiments.sh
351 lines (316 loc) · 14.5 KB
/
experiments.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
# -*- coding: utf-8 -*-
# This is a log of the experiments run under PVNet2.1
set -e
if false
then
################################################################################################
# These have already been run.
#
# Note that this library has been refactored since these runs. So they will not work as they
# are written here
#
# A few small changes would bes required to re-run these. For example, in the first model below
# `pvnet.models.conv3d.encoders.DefaultPVNet2` should be replaced with
# `pvnet.models.multimodal.encoders.encoders3d.DefaultPVNet2` and
# `pvnet.models.conv3d.dense_networks.ResFCNet2` should be replaced with
# `pvnet.models.multimodal.linear_networks.networks.ResFCNet2`
################################################################################################
# Save pre-made batches
cd scripts
python save_batches.py \
+batch_output_dir="/mnt/disks/batches2/batches_v3.1" \
+num_train_batches=50_000 +num_val_batches=2_000 \
cd ..
# Set baseline to compare to
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet2 \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet2+ResFC2_slow_regx25_amsgrad_v0"
# Use deep supervision to help break down the sources usefulness
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model._target_=pvnet.models.conv3d.deep_supervision.Model \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet2 \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet2+ResFC2_deepsuper_slow_regx25_amsgrad_v0"
# Was the original encoder network better/worse/same?
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2_slow_regx25_amsgrad_v0"
# Are we using too much regularisation?
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet2 \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.01 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet2+ResFC2_slow_regx1_amsgrad_v0"
# Set this baseline using NWP alone
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=nwp_dwsrf_weighting.yaml \
model.optimizer._target_=pvnet.optimizers.Adam \
model.optimizer.lr=0.0001 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="dwsrf_weighting_slow_v3"
# Try a different encoder network
python run.py \
datamodule.batch_dir="/mnt/disks/batches/batches_v3" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v3.yaml \
model.encoder_kwargs.model_name="efficientnet-b2" \
+model.add_image_embedding_channel=True \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.05 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="EffNet+ResFC2_slow_regx25_amsgrad_v1"
# Use deep supervision and pvnet1 encoder so we can compare to pvnet2+deepsuper
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model._target_=pvnet.models.conv3d.deep_supervision.Model \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2_deepsuper_slow_regx25_amsgrad_v0"
# Reset this baseline for model trained under PVNet2.0 project
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model._target_=pvnet.models.conv3d.weather_residual.Model \
+model.version=1 \
model.add_image_embedding_channel=True \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
+model.optimizer.amsgrad=True \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC_weatherRes_slow_regx25_amsgrad_v1"
# What if we exclude historical GSP as input?
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.include_gsp_yield_history=False \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2_slow_regx25_amsgrad_v1_nohist"
# How about a smaller encoder model?
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.encoder_kwargs.number_of_conv3d_layers=2 \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet_shal+ResFC2_slow_regx25_amsgrad_v1"
# How about a bigger outout model?
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.output_network_kwargs.n_res_blocks=6 \
model.output_network_kwargs.fc_hidden_features=128 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2big_slow_regx25_amsgrad_v1"
# Try the self-regularising neural network as the output network
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v4.yaml \
model.image_encoder._target_=pvnet.models.conv3d.encoders.DefaultPVNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.SNN \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.0001 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+SNN_slow_regx25_amsgrad_v0"
# Try using ResNet encoder
python run.py \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.1" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v5.yaml \
model._target_=pvnet.models.conv3d.deep_supervision.Model \
model.image_encoder._target_=pvnet.models.conv3d.encoders.ResNet \
model.output_network._target_=pvnet.models.conv3d.dense_networks.ResFCNet2 \
model.sat_image_size_pixels=24 \
model.nwp_image_size_pixels=24 \
model.number_nwp_channels=2 \
model.nwp_history_minutes=120 \
model.nwp_forecast_minutes=480 \
model.history_minutes=120 \
model.optimizer._target_=pvnet.optimizers.AdamWReduceLROnPlateau \
model.optimizer.lr=0.00005 \
+model.optimizer.weight_decay=0.25 \
+model.optimizer.amsgrad=True \
+model.optimizer.patience=5 \
+model.optimizer.factor=0.1 \
+model.optimizer.threshold=0.002 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=4 \
trainer.accumulate_grad_batches=32 \
model_name="ResNet+ResFC2_deepsup_slow_regx25_amsgrad_v1"
cd scripts
# Re-train this model after refactoring
python save_batches.py \
+batch_output_dir="/mnt/disks/batches2/batches_v3.2" \
+num_train_batches=200_000 \
+num_val_batches=4_000
cd ..
python run.py \
datamodule=premade_batches \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.2" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=conv3d_sat_nwp_v6.yaml \
callbacks.early_stopping.patience=20 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2+_slow_regx25_amsgrad_v4"
fi
cd scripts
# Changes in datapipes make this more like production
python save_batches.py \
+batch_output_dir="/mnt/disks/batches2/batches_v3.4" \
+num_train_batches=200_000 \
+num_val_batches=4_000
cd ..
python run.py \
datamodule=premade_batches \
datamodule.batch_dir="/mnt/disks/batches2/batches_v3.4" \
+trainer.val_check_interval=10_000 \
trainer.log_every_n_steps=200 \
model=multimodal.yaml \
model.include_gsp_yield_history=False \
+model.min_sat_delay_minutes=30 \
callbacks.early_stopping.patience=20 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
callbacks.early_stopping.patience=10 \
datamodule.batch_size=32 \
trainer.accumulate_grad_batches=4 \
model_name="pvnet+ResFC2+_slow_regx25_amsgrad_v5_nohist"