def evaluate_tf(model, val_dataset, frame_skip, fluid_errors=None, scale=1):
print('evaluating.. ', end='')
if fluid_errors is None:
fluid_errors = FluidErrors()
skip = frame_skip
last_scene_id = 0
frames = []
for data in val_dataset:
if data['frame_id0'][0] == 0:
frames = []
if data['frame_id0'][0] % skip < 3:
frames.append(data)
if data['frame_id0'][0] % skip == 3:
if len(
set([
frames[0]['scene_id0'][0], frames[1]['scene_id0'][0],
frames[2]['scene_id0'][0]
])) == 1:
scene_id = frames[0]['scene_id0'][0]
if last_scene_id != scene_id:
last_scene_id = scene_id
print(scene_id, end=' ', flush=True)
frame0_id = frames[0]['frame_id0'][0]
frame1_id = frames[1]['frame_id0'][0]
frame2_id = frames[2]['frame_id0'][0]
box = frames[0]['box'][0]
box_normals = frames[0]['box_normals'][0]
gt_pos1 = frames[1]['pos0'][0]
gt_pos2 = frames[2]['pos0'][0]
inputs = (frames[0]['pos0'][0], frames[0]['vel0'][0], None,
box, box_normals)
pr_pos1, pr_vel1 = model(inputs)
inputs = (pr_pos1, pr_vel1, None, box, box_normals)
pr_pos2, pr_vel2 = model(inputs)
fluid_errors.add_errors(scene_id, frame0_id, frame1_id,
scale * pr_pos1, scale * gt_pos1)
fluid_errors.add_errors(scene_id, frame0_id, frame2_id,
scale * pr_pos2, scale * gt_pos2)
frames = []
result = {}
result['err_n1'] = np.mean([
v['mean'] for k, v in fluid_errors.errors.items() if k[1] + 1 == k[2]
])
result['err_n2'] = np.mean([
v['mean'] for k, v in fluid_errors.errors.items() if k[1] + 2 == k[2]
])
print(result)
print('done')
return result
def evaluate_whole_sequence_torch(model,
val_dataset,
frame_skip,
device,
fluid_errors=None):
import torch
print('evaluating.. ', end='')
if fluid_errors is None:
fluid_errors = FluidErrors()
skip = frame_skip
last_scene_id = None
for data in val_dataset:
scene_id = data['scene_id0'][0]
if last_scene_id is None or last_scene_id != scene_id:
print(scene_id, end=' ', flush=True)
last_scene_id = scene_id
box = torch.from_numpy(data['box'][0]).to(device)
box_normals = torch.from_numpy(data['box_normals'][0]).to(device)
init_pos = torch.from_numpy(data['pos0'][0]).to(device)
init_vel = torch.from_numpy(data['vel0'][0]).to(device)
inputs = (init_pos, init_vel, None, box, box_normals)
else:
inputs = (pr_pos, pr_vel, None, box, box_normals)
pr_pos, pr_vel = model(inputs)
frame_id = data['frame_id0'][0]
if frame_id > 0 and frame_id % skip == 0:
gt_pos = data['pos0'][0]
fluid_errors.add_errors(scene_id,
0,
frame_id,
pr_pos.cpu().numpy(),
gt_pos,
compute_gt2pred_distance=True)
result = {}
result['whole_seq_err'] = np.mean([
v['gt2pred_mean']
for k, v in fluid_errors.errors.items()
if 'gt2pred_mean' in v
])
print(result)
print('done')
return result
def evaluate_ts(model, val_dataset, frame_skip, device, fluid_errors=None):
import torch
print('evaluating.. ', end='')
if fluid_errors is None:
fluid_errors = FluidErrors(flag='ts')
skip = frame_skip
last_scene_id = 0
frames = []
for data in val_dataset:
if data['frame_id0'][0] == 0:
frames = []
if data['frame_id0'][0] % skip < 2:
frames.append(data)
if data['frame_id0'][0] % skip == 2:
if len(
set([
frames[0]['scene_id0'][0], frames[1]['scene_id0'][0]])) == 1:
scene_id = frames[0]['scene_id0'][0]
if last_scene_id != scene_id:
last_scene_id = scene_id
print(scene_id, end=' ', flush=True)
frame0_id = frames[0]['frame_id0'][0]
frame1_id = frames[1]['frame_id0'][0]
box = torch.from_numpy(frames[0]['box'][0]).to(device)
box_normals = torch.from_numpy(
frames[0]['box_normals'][0]).to(device)
inputs = (torch.from_numpy(frames[0]['pos0'][0]).to(device),
torch.from_numpy(frames[0]['vel0'][0]).to(device),
None, box, box_normals)
pr_timestep = model(inputs)
gt = frames[1]['time0'][0] - frames[0]['time0'][0]
fluid_errors.add_errors(scene_id, frame0_id, frame1_id,
pr_timestep.cpu().detach().numpy(), gt[0])
frames = []
result = {}
result['err'] = np.mean(
[v['err'] for k, v in fluid_errors.errors.items() if k[1] + 1 == k[2]])
print(result)
print('done')
return result
def evaluate_torch(model, val_dataset, frame_skip, device, fluid_errors=None):
import torch
print('evaluating.. ', end='')
if fluid_errors is None:
fluid_errors = FluidErrors()
skip = frame_skip
last_scene_id = 0
frames = []
for data in val_dataset:
if data['frame_id0'][0] == 0:
frames = []
if data['frame_id0'][0] % skip < 3:
frames.append(data)
if data['frame_id0'][0] % skip == 3:
if len(
set([
frames[0]['scene_id0'][0], frames[1]['scene_id0'][0],
frames[2]['scene_id0'][0]
])) == 1:
scene_id = frames[0]['scene_id0'][0]
if last_scene_id != scene_id:
last_scene_id = scene_id
print(scene_id, end=' ', flush=True)
frame0_id = frames[0]['frame_id0'][0]
frame1_id = frames[1]['frame_id0'][0]
frame2_id = frames[2]['frame_id0'][0]
box = torch.from_numpy(frames[0]['box'][0]).to(device)
box_normals = torch.from_numpy(
frames[0]['box_normals'][0]).to(device)
gt_pos1 = frames[1]['pos0'][0]
gt_pos2 = frames[2]['pos0'][0]
inputs = (torch.from_numpy(frames[0]['pos0'][0]).to(device),
torch.from_numpy(frames[0]['vel0'][0]).to(device),
None, box, box_normals)
pr_pos1, pr_vel1 = model(inputs)
inputs = (pr_pos1, pr_vel1, None, box, box_normals)
pr_pos2, pr_vel2 = model(inputs)
fluid_errors.add_errors(scene_id, frame0_id, frame1_id,
pr_pos1.cpu().detach().numpy(), gt_pos1)
fluid_errors.add_errors(scene_id, frame0_id, frame2_id,
pr_pos2.cpu().detach().numpy(), gt_pos2)
frames = []
result = {}
result['err_n1'] = np.mean(
[v['mean'] for k, v in fluid_errors.errors.items() if k[1] + 1 == k[2]])
result['err_n2'] = np.mean(
[v['mean'] for k, v in fluid_errors.errors.items() if k[1] + 2 == k[2]])
print(result)
print('done')
return result