img_count = 0
outrange = len(tiCr.tile_inputs_all_complete) / tilesPerImg
# use int to avoid TypeError: 'float' object cannot be interpreted as an integer
for currOut in range( int(outrange) ):
batch_xs = []
batch_ys = []
for curr_tile in range(tilesPerImg):
idx = currOut * tilesPerImg + curr_tile
batch_xs.append(tiCr.tile_inputs_all_complete[idx])
batch_ys.append(np.zeros((tileSizeHigh * tileSizeHigh), dtype='f'))
resultTiles = y_pred.eval(feed_dict={x: batch_xs, y_true: batch_ys, keep_prob: 1.})
tiCr.debugOutputPngsCrop(resultTiles, tileSizeHigh, simSizeHigh, test_path, \
imageCounter=currOut, cut_output_to=tileSizeHiCrop, tiles_in_image=tilesPerImg)
# optionally, output references
#tiCr.debugOutputPngsCrop(batch_ys, tileSizeHigh, simSizeHigh, test_path+"_ref", imageCounter=currOut, cut_output_to=tileSizeHiCrop, tiles_in_image=tilesPerImg)
img_count += 1
print('Test finished, %d pngs written to %s.' % (img_count, test_path) )
# write summary to test overview
loaded_model = ''
if not loadModelTest == -1:
loaded_model = ', Loaded %04d, %d' % (loadModelTest , loadModelNo)
with open(basePath + 'test_overview.txt', "a") as text_file:
text_file.write(test_path[-10:-1] + ': {:.2f} min, {} Epochs, cost {:.4f}, {}'.format(training_duration, trainingEpochs, cost, " ") + loaded_model + '\n')
break
resultTiles = y_pred.eval(feed_dict={
x: batch_xs,
y_true: batch_ys,
training: False
})
if brightenOutput > 0:
for i in range(len(resultTiles)):
resultTiles[i] *= brightenOutput
for i in range(len(batch_ys)):
batch_ys[i] *= brightenOutput
tiCr.debugOutputPngsCrop(resultTiles,
tileSizeHigh,
simSizeHigh,
test_path,
imageCounter=currOut,
cut_output_to=tileSizeHiCrop,
tiles_in_image=tilesPerImg,
name='output')
tiCr.debugOutputPngsCrop(batch_ys,
tileSizeHigh,
simSizeHigh,
test_path,
imageCounter=currOut,
cut_output_to=tileSizeHiCrop,
tiles_in_image=tilesPerImg,
name='expected_out')
if outputInputs:
if not useVelocities:
tiCr.debugOutputPngsSingle(batch_xs,
tdataSize = len(batch_xs)
# for prediction, we have to get rid of the third spatial dim
batch_xs = np.reshape( batch_xs, [tdataSize,tileSizeLow,tileSizeLow, n_inputChannels] )
batch_ys = np.reshape( batch_ys, [tdataSize,tileSizeHigh,tileSizeHigh,1 ] )
resultTiles = model.predict( batch_xs )
# now restore it...
batch_xs = np.reshape( batch_xs, [tdataSize,1, tileSizeLow,tileSizeLow, n_inputChannels] )
batch_ys = np.reshape( batch_ys, [tdataSize,1, tileSizeHigh,tileSizeHigh,1 ] )
resultTiles = np.reshape( resultTiles, [tdataSize,1, tileSizeHigh,tileSizeHigh,1 ] )
# simply concat tiles into images...
tileSizeHiCrop = upRes * cropTileSizeLow
tilesPerImg = (simSizeHigh // tileSizeHiCrop) ** 2
imgCnt = len(tiCr.tile_inputs_all_complete) / tilesPerImg
tiCr.debugOutputPngsCrop(resultTiles, tileSizeHigh, simSizeHigh, test_path, imageCounter=0, cut_output_to=tileSizeHiCrop, \
tiles_in_image=tilesPerImg, name='output')
if outputInputs:
tiCr.debugOutputPngsSingle(batch_xs, tileSizeLow, simSizeLow, test_path, imageCounter=0, name='input', channel=0)
if useVelocities:
tiCr.debugOutputPngsSingle(batch_xs, tileSizeLow, simSizeLow, test_path, imageCounter=0, name='in_vel_x', channel=1)
tiCr.debugOutputPngsSingle(batch_xs, tileSizeLow, simSizeLow, test_path, imageCounter=0, name='in_vel_y', channel=2)
print('Output finished, %d pngs written to %s.' % (imgCnt, test_path) )