|
| 1 | +import yaml |
| 2 | +import os |
| 3 | +import seaborn as sn |
| 4 | +import numpy as np |
| 5 | +import kwcoco |
| 6 | +import matplotlib.pyplot as plt |
| 7 | +from sklearn.metrics import confusion_matrix |
| 8 | +import scipy.ndimage as ndi |
| 9 | + |
| 10 | +def sanitize_str(str_: str): |
| 11 | + """ |
| 12 | + Convert string to lowercase and emove trailing whitespace and period. |
| 13 | +
|
| 14 | + :param str_: Input text |
| 15 | +
|
| 16 | + :return: ``str_`` converted to lowercase and stripped of trailing whitespace and period. |
| 17 | + :rtype: str |
| 18 | + """ |
| 19 | + return str_.lower().strip(" .") |
| 20 | + |
| 21 | +def plot_positive_GT_conf_distributions(activity_confs, activity_gt): |
| 22 | + """ |
| 23 | + plot_TP_conf_distributions: |
| 24 | + For each activity, plot the distribution of confidences when ground |
| 25 | + truth indicates that activity is happening. |
| 26 | +
|
| 27 | + i.e.: for activity x, for frames in which ground truth = x, plot |
| 28 | + the distribution of confidences. |
| 29 | +
|
| 30 | + Inputs: |
| 31 | + activity_confs: frames x class-wise-confidences. Given a kwcoco |
| 32 | + dataset called "coco": |
| 33 | + ``` |
| 34 | + activity_confs = torch.asarray(coco.images().lookup("activity_conf")) |
| 35 | + ``` |
| 36 | + (49K x 25 for coffee val set.) |
| 37 | + activity_gt: frames x ground truth activity_id. |
| 38 | + Given a kwcoco dataset called "coco": |
| 39 | + ``` |
| 40 | + activity_gt = torch.asarray(coco.images().lookup("activity_gt")) |
| 41 | + ``` |
| 42 | + """ |
| 43 | + |
| 44 | + sns.set_theme(style="white", rc={"axes.facecolor": (0, 0, 0, 0)}) |
| 45 | + |
| 46 | + # Get data together |
| 47 | + true_confs = [float(activity_confs[i,truth_ind]) for i, truth_ind in enumerate(activity_gt)] |
| 48 | + data = {"true_conf":true_confs, "gt":activity_gt} |
| 49 | + df = pd.DataFrame(data) |
| 50 | + |
| 51 | + false_confs = np.array([[a for i, a in enumerate(act_conf) if i != gt] for act_conf, gt in zip(activity_confs, activity_gt)]).flatten() |
| 52 | + false_gt = np.array([[gt for i, a in enumerate(act_conf) if i != gt] for act_conf, gt in zip(activity_confs, activity_gt)]).flatten() |
| 53 | + data_opposite = {"true_conf":false_confs, "gt":false_gt} |
| 54 | + df_opposite = pd.DataFrame(data_opposite) |
| 55 | + |
| 56 | + def plot(df): |
| 57 | + # Initialize the FacetGrid object |
| 58 | + pal = sns.cubehelix_palette(10, rot=-.25, light=.7) |
| 59 | + g = sns.FacetGrid(df, row="gt", hue="gt", aspect=15, height=.5, palette=pal) |
| 60 | + |
| 61 | + # Draw the densities in a few steps |
| 62 | + g.map(sns.kdeplot, "true_conf", |
| 63 | + bw_adjust=.5, clip_on=False, |
| 64 | + fill=True, alpha=1, linewidth=1.5) |
| 65 | + g.map(sns.kdeplot, "true_conf", clip_on=False, color="w", lw=2, bw_adjust=.5) |
| 66 | + |
| 67 | + # passing color=None to refline() uses the hue mapping |
| 68 | + g.refline(y=0, linewidth=2, linestyle="-", color=None, clip_on=False) |
| 69 | + |
| 70 | + # Define and use a simple function to label the plot in axes coordinates |
| 71 | + def label(x, color, label): |
| 72 | + ax = plt.gca() |
| 73 | + ax.text(0, .2, label, fontweight="bold", color=color, |
| 74 | + ha="left", va="center", transform=ax.transAxes) |
| 75 | + g.map(label, "true_conf") |
| 76 | + |
| 77 | + # Set the subplots to overlap |
| 78 | + g.figure.subplots_adjust(hspace=-.25) |
| 79 | + |
| 80 | + # Remove axes details that don't play well with overlap |
| 81 | + g.set_titles("") |
| 82 | + g.set(yticks=[], ylabel="") |
| 83 | + g.despine(bottom=True, left=True) |
| 84 | + |
| 85 | + # save |
| 86 | + plt.savefig("./outputs/plot_positive_GT_conf_distributions.png") |
| 87 | + |
| 88 | + |
| 89 | +def bilateralFtr1D(y, sSpatial = 5, sIntensity = 1): |
| 90 | + ''' |
| 91 | + The equation of the bilateral filter is |
| 92 | + |
| 93 | + ( dx ^ 2 ) ( dI ^2 ) |
| 94 | + F = exp (- ----------------- ) * exp (- ------------------- ) |
| 95 | + ( sigma_spatial ^ 2 ) ( sigma_Intensity ^ 2 ) |
| 96 | + ~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 97 | + This is a guassian filter! |
| 98 | + dx - The 'geometric' distance between the 'center pixel' and the pixel |
| 99 | + to sample |
| 100 | + dI - The difference between the intensity of the 'center pixel' and |
| 101 | + the pixel to sample |
| 102 | + sigma_spatial and sigma_Intesity are constants. Higher values mean |
| 103 | + that we 'tolerate more' higher value of the distances dx and dI. |
| 104 | + |
| 105 | + Dependencies: numpy, scipy.ndimage.gaussian_filter1d |
| 106 | + |
| 107 | + calc gaussian kernel size as: filterSize = (2 * radius) + 1; radius = floor (2 * sigma_spatial) |
| 108 | + y - input data |
| 109 | + ''' |
| 110 | + |
| 111 | + # gaussian filter and parameters |
| 112 | + radius = np.floor (2 * sSpatial) |
| 113 | + filterSize = ((2 * radius) + 1) |
| 114 | + ftrArray = np.zeros(int(filterSize)) |
| 115 | + ftrArray[int(radius)] = 1 |
| 116 | + |
| 117 | + # Compute the Gaussian filter part of the Bilateral filter |
| 118 | + gauss = ndi.gaussian_filter1d(ftrArray, sSpatial) |
| 119 | + |
| 120 | + # 1d data dimensions |
| 121 | + width = y.size |
| 122 | + |
| 123 | + # 1d resulting data |
| 124 | + ret = np.zeros (width) |
| 125 | + |
| 126 | + for i in range(width): |
| 127 | + |
| 128 | + ## To prevent accessing values outside of the array |
| 129 | + # The left part of the lookup area, clamped to the boundary |
| 130 | + xmin = max(i - radius, 1); |
| 131 | + # How many columns were outside the image, on the left? |
| 132 | + dxmin = xmin - (i - radius); |
| 133 | + |
| 134 | + # The right part of the lookup area, clamped to the boundary |
| 135 | + xmax = min(i + radius, width); |
| 136 | + # How many columns were outside the image, on the right? |
| 137 | + dxmax = (i + radius) - xmax; |
| 138 | + |
| 139 | + # The actual range of the array we will look at |
| 140 | + area = y [int(xmin):int(xmax)] |
| 141 | + |
| 142 | + # The center position |
| 143 | + center = y[i] |
| 144 | + |
| 145 | + # The left expression in the bilateral filter equation |
| 146 | + # We take only the relevant parts of the matrix of the |
| 147 | + # Gaussian weights - we use dxmin, dxmax, dymin, dymax to |
| 148 | + # ignore the parts that are outside the image |
| 149 | + expS = gauss[int((1+dxmin)):int((filterSize-dxmax))] |
| 150 | + |
| 151 | + # The right expression in the bilateral filter equation |
| 152 | + dy = y [int(xmin):int(xmax)] - y[i] |
| 153 | + dIsquare = (dy * dy) |
| 154 | + expI = np.exp (- dIsquare / (sIntensity * sIntensity)) |
| 155 | + |
| 156 | + # The bilater filter (weights matrix) |
| 157 | + F = expI * expS |
| 158 | + |
| 159 | + # Normalized bilateral filter |
| 160 | + Fnormalized = F / sum(F) |
| 161 | + |
| 162 | + # Multiply the area by the filter |
| 163 | + tempY = y [int(xmin):int(xmax)] * Fnormalized |
| 164 | + |
| 165 | + # The resulting pixel is the sum of all the pixels in |
| 166 | + # the area, according to the weights of the filter |
| 167 | + # ret(i,j,R) = sum (tempR(:)) |
| 168 | + ret[i] = sum (tempY) |
| 169 | + |
| 170 | + return ret |
| 171 | + |
| 172 | + |
| 173 | +def get_average_TP_activations(coco): |
| 174 | + # For each activity, given the Ground Truth-specified |
| 175 | + # frame subset where that activity is happening, get the |
| 176 | + # average activation of that class. |
| 177 | + |
| 178 | + all_activity_ids = np.unique(np.asarray(coco.images().lookup('activity_gt'))) |
| 179 | + all_vid_ids = np.unique(np.asarray(coco.images().lookup('video_id'))) |
| 180 | + |
| 181 | + avg_probs = np.zeros(max(all_activity_ids) + 1) |
| 182 | + |
| 183 | + for activity_id in all_activity_ids: |
| 184 | + #image_ids = coco.index.vidid_to_gids[vid_id] |
| 185 | + image_ids = [img['id'] for img in coco.videos(video_ids=all_vid_ids).images[0].objs if img['activity_gt'] == activity_id] |
| 186 | + sub_dset = coco.subset(gids=image_ids, copy=True) |
| 187 | + probs_for_true_inds = np.asarray( |
| 188 | + sub_dset.images().lookup("activity_conf"))[:,activity_id] |
| 189 | + avg_prob = np.mean(probs_for_true_inds) |
| 190 | + avg_probs[activity_id] = avg_prob |
| 191 | + |
| 192 | + return avg_probs |
| 193 | + |
| 194 | +config_fn = "config/tasks/task_steps_cofig-recipe-coffee-shortstrings.yaml" |
| 195 | +with open(config_fn, "r") as stream: |
| 196 | + config = yaml.safe_load(stream) |
| 197 | +labels = [sanitize_str(l["description"]) for l in config["steps"]] |
| 198 | +steps = config['steps'] |
| 199 | +if steps[0]['id'] == 1: |
| 200 | + config['steps'].insert(0, {'id':0, |
| 201 | + 'activity_id':0, |
| 202 | + 'description':'background', |
| 203 | + 'median_duration_seconds':0.5, |
| 204 | + 'mean_conf':0.5, |
| 205 | + 'std_conf':0.2, |
| 206 | + }) |
| 207 | + |
| 208 | +coco_val = kwcoco.CocoDataset("model_files/val_activity_preds_epoch40.mscoco.json") |
| 209 | +coco_test = kwcoco.CocoDataset("model_files/test_activity_preds.mscoco.json") |
| 210 | + |
| 211 | +image_ids = coco_test.index.vidid_to_gids[3] |
| 212 | +video_dset = coco_test.subset(gids=image_ids, copy=True) |
| 213 | + |
| 214 | +# "Training": for each activity class, see what the average "true positive" |
| 215 | +# activation was. |
| 216 | +avg_probs = get_average_TP_activations(coco_test) |
| 217 | +print(f"average_probs = {avg_probs}") |
| 218 | + |
| 219 | +all_vid_ids = np.unique(np.asarray(coco_val.images().lookup('video_id'))) |
| 220 | + |
| 221 | +for vid_id in all_vid_ids: |
| 222 | + print(f"vid_id {vid_id}") |
| 223 | + |
| 224 | + image_ids = coco_test.index.vidid_to_gids[vid_id] |
| 225 | + video_dset = coco_test.subset(gids=image_ids, copy=True) |
| 226 | + |
| 227 | + # All N activity confs x each video frame |
| 228 | + activity_confs = video_dset.images().lookup("activity_conf") |
| 229 | + |
| 230 | + next_step = 1 |
| 231 | + step_predictions = [] |
| 232 | + num_frames_activated = 0 |
| 233 | + |
| 234 | + # Predicted step: confidence has been above threshold for 5 frames. |
| 235 | + threshold_frame_count = 8 |
| 236 | + for i, activity_conf in enumerate(activity_confs): |
| 237 | + |
| 238 | + # Check if we're done: if so, append last step & continue |
| 239 | + if next_step == len(steps): |
| 240 | + step_predictions.append(next_step-1) |
| 241 | + continue |
| 242 | + # Next step |
| 243 | + next_activity_id = steps[next_step]['activity_id'] |
| 244 | + next_next_activity_id = steps[min(len(steps)-1,next_step + 1)][ |
| 245 | + 'activity_id'] |
| 246 | + |
| 247 | + next_activity_conf = activity_conf[next_activity_id] |
| 248 | + next_next_activity_conf = activity_conf[next_next_activity_id] |
| 249 | + |
| 250 | + avg_prob_next_activity = avg_probs[next_activity_id] |
| 251 | + avg_prob_next_next_activity = avg_probs[next_next_activity_id] |
| 252 | + ''' |
| 253 | + if next_activity_id == 16 and vid_id == 2: |
| 254 | + print(f"next_activity_id = {next_activity_id}") |
| 255 | + print(f"avg_prob_next_activity = {avg_prob_next_activity}") |
| 256 | + ''' |
| 257 | + if i > 15: |
| 258 | + threshold_frame_count = 16 |
| 259 | + |
| 260 | + if next_activity_conf > 0.8 * avg_prob_next_activity: |
| 261 | + num_frames_activated += 1 |
| 262 | + ''' |
| 263 | + if next_activity_id == 16 and vid_id == 2: |
| 264 | + print(f"num_frames_activated = {num_frames_activated}. prob = {next_activity_conf}") |
| 265 | + ''' |
| 266 | + else: |
| 267 | + num_frames_activated = 0 |
| 268 | + if next_next_activity_conf > 0.8 * avg_prob_next_activity: |
| 269 | + num_skip2_frames_activated += 1 |
| 270 | + else: |
| 271 | + num_skip2_frames_activated = 0 |
| 272 | + |
| 273 | + if num_frames_activated >= threshold_frame_count: |
| 274 | + #if next_step < 23: |
| 275 | + #next_step += 1 |
| 276 | + next_step += 1 |
| 277 | + num_frames_activated = 0 |
| 278 | + num_skip2_frames_activated = 0 |
| 279 | + elif num_skip2_frames_activated >= threshold_frame_count: |
| 280 | + next_step = min(next_step + 2, len(steps)) |
| 281 | + num_frames_activated = 0 |
| 282 | + num_skip2_frames_activated = 0 |
| 283 | + print("hit a skip-step!!") |
| 284 | + |
| 285 | + step_predictions.append(next_step-1) |
| 286 | + |
| 287 | + # Ground truth step: |
| 288 | + activity_gts = video_dset.images().lookup("activity_gt") |
| 289 | + step_gts = [] |
| 290 | + step_gts_no_background = [] |
| 291 | + current_step = 0 |
| 292 | + for activity_gt in activity_gts: |
| 293 | + # convert activity id to step id |
| 294 | + step_id = next(int(item['id']) for item in steps if item['activity_id'] == activity_gt) |
| 295 | + step_gts.append(step_id) |
| 296 | + |
| 297 | + # A version of GT that never jumps back to 0 |
| 298 | + if step_id > 0: |
| 299 | + current_step = step_id |
| 300 | + step_gts_no_background.append(current_step) |
| 301 | + |
| 302 | + |
| 303 | + # Plot confusion matrix |
| 304 | + fig, ax = plt.subplots(figsize=(100, 100)) |
| 305 | + cm = confusion_matrix(step_gts_no_background, step_predictions, normalize="true") |
| 306 | + sn.heatmap(cm, annot=True, fmt="0.0%", ax=ax, linewidth=.5) |
| 307 | + sn.set(font_scale=4) |
| 308 | + ax.set( |
| 309 | + title="Confusion Matrix", |
| 310 | + xlabel="Predicted Label", |
| 311 | + ylabel="True Label",) |
| 312 | + fig.savefig(f"./outputs/plot_confusion_mat_vid{vid_id}.png") |
| 313 | + |
| 314 | + # Plot gt vs predicted class across all vid frames |
| 315 | + fig = plt.figure() |
| 316 | + sn.set(font_scale=1) |
| 317 | + step_gts = [float(i) for i in step_gts] |
| 318 | + plt.plot(step_gts, label = 'gt') |
| 319 | + plt.plot(step_predictions, label = 'estimated') |
| 320 | + #plt.plot(inliers-0.5, label = 'inliers') |
| 321 | + plt.plot(10*np.asarray(activity_confs)[:,17]-5, label = 'act_preds[17]') |
| 322 | + plt.plot(10*np.asarray(activity_confs)[:,18]-5, label = 'act_preds[18]') |
| 323 | + plt.plot(10*np.asarray(activity_confs)[:,19]-5, label = 'act_preds[19]') |
| 324 | + |
| 325 | + plt.plot(bilateralFtr1D(10*np.asarray(activity_confs)[:,17])-10, label = 'act_preds_bilateral[17]') |
| 326 | + plt.plot(bilateralFtr1D(10*np.asarray(activity_confs)[:,18])-10, label = 'act_pred_bilateral[18]') |
| 327 | + plt.plot(bilateralFtr1D(10*np.asarray(activity_confs)[:,19])-10, label = 'act_preds_bilateral[19]') |
| 328 | + #plt.plot(10*X_conf_incremental, label = 'confidence') |
| 329 | + #plt.plot(10*vid_acts[:,10], label = act_labels[10]) |
| 330 | + #plt.plot(10*vid_acts[:,11], label = act_labels[11]) |
| 331 | + #plt.plot(10*vid_acts[:,12], label = act_labels[12]) |
| 332 | + plt.legend() |
| 333 | + fig.savefig(f"./outputs/plot_pred_vs_gt_vid{vid_id}.png") |
| 334 | + |
| 335 | + if False: |
| 336 | + plot_positive_GT_conf_distributions(activity_confs, activity_gt) |
| 337 | + |
| 338 | + |
| 339 | + |
0 commit comments