def validate_dcn(dcn, data, save_dir=False, epoch=0, show_ref=False):
"""
Computes validation metrics for a compression model (DCN). (If not a DCN, the function returns immediately).
If requested, plot compressed images to a JPEG image.
:param dcn: the DCN model
:param data: the dataset (instance of Dataset)
:param data: the dataset (instance of Dataset)
:param save_dir: path to the directory where figures should be generated
:param epoch: epoch counter to be appended to the output filename
:param show_ref: whether to show only the compressed image or also the input image as reference
:return: tuple of lists with per-image measurements of (ssims, psnrs, losses, entropies)
"""
if not isinstance(dcn, DCN):
return
# Compute latent representations and compressed output
batch_x = data.next_validation_batch(0, data.count_validation)
if isinstance(batch_x, tuple):
batch_x = batch_x[-1]
batch_y, entropy = dcn.process(batch_x, return_entropy=True)
entropy = float(entropy.numpy())
ssim = metrics.ssim(batch_x, batch_y.numpy()).tolist()
psnr = metrics.psnr(batch_x, batch_y.numpy()).tolist()
loss = float(dcn.loss(batch_x, batch_y, entropy).numpy())
# If requested, plot a figure with input/output pairs
if save_dir is not None:
images_x = np.minimum(data.count_validation, 10 if not show_ref else 5)
images_y = np.ceil(data.count_validation / images_x)
fig = Figure(figsize=(20, 20 / images_x * images_y *
(1 if not show_ref else 0.5)))
for b in range(data.count_validation):
ax = fig.add_subplot(images_y, images_x, b + 1)
plots.image(np.concatenate(
(batch_x[b], batch_y[b]), axis=1) if show_ref else batch_y[b],
'{:.1f} / {:.2f}'.format(psnr[b], ssim[b]),
axes=ax)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
fig.savefig('{}/dcn_validation_{:05d}.jpg'.format(save_dir, epoch),
bbox_inches='tight',
dpi=100,
quality=90)
del fig
return {
'ssim': float(np.mean(ssim)),
'psnr': float(np.mean(psnr)),
'loss': loss,
'entropy': entropy
}
def compare_batches(batch_a, batch_b, labels=None, fig=None, figwidth=4, nrows=3, transpose=False):
n_images = min(len(batch_a), len(batch_b))
labels = labels or ['', '']
fig, axes = plots.sub(n_images * nrows, figwidth, ncols=n_images, fig=fig, transpose=transpose)
for i, (img_a, img_b) in enumerate(zip(batch_a, batch_b)):
label_a = f'(A) {labels[0]}'
plots.image(img_a, label_a, axes=axes[i + n_images * 0])
psnr = metrics.psnr(img_a, img_b)
ssim = metrics.ssim(img_a, img_b)
label_b = f'(B) {labels[1]}: {psnr:.1f} dB / {ssim:.3f}'
plots.image(img_b, label_b, axes=axes[i + n_images * 1])
diff_ab = img_b - img_a
diff_abs = np.abs(img_b - img_a)
diff_mean = diff_ab.mean()
if nrows > 2:
plots.image(image.normalize(diff_ab), f'A - B: mean abs {diff_mean:.3f}', axes=axes[i + n_images*2])
if nrows > 3:
plots.image(image.normalize(diff_abs, 0.1), f'|A - B|: mean abs {diff_mean:.3f}', axes=axes[i + n_images*3])
return fig
def test_output(image, jpeg_quality=50, rounding_approximation=None):
jpg = DJPG(rounding_approximation=rounding_approximation)
print(jpg)
batch_x = np.expand_dims(image, 0)
batch_y = jpg.process(batch_x / 255, jpeg_quality)
n_images = batch_x.shape[0]
batch_j = np.zeros_like(batch_x)
for n in range(n_images):
io.imwrite('/tmp/patch_{}.jpg'.format(n),
(batch_x[n].squeeze()).astype(np.uint8),
quality=jpeg_quality,
subsampling='4:4:4')
batch_j[n] = io.imread('/tmp/patch_{}.jpg'.format(n))
for n in range(n_images):
plt.subplot(n_images, 3, 1 + n * 3)
plots.image(batch_x[n].squeeze() / np.max(np.abs(batch_x)), 'Input')
plt.subplot(n_images, 3, 2 + n * 3)
plots.image(batch_y[n].squeeze() / np.max(np.abs(batch_y)),
'dJPEG Model')
plt.subplot(n_images, 3, 3 + n * 3)
plots.image(batch_j[n].squeeze() / np.max(np.abs(batch_j)),
'libJPG Codec')
plt.show()
def match_jpeg(model, batch_x, axes=None, match='ssim'):
# Compress using DCN and get number of bytes
batch_y, bytes_dcn = codec.simulate_compression(batch_x, model)
ssim_dcn = metrics.ssim(batch_x.squeeze(), batch_y.squeeze()).mean()
bpp_dcn = 8 * bytes_dcn / np.prod(batch_x.shape[1:-1])
target = ssim_dcn if match == 'ssim' else bpp_dcn
try:
jpeg_quality = jpeg_helpers.match_quality(batch_x.squeeze(),
target,
match=match)
except:
if match == 'ssim':
jpeg_quality = 95 if ssim_dcn > 0.8 else 10
else:
jpeg_quality = 95 if bpp_dcn > 3 else 10
print(
'WARNING Could not find a matching JPEG quality factor - guessing {}'
.format(jpeg_quality))
# Compress using JPEG
batch_j, bytes_jpeg = jpeg_helpers.compress_batch(batch_x[0],
jpeg_quality,
effective=True)
ssim_jpeg = metrics.ssim(batch_x.squeeze(), batch_j.squeeze()).mean()
bpp_jpg = 8 * bytes_jpeg / np.prod(batch_x.shape[1:-1])
# Get stats
code_book = model.get_codebook()
batch_z = model.compress(batch_x).numpy()
counts = helpers.stats.hist(batch_z, code_book)
counts = counts.clip(min=1)
probs = counts / counts.sum()
entropy = -np.sum(probs * np.log2(probs))
# Print report
print('DCN : {}'.format(model.model_code))
print('Pixels : {}x{} = {:,} px'.format(
batch_x.shape[1], batch_x.shape[2], np.prod(batch_x.shape[1:-1])))
print('Bitmap : {:,} bytes'.format(np.prod(batch_x.shape)))
print('Code-book size : {} elements from {} to {}'.format(
len(code_book), min(code_book), max(code_book)))
print('Entropy : {:.2f} bits per symbol'.format(entropy))
print('Latent size : {:,}'.format(np.prod(batch_z.shape)))
print('PPF Naive : {:,.0f} --> {:,.0f} bytes [{} bits per element]'.
format(
np.prod(batch_z.shape) * np.log2(len(code_book)) / 8,
np.prod(batch_z.shape) * np.ceil(np.log2(len(code_book))) / 8,
np.ceil(np.log2(len(code_book)))))
print('PPF Theoretical : {:,.0f} bytes ({:.2f} bpp)'.format(
np.prod(batch_z.shape) * entropy / 8,
np.prod(batch_z.shape) * entropy / np.prod(batch_x.shape[1:-1])))
print('FSE Coded : {:,} bytes ({:.2f} bpp) --> ssim: {:.3f}'.format(
bytes_dcn, bpp_dcn, ssim_dcn))
print(
'JPEG (Q={:2d}) : {:,} bytes ({:0.2f} bpp) --> ssim: {:.3f} // effective size disregarding JPEG headers'
.format(jpeg_quality, bytes_jpeg, bpp_jpg, ssim_jpeg))
# Plot results
if axes is None:
fig, axes = plots.sub(6, ncols=3)
fig.set_size_inches(12, 10)
fig.tight_layout()
else:
fig = axes[0].figure
# Plot full-resolution
plots.image(batch_x,
'Original ({0}x{0})'.format(batch_x.shape[1]),
axes=axes[0])
plots.image(batch_y,
'DCN ssim:{:.2f} bpp:{:.2f}'.format(ssim_dcn, bpp_dcn),
axes=axes[1])
plots.image(batch_j,
'JPEG {} ssim:{:.2f} bpp:{:.2f}'.format(
jpeg_quality, ssim_jpeg, bpp_jpg),
axes=axes[2])
# Plot zoom
crop_size = max([64, batch_x.shape[1] // 4])
plots.image(helpers.image.crop_middle(batch_x, crop_size),
'Original crop ({0}x{0})'.format(crop_size),
axes=axes[3])
plots.image(helpers.image.crop_middle(batch_y, crop_size),
'DCN crop ({0}x{0})'.format(crop_size),
axes=axes[4])
plots.image(helpers.image.crop_middle(batch_j, crop_size),
'JPEG crop ({0}x{0})'.format(crop_size),
axes=axes[5])
return fig
def show_example(model, batch_x):
# Compress and decompress model
batch_z = model.compress(batch_x).numpy()
batch_y = model.decompress(batch_z).numpy()
# Get empirical histogram of the latent representation
codebook = model.get_codebook()
qmin = np.floor(codebook[0])
qmax = np.ceil(codebook[-1])
bin_centers = np.arange(qmin - 1, qmax + 1, 0.1)
bin_boundaries = np.convolve(bin_centers, [0.5, 0.5], mode='valid')
bin_centers = bin_centers[1:-1]
hist_emp = np.histogram(batch_z.reshape((-1, )),
bins=bin_boundaries,
density=True)[0]
hist_emp = np.maximum(hist_emp, 1e-9)
hist_emp = hist_emp / hist_emp.sum()
# Get TF histogram estimate based on soft quantization
hist = helpers.stats.hist(batch_z, codebook)
hist = hist / hist.sum()
# Entropy
entropy = -np.sum(hist * np.log2(hist))
entropy_emp = -np.sum(hist_emp * np.log2(hist_emp))
fig, axes = plots.sub(2, ncols=1)
fig.set_size_inches(12, 10)
axes[0].plot(bin_centers, hist_emp / hist_emp.max(), 'r-')
axes[0].plot(codebook, hist / hist.max(), '-bo')
axes[0].legend([
'Empirical H={:.2f}'.format(entropy_emp),
'TF estimate (soft) H={:.2f}'.format(entropy)
])
axes[0].set_ylabel('normalized frequency')
axes[0].set_xlabel('latent values')
# Thumbnails
indices = np.argsort(np.var(batch_x, axis=(1, 2, 3)))[::-1]
thumbs_pairs_few = np.concatenate((batch_x[indices], batch_y[indices]),
axis=0)
thumbs_few = (
255 * plots.thumbnails(thumbs_pairs_few, ncols=len(batch_x))).astype(
np.uint8)
ssim_values = [
metrics.ssim(batch_x[i], batch_y[i]).mean()
for i in range(len(batch_x))
]
plots.image(thumbs_few,
'Sample reconstructions, ssim={:.3f}'.format(
np.mean(ssim_values)),
axes=axes[1])
fig.tight_layout()
return fig
def compare_images_ab_ref(img_ref, img_a, img_b, labels=None, extras=False, fig=None):
labels = labels or ['target', '', '']
img_a = img_a.squeeze()
img_b = img_b.squeeze()
img_ref = img_ref.squeeze()
fig, axes = plots.sub(9 if extras else 3, ncols=3, fig=fig)
# Index of the last axes
j = 3 if extras else 2
plots.image(img_ref, '(T) {}'.format(labels[0]), axes=axes[0])
label_a = '(A) {}: {:.1f} dB / {:.3f}'.format(labels[1], metrics.psnr(img_ref, img_a), metrics.ssim(img_ref, img_a))
plots.image(img_a, label_a, axes=axes[1])
label_b = '(B) {}: {:.1f} dB / {:.3f}'.format(labels[2], metrics.psnr(img_ref, img_b), metrics.ssim(img_ref, img_b))
plots.image(img_b, label_b, axes=axes[j])
# A hack to allow image axes to zoom together
axes[1].get_shared_x_axes().join(axes[0], axes[1])
axes[j].get_shared_x_axes().join(axes[0], axes[j])
axes[1].get_shared_y_axes().join(axes[0], axes[1])
axes[j].get_shared_y_axes().join(axes[0], axes[j])
if not extras:
return fig
# Compute and plot difference images
diff_a = np.abs(img_a - img_ref)
diff_a_mean = diff_a.mean()
diff_a = image.normalize(diff_a, 0.1)
diff_b = np.abs(img_b - img_ref)
diff_b_mean = diff_b.mean()
diff_b = image.normalize(diff_b, 0.1)
diff_ab = np.abs(img_b - img_a)
diff_ab_mean = diff_ab.mean()
diff_ab = image.normalize(diff_ab, 0.1)
plots.image(diff_a, 'T - A: mean abs {:.3f}'.format(diff_a_mean), axes=axes[2])
plots.image(diff_b, 'T - B: mean abs {:.3f}'.format(diff_b_mean), axes=axes[6])
plots.image(diff_ab, 'A - B: mean abs {:.3f}'.format(diff_ab_mean), axes=axes[4])
# Compute and plot spectra
fft_a = fft_log_norm(diff_a)
fft_b = fft_log_norm(diff_b)
# fft_ab = utils.normalize(np.abs(fft_a - fft_b))
fft_ab = image.normalize(np.abs(fft_log_norm(img_b) - fft_log_norm(img_a)), 0.01)
plots.image(fft_a, 'FFT(T - A)', axes=axes[5])
plots.image(fft_b, 'FFT(T - B)', axes=axes[7])
plots.image(fft_ab, 'FFT(A) - FFT(B)', axes=axes[8])
return fig
def develop_image(pipeline,
camera=None,
batch=None,
image=None,
patch_size=0,
patches=2,
root_dir='./data',
pipeline_args=None):
"""
Display a patch developed by a neural imaging pipeline.
"""
if camera is not None:
supported_cameras = fsutil.listdir(
os.path.join(root_dir, 'models', 'nip'), '.*')
if camera not in supported_cameras:
raise ValueError(
'Camera data not found ({})! Available cameras: {}'.format(
camera, ', '.join(supported_cameras)))
root_dirname = os.path.join(root_dir, 'models', 'nip', camera)
data_dirname = os.path.join(root_dir, 'raw', 'training_data', camera)
if patch_size != 0 and (patch_size < 4 or patch_size > 2048):
raise ValueError('Patch size seems to be invalid!')
# Lazy imports to minimize delay for invalid command line parameters
import numpy as np
import imageio as io
import matplotlib.pyplot as plt
import tensorflow as tf
from models import pipelines
# Construct the NIP model ---------------------------------------------------------------------
if os.path.isdir(pipeline):
# Restore a NIP model from a training log
model = tfmodel.restore(pipeline, pipelines)
else:
# Construct the NIP model from class name (and optional arguments)
if pipeline_args is None:
model = getattr(pipelines, pipeline)()
else:
model = getattr(pipelines, pipeline)(**pipeline_args)
loaded_model = False
candidate_dirs = [
os.path.join(root_dirname, model.model_code),
os.path.join(root_dirname)
]
for candidate in candidate_dirs:
if os.path.isdir(candidate):
model.load_model(candidate)
loaded_model = True
break
if not loaded_model:
raise FileNotFoundError(
f'Could not find the corresponding model: {candidate_dirs}')
# Load image(s) -------------------------------------------------------------------------------
if image is None and batch is not None:
print('Loading a batch of {} images'.format(batch))
data = dataset.Dataset(data_dirname,
n_images=0,
v_images=batch,
val_rgb_patch_size=patch_size or 256,
val_n_patches=patches)
sample_x, sample_y = data.next_validation_batch(
0, data.count_validation)
with open('config/cameras.json') as f:
cameras = json.load(f)
cfa, srgb = cameras[camera]['cfa'], np.array(
cameras[camera]['srgb'])
elif image is not None:
print('Loading a RAW image {}'.format(image))
sample_x, cfa, srgb, _ = raw.unpack(image, expand=True)
sample_y = raw.process(image, brightness=None, expand=True)
if isinstance(model, pipelines.ClassicISP):
print('Configuring ISP to CFA: {} & sRGB {}'.format(
cfa,
srgb.round(2).tolist()))
model.set_cfa_pattern(cfa)
model.set_srgb_conversion(srgb)
sample_Y = model.process(sample_x).numpy()
if patch_size > 0:
xx = (sample_y.shape[2] - patch_size) // 2
yy = (sample_y.shape[1] - patch_size) // 2
sample_y = sample_y[:, yy:yy + patch_size, xx:xx + patch_size, :]
sample_Y = sample_Y[:, yy:yy + patch_size, xx:xx + patch_size, :]
psnrs = metrics.psnr(sample_y, sample_Y)
ssims = metrics.ssim(sample_y, sample_Y)
print('sample x: {}'.format(sample_x.shape))
print('sample y: {}'.format(sample_y.shape))
print('sample Y: {}'.format(sample_Y.shape))
# Plot images ---------------------------------------------------------------------------------
if len(sample_y) > 1:
sample_y = plots.thumbnails(sample_y, batch, True)
sample_Y = plots.thumbnails(sample_Y, batch, True)
else:
sample_y = sample_y.squeeze()
sample_Y = sample_Y.squeeze()
print('thumbnails: {}'.format(sample_y.shape))
ncols = 1 if sample_y.shape[1] > sample_y.shape[0] else 2
nrows = 2 if ncols == 1 else 1
fig, axes = plt.subplots(nrows, ncols)
plots.image(sample_Y,
'{}, PSNR={:.1f} dB, SSIM={:.2f} : {{}}'.format(
model.model_code, float(psnrs.mean()),
float(ssims.mean())),
axes=axes[0])
plots.image(sample_y, 'Target RGB images () : {}', axes=axes[1])
plt.show()
plt.close()
def validate_nip(model,
data,
save_dir=False,
epoch=0,
show_ref=False,
loss_type='L2'):
"""
Develops image patches using the given NIP and returns standard image quality measures.
If requested, resulting patches are visualized as thumbnails and saved to a directory.
:param model: the NIP model
:param data: the dataset (instance of Dataset)
:param data: the dataset (instance of Dataset)
:param save_dir: path to the directory where figures should be generated
:param epoch: epoch counter to be appended to the output filename
:param show_ref: whether to show only the developed image or also the GT target
:param loss_type: L1 or L2
:return: tuple of lists with per-image measurements of (ssims, psnrs, losss)
"""
ssims = []
psnrs = []
losss = []
# If requested, plot a figure with output/target pairs
if save_dir is not None:
images_x = np.minimum(data.count_validation, 10 if not show_ref else 5)
images_y = np.ceil(data.count_validation / images_x)
fig = Figure(figsize=(20, 20 / images_x * images_y *
(1 if not show_ref else 0.5)))
developed_out = np.zeros_like(data['validation']['y'], dtype=np.float32)
for b in range(data.count_validation):
example_x, example_y = data.next_validation_batch(b, 1)
developed = model.process(example_x).numpy().clip(0, 1)
developed_out[b, :, :, :] = developed
developed = developed[:, :, :, :].squeeze()
reference = example_y.squeeze()
# Compute stats
ssim = metrics.ssim(reference, developed).mean()
psnr = metrics.psnr(reference, developed).mean()
if loss_type == 'L2':
loss = np.mean(np.power(reference - developed, 2.0))
elif loss_type == 'L1':
loss = np.mean(np.abs(reference - developed))
else:
raise ValueError('Invalid loss! Use either L1 or L2.')
ssims.append(ssim)
psnrs.append(psnr)
losss.append(loss)
# Add images to the plot
if save_dir is not None:
ax = fig.add_subplot(images_y, images_x, b + 1)
plots.image(np.concatenate(
(reference, developed), axis=1) if show_ref else developed,
'{:.1f} dB / {:.2f}'.format(psnr, ssim),
axes=ax)
if save_dir is not None:
if not os.path.exists(save_dir):
os.makedirs(save_dir)
fig.savefig('{}/nip_validation_{:05d}.jpg'.format(save_dir, epoch),
bbox_inches='tight',
dpi=100,
quality=90)
del fig
return ssims, psnrs, losss