File size: 1,866 Bytes
41dd7e0 | 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 | # Evaluation settings shared by the five systems: csFID-10k (the paper's image metric) and the sampling of the test read.
fid:
n_truth: 10000 # truth realizations per test condition (the first 10,000 members of the truth bank)
n_generated: 10000 # generated samples per model and test condition
backend: clean-fid
clean_fid_version: "0.1.35" # checked at run time
mode: clean # clean-fid "clean" mode: one PIL bicubic resize to 299 x 299
inception:
file: inception-2015-12-05.pt # the TF Inception-v3 weights distributed with clean-fid (sha256 checked)
sha256: f58cb9b6ec323ed63459aa4fb441fe750cfe39fafad6da5cb504a16f19e958f4
feature_layer: pool3 # 2048-D, used by cFID / cKID
resize: {size: 299, filter: bicubic, backend: PIL, antialias: true, once: true}
kid: {num_subsets: 100, max_subset_size: 1000}
clip_quantile: 0.9999 # image range: c = Q_0.9999(|x|) over the training fields (records/<system>/fid_adapter.json)
feature_batch: 200
primary_metric: csfid # csFID: the Frechet distance of the spatial Inception feature below
secondary_metrics: [cfid, ckid]
spatial:
layer: mixed_6/conv # the first branch of the mixed_6 block, before concatenation (192 x 17 x 17)
channels: 7 # the first 7 channels
grid: 17
dim: 2023 # 17 * 17 * 7, flattened NHWC, no spatial pooling
flatten: NHWC
pooling: none
reference: "Nash et al., ICML 2021; openai/guided-diffusion evaluations/evaluator.py FID_SPATIAL_NAME"
sampling:
crn_base_seed: 777 # the noise of test condition gid is drawn from the stream seeded with 777 * 100003 + gid
sample_chunk: 2000 # the 10,000 samples of a condition are drawn in chunks of 2,000 from that stream
|