def eval(gt, pred):
'''
Input size : (BS, C, H, W)
'''
gt = (gt + 1.) * 127.5
pred = (pred + 1.) * 127.5
gt = np.clip(gt, 0., 255.).astype(np.uint8)
pred = np.clip(pred, 0., 255.).astype(np.uint8)
batch_size = np.shape(gt)[0]
ssim = np.zeros([batch_size])
psnr = np.zeros([batch_size])
mse = np.zeros([batch_size])
for bs in range(batch_size):
for c in range(gt[bs].shape[0]):
ssim[bs] += ssim_metric(gt[bs][c], pred[bs][c])
psnr[bs] += psnr_metric(gt[bs][c], pred[bs][c])
mse[bs] += mse_metric(gt[bs], pred[bs])
psnr /= 3
ssim /= 3
return mse, ssim, psnr
def eval_seq(gt, pred):
mse_l = []
ssim_l = []
psnr_l = []
l_pred = []
aa = []
T = len(gt)
print("T",T)
bs = gt[0].shape[0]
print("bs",bs)
ssim = np.zeros((bs, T))
psnr = np.zeros((bs, T))
mse = np.zeros((bs, T))
for i in range(bs):
for t in range(T):
for c in range(gt[t][i].shape[0]):
ssim[i, t] += ssim_metric(gt[t][i][c], pred[t][i][c])
psnr[i, t] += psnr_metric(gt[t][i][c], pred[t][i][c])
ssim[i, t] /= gt[t][i].shape[0]
psnr[i, t] /= gt[t][i].shape[0]
mse[i, t] = mse_metric(gt[t][i], pred[t][i])
mse_l.append(mse)
ssim_l.append(ssim)
psnr_l.append(psnr)
#l_pred=list(map(sum, zip(mse_l, ssim_l, psnr_l)))
print('mse_l',mse_l)
print('ssim_l',ssim_l)
print('psnr_l',psnr_l)
#print("l_pred val:", l_pred)
#print('score at at eval_seq',score(l_pred, gt))
#mse_ll.append(mse)
return mse, ssim, psnr
def eval_seq(gt, pred):
T = len(gt)
bs = gt[0].shape[0]
for t in range(T):
pred[t] = np.maximum(pred[t], 0)
pred[t] = np.minimum(pred[t], 1)
ssim = np.zeros((bs, T))
psnr = np.zeros((bs, T))
mse = np.zeros((bs, T))
mae = np.zeros((bs, T))
sharp = np.zeros((bs, T))
for i in range(bs):
for t in range(T):
mse[i, t] = mse_metric(gt[t][i], pred[t][i])
mae[i, t] = mae_metric(gt[t][i], pred[t][i])
sharp[i, t] = sharp_metric(pred[t][i])
x = np.uint8(gt[t][i] * 255)
gx = np.uint8(pred[t][i] * 255)
for c in range(gt[t][i].shape[0]):
ssim[i, t] += ssim_metric(x[c], gx[c])
psnr[i, t] += psnr_metric(x[c], gx[c])
ssim[i, t] /= gt[t][i].shape[0]
psnr[i, t] /= gt[t][i].shape[0]
return np.sum(mse, axis=0), np.sum(mae, axis=0), np.sum(ssim, axis=0), np.sum(psnr, axis=0), np.sum(sharp, axis=0)
def eval_seq_batch(gt, pred):
T = len(gt)
bs = gt[0].shape[0]
ssim = np.zeros(T)
psnr = np.zeros(T)
mse = np.zeros(T)
mae = np.zeros(T)
sharp = np.zeros(T)
for t in range(T):
x = gt[t][:, 0, :, :]
gx = pred[t][:, 0, :, :]
gx = np.maximum(gx, 0)
gx = np.minimum(gx, 1)
mae[t] = batch_mae_frame_float(gx, x)
mse_t = np.square(x - gx).sum()
mse[t] = mse_t
real_frm = np.uint8(x * 255)
pred_frm = np.uint8(gx * 255)
psnr[t] = batch_psnr(pred_frm, real_frm)
for b in range(bs):
sharp[t] += np.max(
cv2.convertScaleAbs(cv2.Laplacian(pred_frm[b], 3)))
score, _ = ssim_metric(pred_frm[b], real_frm[b], full=True)
ssim[t] += score
return mse, mae, ssim, psnr, sharp
def get_seq_psnr_ssim(self, context_frames, gen_frames):
"""Compute PSNR and SSIM from skimage"""
assert context_frames.shape == gen_frames.shape
n_channels = gen_frames.shape[-1]
n_future = gen_frames.shape[1]
batch_size = gen_frames.shape[0]
_psnr_by_step = np.zeros([batch_size, n_future])
_ssim_by_step = np.zeros([batch_size, n_future])
for b in range(batch_size):
for t in range(n_future):
for chan in range(n_channels):
_psnr_by_step[b, t] += psnr_metric(
context_frames[b, t, :, :, chan],
gen_frames[b, t, :, :, chan])
_ssim_by_step[b, t] += ssim_metric(
context_frames[b, t, :, :, chan],
gen_frames[b, t, :, :, chan])
_psnr_by_step[b, t] /= n_channels
_ssim_by_step[b, t] /= n_channels
return _psnr_by_step, _ssim_by_step
def eval_seq(gt, pred):
T = len(gt)
bs = gt[0].shape[0]
ssim = np.zeros((bs, T))
psnr = np.zeros((bs, T))
mse = np.zeros((bs, T))
for i in range(bs):
for t in range(T):
for c in range(gt[t][i].shape[0]):
ssim[i, t] += ssim_metric(gt[t][i][c], pred[t][i][c])
psnr[i, t] += psnr_metric(gt[t][i][c], pred[t][i][c])
ssim[i, t] /= gt[t][i].shape[0]
psnr[i, t] /= gt[t][i].shape[0]
mse[i, t] = mse_metric(gt[t][i], pred[t][i])
return mse, ssim, psnr
