| Geant4 Cross Reference |
1 """
2 ** generate **
3 generate showers using a saved VAE model
4 """
5 import argparse
6
7 import numpy as np
8 import tensorflow as tf
9 from tensorflow.python.data import Dataset
10
11 from core.constants import GLOBAL_CHECKPOINT_DIR, GEN_DIR, BATCH_SIZE_PER_REPLICA, MAX_GPU_MEMORY_ALLOCATION, GPU_IDS
12 from utils.gpu_limiter import GPULimiter
13 from utils.preprocess import get_condition_arrays
14
15
16 def parse_args():
17 argument_parser = argparse.ArgumentParser()
18 argument_parser.add_argument("--geometry", type=str, default="")
19 argument_parser.add_argument("--energy", type=int, default="")
20 argument_parser.add_argument("--angle", type=int, default="")
21 argument_parser.add_argument("--events", type=int, default=10000)
22 argument_parser.add_argument("--epoch", type=int, default=None)
23 argument_parser.add_argument("--study-name", type=str, default="default_study_name")
24 argument_parser.add_argument("--max-gpu-memory-allocation", type=int, default=MAX_GPU_MEMORY_ALLOCATION)
25 argument_parser.add_argument("--gpu-ids", type=str, default=GPU_IDS)
26 args = argument_parser.parse_args()
27 return args
28
29
30 # main function
31 def main():
32 # 0. Parse arguments.
33 args = parse_args()
34 energy = args.energy
35 angle = args.angle
36 geometry = args.geometry
37 events = args.events
38 epoch = args.epoch
39 study_name = args.study_name
40 max_gpu_memory_allocation = args.max_gpu_memory_allocation
41 gpu_ids = args.gpu_ids
42
43 # 1. Set GPU memory limits.
44 GPULimiter(_gpu_ids=gpu_ids, _max_gpu_memory_allocation=max_gpu_memory_allocation)()
45
46 # 2. Load a saved model.
47
48 # Create a handler and build model.
49 # This import must be local because otherwise it is impossible to call GPULimiter.
50 from core.model import VAEHandler
51 vae = VAEHandler()
52
53 # Load the saved weights
54 weights_dir = f"VAE_epoch_{epoch:03}" if epoch is not None else "VAE_best"
55 vae.model.load_weights(f"{GLOBAL_CHECKPOINT_DIR}/{study_name}/{weights_dir}/model_weights").expect_partial()
56
57 # The generator is defined as the decoder part only
58 generator = vae.model.decoder
59
60 # 3. Prepare data. Get condition values. Sample from the prior (normal distribution) in d dimension (d=latent_dim,
61 # latent space dimension). Gather them into tuples. Wrap data in Dataset objects. The batch size must now be set
62 # on the Dataset objects. Disable AutoShard.
63 e_cond, angle_cond, geo_cond = get_condition_arrays(geometry, energy, events)
64
65 z_r = np.random.normal(loc=0, scale=1, size=(events, vae.latent_dim))
66
67 data = ((z_r, e_cond, angle_cond, geo_cond),)
68
69 data = Dataset.from_tensor_slices(data)
70
71 batch_size = BATCH_SIZE_PER_REPLICA
72
73 data = data.batch(batch_size)
74
75 options = tf.data.Options()
76 options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
77 data = data.with_options(options)
78
79 # 4. Generate showers using the VAE model.
80 generated_events = generator.predict(data) * (energy * 1000)
81
82 # 5. Save the generated showers.
83 np.save(f"{GEN_DIR}/VAE_Generated_Geo_{geometry}_E_{energy}_Angle_{angle}.npy", generated_events)
84
85
86 if __name__ == "__main__":
87 exit(main())