-
Notifications
You must be signed in to change notification settings - Fork 1
/
model_eval.py
385 lines (294 loc) · 16.6 KB
/
model_eval.py
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
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
"""
This file contains functions for running inference on a dataset and printing/plotting the results.
"""
import torch
import numpy as np
from torch.utils.data import DataLoader
import h5py
from torchvision import transforms
from typing import List, Tuple, Dict
from Unet_model import UNet
import matplotlib
from matplotlib import pyplot as plt
import os
import shutil
import cv2
from tqdm import tqdm
from data_loader import BeadSightDataset, get_press_indicies, get_valid_indices, decompress_h5py
from model_training import plot_pressure_maps
from multiprocessing import Pool
def otzu_single(image:np.ndarray) -> np.ndarray:
"""
Run otzu thresholding on a single image. This function is used for multiprocessing.
"""
# convert to uint16 for openCV to work
image = ((65535/image.max())*image).astype(np.uint16)
return cv2.threshold(image, 0, 1, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]
def run_inference(checkpoint_path: str,
data_path: str,
save_path: str,
indicies: str,
batch_size: int = 64,
otzu_batch_size: int = 1024):
"""
Run inference on the dataset and save the results to a file.
"""
# Load the dataset
checkpoint = torch.load(checkpoint_path, map_location='cpu')
window_size:int = checkpoint['window_size']
pixel_mean: List[float] = checkpoint['pixel_mean']
pixel_std: List[float] = checkpoint['pixel_std']
average_force: float = checkpoint['data_loader_info']['average_force']
model = UNet(window_size)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
if indicies == "test": # If no indicies are provided, use the saved test indicies
indicies = checkpoint['test_indices']
elif indicies == "all": # If all indicies are provided, use all the indicies
indicies = get_valid_indices(data_path, window_size)
else:
raise ValueError("indicies must be 'test' or 'all'")
dataset = BeadSightDataset(hdf5_file=data_path,
indicies=indicies,
pixel_mean=pixel_mean,
pixel_std=pixel_std,
average_force=average_force,
train=False,
window_size=window_size)
avg_pressure = dataset.avg_pressure
test_data_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
prefetch_factor=2,
pin_memory=True)
# get the max force for each press:
press_indicies = get_press_indicies(data_path, window_size)
with h5py.File(data_path, 'r') as data_file:
forces = data_file['forces'][:]
sensor_dim = data_file.attrs['sensor_size'] # mm
sensor_area = sensor_dim[0]*sensor_dim[1]/(1000**2) # m^2
press_max_pressures = []
for idx_list in press_indicies:
press_max_force = np.max(forces[idx_list])
press_max_pressures.append(dataset.pressure_mapper.pressure_from_force(press_max_force))
with h5py.File(save_path, 'w') as save_file:
save_file.attrs['checkpoint_path'] = checkpoint_path
save_file.attrs['data_path'] = data_path
save_file.attrs['avg_pressure'] = avg_pressure
save_file.attrs['sensor_size'] = sensor_dim
#copy the checkpoint info the the attributes, excluding the model and optimizer:
# save_file.attrs['checkpoint_info'] = checkpoint_info
n_samples = len(indicies)
save_file.create_dataset(name='press_max_pressures',
data=press_max_pressures)
idx_data = save_file.create_dataset(name='indicies',
shape=(n_samples,),
dtype=np.int32)
gt_pressure_map = save_file.create_dataset(name='gt_pressure_map',
shape=(n_samples, 256, 256),
chunks=(1, 256, 256),
dtype=np.float32,
compression=9)
gt_pressure_values = save_file.create_dataset(name='gt_pressure_values',
shape = (n_samples,),
dtype=np.float32)
pred_pressure_map = save_file.create_dataset(name='pred_pressure_map',
shape=(n_samples, 256, 256),
chunks=(1, 256, 256),
dtype=np.float32,
compression=9)
gt_center_of_pressure = save_file.create_dataset(name='gt_center_of_pressure',
shape=(n_samples, 2),
dtype=np.float32)
pred_center_of_pressure = save_file.create_dataset(name='pred_center_of_pressure',
shape=(n_samples, 2),
dtype=np.float32)
gt_total_force = save_file.create_dataset(name='gt_total_force',
shape=(n_samples,),
dtype=np.float32)
pred_total_force = save_file.create_dataset(name='pred_total_force',
shape=(n_samples,),
dtype=np.float32)
mse_values = save_file.create_dataset(name='mse_values',
shape=(n_samples,),
dtype=np.float32)
mae_values = save_file.create_dataset(name='mae_values',
shape=(n_samples,),
dtype=np.float32)
for i, data in enumerate(tqdm(test_data_loader)):
video_frames, norm_gt_maps, idxs = data
video_frames: torch.Tensor = video_frames.to(device)
norm_gt_maps: torch.Tensor = norm_gt_maps.to(device)
with torch.no_grad():
# run inference
norm_pred_maps: torch.Tensor = model(video_frames)
# un_normalize the pressure maps
pred_maps = norm_pred_maps*avg_pressure
gt_maps = norm_gt_maps*avg_pressure
# calculate the loss
mae_values[i*batch_size:(i+1)*batch_size] = torch.abs(pred_maps - gt_maps).mean(dim=(1,2)).cpu().numpy()
mse_values[i*batch_size:(i+1)*batch_size] = ((pred_maps - gt_maps)**2).mean(dim=(1,2)).cpu().numpy()
# calculate the center of pressure
x_locations = torch.arange(256).to(device)
y_locations = torch.arange(256).to(device)
pred_x_center = torch.sum(torch.sum(pred_maps, dim=2)*x_locations, dim=1)/torch.sum(pred_maps, dim=(1,2))
pred_y_center = torch.sum(torch.sum(pred_maps, dim=1)*y_locations, dim=1)/torch.sum(pred_maps, dim=(1,2))
gt_x_center = torch.sum(torch.sum(gt_maps, dim=2)*x_locations, dim=1)/torch.sum(gt_maps, dim=(1,2))
gt_y_center = torch.sum(torch.sum(gt_maps, dim=1)*y_locations, dim=1)/torch.sum(gt_maps, dim=(1,2))
# calculate the total force
pred_total_force[i*batch_size:(i+1)*batch_size] = torch.mean(pred_maps, dim=(1,2)).cpu().numpy()*sensor_area # total force, N
gt_total_force[i*batch_size:(i+1)*batch_size] = torch.mean(gt_maps, dim=(1,2)).cpu().numpy()*sensor_area # total force, N
# save the results
idx_data[i*batch_size:(i+1)*batch_size] = idxs.cpu().numpy()
gt_pressure_map[i*batch_size:(i+1)*batch_size] = gt_maps.cpu().numpy()
gt_pressure_values[i*batch_size:(i+1)*batch_size] = gt_maps.amax(dim=(1,2)).cpu().numpy()
pred_pressure_map[i*batch_size:(i+1)*batch_size] = pred_maps.cpu().numpy()
gt_center_of_pressure[i*batch_size:(i+1)*batch_size] = torch.stack((gt_x_center, gt_y_center), dim=1).cpu().numpy()
pred_center_of_pressure[i*batch_size:(i+1)*batch_size] = torch.stack((pred_x_center, pred_y_center), dim=1).cpu().numpy()
# finally, run otzu thresholding, which uses a different batch size (much larger)
intersections = []
unions = []
for batch_idx in tqdm(range(0, n_samples, otzu_batch_size)):
start_idx = batch_idx
end_idx = min(batch_idx + batch_size, n_samples)
pred = pred_pressure_map[start_idx:end_idx]
gt = gt_pressure_map[start_idx:end_idx]
# run the otzu thresholding, using multiprocessing
with Pool() as p:
otzu_maps = p.map(otzu_single, [pred[i] for i in range(pred.shape[0])])
otzu_maps = np.stack(otzu_maps)
pred_bin = otzu_maps.astype(bool)
gt_bin = gt.astype(bool)
intersection = np.logical_and(pred_bin, gt_bin)
union = np.logical_or(pred_bin, gt_bin)
intersections.append(intersection.sum(axis=(1,2)))
unions.append(union.sum(axis=(1,2)))
intersections = np.concatenate(intersections)
unions = np.concatenate(unions)
save_file.create_dataset(name='otzu_intersection',
data=intersections,
dtype=np.float32)
save_file.create_dataset(name='otzu_union',
data=unions,
dtype=np.float32)
def plot_inference_results(data_path: str,
save_folder: str,
num_samples: int = 5,
indicies: List[int] = None):
os.makedirs(save_folder, exist_ok=True)
with h5py.File(data_path, 'r') as data_file:
if indicies is None:
indicies = np.random.choice(len(data_file['predictions']), num_samples)
for idx in indicies:
ground_truth = data_file['ground_truth'][idx]
prediction = data_file['predictions'][idx]
save_path = os.path.join(save_folder, f'{data_file['indicies'][idx]}.png')
plot_pressure_maps(ground_truth, prediction, save_path)
def print_eval_metrics(results_path: str):
with h5py.File(results_path, 'r') as results_file:
mse_values = results_file['mse_values'][:]
mae_values = results_file['mae_values'][:]
otzu_intersections = results_file['otzu_intersection'][:]
otzu_unions = results_file['otzu_union'][:]
gt_total_force = results_file['gt_total_force'][:]
pred_total_force = results_file['pred_total_force'][:]
gt_center_of_pressure = results_file['gt_center_of_pressure'][:]
pred_center_of_pressure = results_file['pred_center_of_pressure'][:]
press_max_pressures = results_file['press_max_pressures'][:]
avg_max_pressure = np.mean(press_max_pressures)
avg_pressure = results_file.attrs['avg_pressure']
print(f'Average Pressure: {avg_pressure} Pa')
print(f"Max Max Pressure: {np.max(press_max_pressures)} Pa")
print(f"Average Max Pressure: {avg_max_pressure} Pa")
print(f'MAE: {mae_values.mean()} Pa')
print(f'Percent MAE: {100*mae_values.mean()/avg_max_pressure}%')
print(f'MSE: {mse_values.mean()} Pa^2')
print(f'RMSE: {np.sqrt(mse_values.mean())} Pa')
print(f'Percent RMSE: {100*np.sqrt(mse_values.mean())/avg_max_pressure}%')
print(f"Average Force Error: {100*np.mean(np.abs(gt_total_force - pred_total_force))/np.mean(gt_total_force)}%")
# only calculate IOU for pressure maps with a total force greater than 2N
mask = gt_total_force[:-1] > 2
print('mask intersection over union:', np.mean(otzu_intersections[mask]/otzu_unions[mask]))
distances = np.sqrt(np.sum((gt_center_of_pressure - pred_center_of_pressure)**2, axis=1))
avg_dist_error = distances.mean()
sensor_size = 41 # mm
image_size = results_file['gt_pressure_map'].shape[1]
print(f"Average distance error: {avg_dist_error} pix, {avg_dist_error*sensor_size/image_size} mm")
def error_location_analysis(inference_path: str):
grid_size = 4
# make 3x3 grid for analysis:
total_errors = np.zeros((grid_size, grid_size))
num_presses = np.zeros((grid_size, grid_size))
# Load the generated pressure maps
with h5py.File(inference_path, 'r') as data_file:
all_gen_pressure_maps = data_file['pred_pressure_map'][:]/1000 #kPa
all_gt_pressure_maps = data_file['gt_pressure_map'][:]/1000 #kPa
print(f'all_gen_pressure_maps shape: {all_gen_pressure_maps.shape}')
print(f'all_gt_pressure_maps shape: {all_gt_pressure_maps.shape}')
for i in tqdm(range(all_gen_pressure_maps.shape[0])):
if all_gt_pressure_maps[i].sum() == 0:
continue
for x in range(grid_size):
for y in range(grid_size):
if all_gt_pressure_maps[i][int(x*256/grid_size):int((x+1)*256/grid_size), int(y*256/grid_size):int((y+1)*256/grid_size)].sum() > 0:
num_presses[x, y] += 1
total_errors[x, y] += np.mean(np.abs(all_gen_pressure_maps[i] - all_gt_pressure_maps[i]))
matplotlib.rcParams['font.size'] = 16 # Adjust to change default font size for all text
plt.figure(figsize=(5, 5))
plt.imshow(total_errors/num_presses, vmin=0)
cbar = plt.colorbar(shrink=0.8)
cbar.ax.tick_params(labelsize=14)
cbar.set_label('Error (kPa)', rotation=270, labelpad=20)
plt.xticks([])
plt.yticks([])
plt.savefig('Error_vs_Location.jpg', dpi = 300)
print('max error:', np.max(total_errors/num_presses))
def make_horizon_graph(inference_paths: List[str]):
mae_values = []
gt_total_forces = []
IOUs = []
avg_distances = []
n_evals = len(inference_paths)
for inference_path in inference_paths:
with h5py.File(inference_path, 'r') as f:
mae_values.append(f['mae_values'][:].mean())
gt_total_forces.append(f['gt_total_force'][:])
# only calculate IOU for pressure maps with a max pressure greater than 2N
mask = gt_total_forces[-1][:-1] > 2
IOUs.append(np.mean(f['otzu_intersection'][mask]/f['otzu_union'][mask]))
distances = np.sqrt(np.sum((f['gt_center_of_pressure'][:] - f['pred_center_of_pressure'][:])**2, axis=1))
avg_distances.append(distances.mean())
# plot the normalized MAE and IOU
normalized_MAE = mae_values/mae_values[0]*100
normalized_IOU = IOUs/IOUs[0]*100
normalized_dist = avg_distances/avg_distances[0]*100
# print the average and final percentage values:
print("Average Percentage Values:")
print(f"MAE: {np.mean(normalized_MAE)}")
print(f"IOU: {np.mean(normalized_IOU)}")
print(f"Distance: {np.mean(normalized_dist)}")
print("Final Percentage Values:")
print(f"MAE: {normalized_MAE[-1]}")
print(f"IOU: {normalized_IOU[-1]}")
print(f"Distance: {normalized_dist[-1]}")
plt.figure()
# plt.title("Normalized Performace Over Time")
plt.plot(range(0, n_evals), normalized_MAE, label='MAE Error')
plt.plot(range(0, n_evals), normalized_IOU, label='IOU Errror')
plt.plot(range(0, n_evals), normalized_dist, label='Distance Error')
plt.plot(range(0, n_evals), np.ones(n_evals)*100, '--', c = 'gray')
plt.xlabel('Hours')
plt.ylabel('Percent of Initial Value')
plt.legend()
plt.show()
if __name__ == '__main__':
checkpoint_path = "data/12_hr_75/trained_models/hours_1_to_3_same_as_paper_0/checkpoints/checkpoint_99.pt"
data_path = 'data/12_hr_75/hours_3_to_4.hdf5'
save_folder = 'data/12_hr_75/trained_models/hours_1_to_3_same_as_paper_0/eval'
save_name = "3_to_4.hdf5"
save_path = os.path.join(save_folder, save_name)
run_inference(checkpoint_path, data_path, save_path, indicies='all')
print_eval_metrics(save_path)