-
Notifications
You must be signed in to change notification settings - Fork 1
/
app.py
1195 lines (1023 loc) · 47.6 KB
/
app.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
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# -*- coding: utf-8 -*-
import dash
import dash_core_components as dcc
import dash_html_components as html
from dash.dependencies import Input, Output, State
import pandas as pd
import plotly.graph_objs as go
import plotly.plotly as py
from plotly import tools
import matplotlib.pyplot as plt
import PIL
import h5py
import numpy as np
import os, base64
from io import BytesIO
from scipy.stats import entropy
import gym
from visualize_atari import *
external_stylesheets = ['https://codepen.io/chriddyp/pen/bWLwgP.css']
app = dash.Dash(__name__, external_stylesheets=external_stylesheets)
server = app.server
# Total training log
log_data = pd.read_csv("baby-a3c/breakout-v4/log-model7-02-17-20-41.txt")
log_data.columns = log_data.columns.str.replace(" ", "")
# Load data into memory by default, otherwise delete 'driver='core'' to read from disk
replays = h5py.File('static/model_rollouts_5.h5','r', driver='core')
# list of epoch numbers we took (number * 500k is the number of frames trained at that point)
snapshots = [1,19,30,40,50,60,70,80,90,100]
# HTML Page layout
app.layout = html.Div(children=[
html.H5(children='Interactive Atari RL', id = 'null',style = {'padding-bottom':'0px', 'margin':'0'}),
# Top Row
html.Div([
html.Div([
dcc.Graph(id = 'rewards-candlestick',
style={'height':'21em'}
)
], style={'border':'1px solid black', 'display':'inline-block'}),
html.Div([
dcc.Graph(id = 'all-cum-rewards',
style={'height':'21em'}
)
], style={'border':'1px solid black', 'display':'inline-block'}),
], style = {'height':'23em','overflow':'auto','display':'block','width':'150em'}),
html.Div([ # Big bottom group with 3 columns
html.Div([ # column 1
html.Div([
html.Div(style = {'padding-top':'0em',
'padding-left':'0',
},
id = 'info-box-epoch'
),
html.Div(style = {'padding-top':'0em',
'padding-left':'0',
},
id = 'info-box-frame'
)
],
style={'border':'1px solid black', 'display':'inline-block'}),
html.Div([
html.Button('5 frames back', id='back-frame', style={'display':'inline-block','width':'50%'}),
html.Button('5 frames next', id='forward-frame', style={'display':'inline-block','width':'50%'}),
],style={'border':'1px solid black', 'display':'block'}),
html.Div(html.Img(id = 'screen-ins',
style = {'max-width':'100%', 'max-height':'100%','height':'30em'}),
style = {'border':'1px solid black','width':'20em','height':'100%'}),
], style = {'position':'absolute','top':'8px','border':'1px solid black', 'display':'inline-block','width':'20em','border-bottom':'20em'}),
html.Div([ # column 2a (gantt and parallel coords plots); (data timeline plots referred to 'gantt' in code because it was my original name for these plots as I thought they looked like gantt charts)
html.Div([
dcc.Graph(id = 'gantt',
style={'border':'1px solid black','height':'15em','display':'block'})
], style = {'width':'60em'}),
html.Div([
dcc.Graph(id = 'gantt2',
style={'border':'1px solid black','height':'15em','display':'block'})
], style = {'width':'60em'}),
html.Div([
dcc.Graph(id = 'parallel-sal',
style={'border':'1px solid black','height':'30em','display':'block'})
], style = {'width':'60em'}),
], style = {'position':'absolute','margin-left':'20em','border':'1px solid black', 'display':'inline-block'}),
html.Div([ # column 2b (dropdowns)
html.Div(["Select episodes for 'gantt' and parallel coordinates plots"], style = {'word-wrap':'break-word'}),
html.Div([
dcc.Dropdown(
id='gantt-select1',
options=[{'label':x, 'value':x} for x in snapshots],
value='90'
),
dcc.Dropdown(
id='gantt-select2',
options=[{'label':x, 'value':x} for x in snapshots],
value='60'
),], style = {'padding-bottom':'15em'}),
], style = {'position':'absolute','margin-left':'80em','width':'6em','border':'1px solid black', 'display':'inline-block'}),
html.Div([ # column 3 (saliency by region heatmap and subplots)
html.Div([
dcc.Graph(id = 'regions_bars',
style = {'border':'1px solid black', 'height':'23em'})
]),
html.Div([
dcc.Graph(id = 'regions-subplots',
style = {'height':'37em'})
], style = {'display':'block','border':'1px solid black'})
], style = {'position':'absolute','border':'1px solid black', 'display':'inline-block','margin-left':'86em'}),
html.Div([ # column 4 (action and trajectory plots)
html.Div([
dcc.Graph(id = 'actions',
style={'border':'1px solid black','height':'23em'})
], ),
html.Div([
dcc.Graph(id = 'trajectory',
style = {'border':'1px solid black'})
])
], style = {'position':'absolute','margin-left':'135em','border':'1px solid black', 'display':'inline-block'}),
], style = {'position':'relative', 'display':'block'}),
html.Div([ # Row 3 (sliders)
html.Div([
html.Div(id='frame-val'),
dcc.Slider(id='frame-slider',
min = 0,
max = 3000,
value = 0,
marks = {i: str(i) for i in range(0, 3000, 100)},
step = 5
)
], style = {'padding-bottom':'50px', 'padding-left':'10px'}),
html.Div([
html.Div(id='snapshot-val'),
dcc.Slider(id='snapshot-slider',
min = 1,
max = 100,
value = 50,
marks = {i: str(i) for i in [1,10,19,30,40,50,60,70,80,90,100]},
step = None
)
])
], style={'position':'relative','padding-bottom':'20px','margin-top':'65em'}),
html.Div([ # Row 4 (Action entropy (sort of legacy))
dcc.Graph(id = 'action-entropy')
], style={'border':'1px solid black'}),
html.Div(50, id = 'current-frame'), # 'hidden' div to store current frame as state
html.Div(0, id = 'gantt-mem') # 'hidden' div to keep track of last frame called from first gantt (needed since second gantt plot couldn't update frame otherwise)
])
#====================== Begin callback functions, serving as the 'backend' to control all plotting =================#
# Each callback function is decorated with the named HTML components used for inputs and outputs
# https://dash.plot.ly/getting-started-part-2
# Update written values on page with the frame slider
@app.callback(
[Output(component_id='frame-val', component_property='children'),
Output(component_id='current-frame', component_property='children'),
Output(component_id='gantt-mem', component_property='children')],
[Input(component_id='frame-slider', component_property='value')]
)
def update_frame_slider(input_value):
return 'Frame number of episode: {}'.format(input_value), input_value, input_value
# Update written values on page with the snapshot slider and mark unavailable values as red on the frame slider
@app.callback(
[Output(component_id='snapshot-val', component_property='children'),
Output(component_id='frame-slider', component_property='marks')],
[Input(component_id='snapshot-slider', component_property='value')]
)
def update_snapshot_slider(snapshot):
length = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0/logits'].shape[0]
d = {i: str(i) for i in range(0, 3000, 100)}
for k in d:
if k > length:
d[k] = {'label': k, 'style':{'color': '#f50'}}
return 'Model iteration (500k frame increments): {}\n Ep Length {}'.format(snapshot, length), d
# Update info box above frame display with current snapshot val
@app.callback(
Output(component_id='info-box-epoch', component_property='children'),
[Input(component_id='snapshot-slider', component_property='value')]
)
def update_info_box(input_value):
return f'Epoch selected: {input_value}'
# Update info box above frame display with current frame val
@app.callback(
Output(component_id='info-box-frame', component_property='children'),
[Input(component_id='frame-slider', component_property='value')]
)
def update_info_box_frame(input_value):
return f'Frame selected: {input_value}'
# Handle all links controlling snapshots by passing to the snapshot slider
@app.callback(
Output(component_id='snapshot-slider', component_property='value'),
[Input(component_id='action-entropy', component_property = 'clickData'),
Input('rewards-candlestick', 'clickData'),
Input(component_id='gantt', component_property='clickData'),
Input(component_id='gantt2', component_property='clickData')],
[ State('gantt-mem', 'children'),
State(component_id='gantt-select1', component_property='value'),
State(component_id='gantt-select2', component_property='value')]
)
def update_link_snapshot(entropy_click, candle_click, gantt_click1, gantt_click2, gantt_click_memory, gantt_epoch1, gantt_epoch2):
if entropy_click:
return (entropy_click['points'][0]['x'])
if candle_click:
return (candle_click['points'][0]['x'])
print(gantt_click1, gantt_click_memory)
# Prevent first gantt's clicks from re-updating if it didn't change
if gantt_click1 and myround(gantt_click1['points'][0]['x']) != gantt_click_memory:
print(gantt_epoch1)
return gantt_epoch1
if gantt_click2:
print(gantt_epoch2)
return gantt_epoch2
return 50
# Round to multiple of 5
def myround(x, base=5):
return base * round(x/base)
# Handle all links controlling frame by passing to the frame slider
@app.callback(
Output(component_id='frame-slider', component_property='value'),
[Input(component_id='regions-subplots', component_property = 'clickData'),
Input('actions', 'clickData'),
Input('trajectory', 'clickData'),
Input('regions_bars', 'clickData'),
Input('gantt', 'clickData'),
Input('gantt2', 'clickData'),
Input(component_id='back-frame', component_property='n_clicks'),
Input(component_id='forward-frame', component_property='n_clicks'),
],
[State(component_id='current-frame', component_property='children'),
State('gantt-mem', 'children'),],
)
def update_link_frame(regions_click, actions_click, trajectory_click, bars_click,
gantt_click1, gantt_click2,
back_click, forward_click,
cur_frame, gantt_click_memory):
ctx = dash.callback_context
# Check which buttons was pressed (since we only have n_clicks data)
for item in ctx.triggered:
if 'back-frame' in item['prop_id'] and item['value']:
return max(0, cur_frame - 5)
if 'forward-frame' in item['prop_id'] and item['value']:
return cur_frame +5
# Else, check through inputs of specified plots
if regions_click:
return (regions_click['points'][0]['x'])
if actions_click:
return myround(actions_click['points'][0]['x'])
if trajectory_click:
return myround(trajectory_click['points'][0]['x'])
if bars_click:
return (bars_click['points'][0]['x'])
# While not buttons, clicking on lines in gantt plots need to be put here for some response errors
if gantt_click1 and myround(gantt_click1['points'][0]['x']) != gantt_click_memory:
return myround((gantt_click1['points'][0]['x']))
if gantt_click2:
print(gantt_click2)
return myround((gantt_click2['points'][0]['x']))
return 50
# Control rewards candlestick chart (row 1, col 1)
@app.callback(
Output(component_id='rewards-candlestick', component_property='figure'),
[Input(component_id='null', component_property='children')]
)
def update_rewards_candlestick(start):
# Get data in increments of 40 (less noise)
epr_xrange = (log_data['frames']/500e3).values[::40]
epr_vals = log_data['mean-epr'].values[::40]
rewards_candle_hovertext = [str(i) for i in epr_vals[:-1]]
# Derive plot with data
trace = go.Ohlc(x = epr_xrange,
open = epr_vals[:-1],
high = epr_vals[:-1],
low = epr_vals[1:],
close = epr_vals[1:],
text = rewards_candle_hovertext,
hoverinfo = 'x+text',
name = 'Mean EPR')
data = [trace]
# Also accumulate total saliency per episode for both types
saliency_toplevel = []
for s in snapshots:
print(s)
history = replays['models_model7-02-17-20-41/model.'+str(s)+'.tar/history/0']
actor_frames, critic_frames = history['actor_sal'].value, history['critic_sal'].value
actor_tot_perframe = actor_frames.sum((1,2)).sum()/actor_frames.shape[0]
critic_tot_perframe = critic_frames.sum((1,2)).sum()/critic_frames.shape[0]
saliency_toplevel.append([actor_tot_perframe, critic_tot_perframe])
saliency_toplevel = np.array(saliency_toplevel)
# Write saliency data into bars
actor_bars = go.Bar(
x = snapshots,
y = saliency_toplevel[:,0],
name = 'A Sal /frame',
yaxis = 'y2',
marker = dict(
color = 'rgba(68, 68, 230, 0.7)'
),
hoverinfo = 'none'
)
critic_bars = go.Bar(
x = snapshots,
y = saliency_toplevel[:,1],
name = 'C Sal /frame',
yaxis='y2',
marker = dict(
color = 'rgba(240, 68, 68, 0.7)'
),
hoverinfo = 'none'
)
data += [actor_bars, critic_bars]
layout = go.Layout(
title = "Mean episode reward",
xaxis = dict(
title = "Frames (500k)",
rangeslider = dict(
visible = False
)
),
yaxis = dict(
title = "Reward"
),
yaxis2=dict(
title='Total saliency per frame',
overlaying='y',
side='right',
range = [0, 1200e3],
rangemode = 'nonnegative'
),
margin = dict(
l = 50,
r = 60,
b = 35,
t = 30,
pad = 4
),
legend = dict(x = 0.2, y = 1)
)
figure = go.Figure(data = data, layout = layout)
return figure
# Control cumulative rewards plot (row 1, col 2)
# Input is 'null' as it stays static
@app.callback(
Output(component_id='all-cum-rewards', component_property='figure'),
[Input(component_id='null', component_property='children')]
)
def update_all_cum_rewards(null):
data = []
for s in snapshots:
rewards = replays['models_model7-02-17-20-41/model.'+str(s)+'.tar/history/0/reward'].value
reward_trace = dict(
y = np.cumsum(rewards),
x = list(range(0, rewards.shape[0])),
name = f'Epoch {s}',
line = dict(width = 3)
)
data.append(reward_trace)
layout = go.Layout(title = 'Cumulative reward by epoch replay',
xaxis=dict(title='Episode Frame'),
yaxis=dict(title='Reward'),
margin = dict(
l = 50,
r = 40,
b = 35,
t = 30,
pad = 4
),
#legend = dict(x = 0.1, y = 1),
showlegend = False
)
figure = go.Figure(data = data, layout= layout)
return figure
# Control action entropy graph at bottom (row 4, col 1)
@app.callback(
Output(component_id='action-entropy', component_property='figure'),
[Input(component_id='null', component_property='children')]
)
def update_actions_entropy(null):
return go.Figure() # turn off
# Get list of available snapshots
iterations = sorted([int(x.split('.')[1]) for x in list(replays['models_model7-02-17-20-41'].keys())])
y_data = []
ep_lengths = {}
x_range = []
y_segments = {i:[] for i in range(4)}
avg_len = 20
actions = ['NOOP', 'FIRE', 'RIGHT', 'LEFT']
# For each iteration, get logits and scale on x-axis with respect to length of episode
# Logits are converted to moving averages of window size 10
for i in iterations:
softmax_logits = replays['models_model7-02-17-20-41/model.'+str(i)+'.tar/history/0/outs'].value
y_data.append(softmax_logits)
ep_lengths[i] = len(softmax_logits)
x_range.append(10*np.arange(0, (len(softmax_logits)/avg_len)))
ids = np.arange(len(softmax_logits))//avg_len
for a in range(4):
y_segments[a].append(np.bincount(ids,softmax_logits[:,a])/np.bincount(ids))
# Also get entropy data - entropy is calculated by action for the entire episode
entropy_data = np.array([entropy(logits) for logits in y_data])
softmax_data = np.vstack(y_data)
# Put stacked logits segments together
x_range = np.hstack(x_range)
for a in range(4):
y_segments[a] = np.hstack(y_segments[a])
# Create plot
ids = np.arange(len(softmax_data))//avg_len
series = [entropy_data[:, i] for i in range(4)]
data = []
# Iterate by action type
for i, t in enumerate(series):
# plot entropy as lines
trace = go.Scatter(
y = t,
x = iterations,
name = actions[i]
)
data.append(trace)
# Plot moving averaged logits as stacked lines
averaged = np.bincount(ids,softmax_data[:,i])/np.bincount(ids)
trace2 = dict(
x = 101*np.arange(0, (len(softmax_data)/avg_len), 1/(len(softmax_data)/avg_len)),
y = averaged,
mode = 'lines',
line = dict(width=0.5),
stackgroup = 'one',
yaxis='y2',
name = actions[i]
)
data.append(trace2)
layout = go.Layout(title = 'Entropy by Action per Iteration Episode',
height = 300,
yaxis=dict(
title='Entropy'
),
yaxis2=dict(
title='Softmax',
overlaying='y',
side='right'
),
clickmode = 'event+select')
figure = go.Figure(data = data, layout = layout)
return figure
# Control actions plot of stacked logits and cumulative reward per episode (row 1, column 4)
@app.callback(
Output(component_id='actions', component_property='figure'),
[Input(component_id='snapshot-slider', component_property='value')]
)
def update_actions(snapshot):
softmax_logits = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0/outs'].value
traces = []
actions = ['NOOP', 'FIRE', 'RIGHT', 'LEFT']
# Just plot each logit for each frame
for a in range(softmax_logits.shape[1]):
trace = dict(
x = list(range(0, softmax_logits.shape[0])),
y = softmax_logits[:, a],
hoverinfo = 'x+y',
mode = 'lines',
line = dict(width=0.5),
stackgroup = 'one',
name = actions[a]
)
traces.append(trace)
# Also plot cumulative rewards by frame
rewards = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0/reward'].value
reward_trace = dict(
y = np.cumsum(rewards),
x = list(range(0, softmax_logits.shape[0])),
name = 'Ep Reward',
yaxis="y2",
line = dict(width = 3)
)
traces.append(reward_trace)
layout = go.Layout(xaxis=dict(title='frame'),
yaxis=dict(title='Softmax value'),
yaxis2=dict(
title='Rewards',
overlaying='y',
side='right'
),
margin = dict(
l = 50,
r = 40,
b = 35,
t = 30,
pad = 4
),
legend = dict(
orientation = "h",
y = 1.13)
)
figure = go.Figure(data = traces, layout= layout)
return figure
# Function from Greydanus to upscale saliency values into visible blots of blue/red
def saliency_on_frame_abbr(S, frame, fudge_factor, sigma = 0, channel = 0):
S = fudge_factor * S / S.max()
I = frame.astype('uint16')
I[35:195,:,channel] += S.astype('uint16')
I = I.clip(1,255).astype('uint8')
return I
# Control frame display (row 2, col 1)
@app.callback(
Output(component_id='screen-ins', component_property='src'),
[Input(component_id='current-frame', component_property='children'),
Input('snapshot-slider', 'value')]
)
def update_frame_in_slider(frame, snapshot):
print('frame update')
# fetch frame based on snapshot and frame
frame = int(frame/5)
ins = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0/ins'].value
img = ins.copy()
history = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
# Get appropriate saliency values
if frame > len(ins):
img = np.zeros((210,160,3))
actor = img.copy(); critic = img.copy()
else:
img = img[frame]
actor_frames = history['actor_sal'].value
critic_frames = history['critic_sal'].value
actor = actor_frames[frame]; critic=critic_frames[frame]
# Overlay saliency on frame
img = saliency_on_frame_abbr(actor, img, 500, 0, 2)
img = saliency_on_frame_abbr(critic, img, 500, 0 , 0)
# Save as base64 string
buffer = BytesIO()
plt.imsave(buffer, img)
img_str = base64.b64encode(buffer.getvalue()).decode()
return 'data:image/png;base64,{}'.format(img_str)
# Helper func to fetch appropriate layout parameters for color, size, and border for an array of actions (0,1,2,3,2 etc) for 'gantt' plots
def actions_to_marker(actions):
action_colors = {0: "rgb(87, 137, 224)", 1:'rgb(247, 227, 116)', 2:'rgb(59, 229, 73)', 3:'rgb(232, 73, 64)'}
colors = [action_colors[i] for i in actions]
return dict(color = colors, size = 8, line = dict(width = 1))
# Big function to generate 'gantt' plots (row 2, col 2)
# Give a snapshot number and object for selected data in the parallel coords plot below
# also take last arg to correct for zooming scale
def gantt_figures(snapshot, parallelSelectedData, range_bounds=None):
ymid, xmid = 80, 80 # size in pixels of midpoints of saliency frames
sal_thresh = 0.5 # threshold for selecting saliency as percentage of max saliency in frame
action_thresh = 0.7 # threshold for selecting actions with softmax val greater than threshold
# Generate saliency boxes
# This version selects frames where the *regional* saliency is greater than some threshold of the max *regional* saliency for the episode, and values are max-normalized by region
def chart_data(snapshot):
# Get data
history1 = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
rewards1 = history1['reward'].value
outs = history1['outs'].value
critic_sal = history1['critic_sal'].value
lower, upper = 0, len(rewards1)
# Subset data if there are range bounds (due to zooming)
if range_bounds:
lower = int(range_bounds[0])
upper = int(range_bounds[1])
rewards1 = rewards1[lower:upper]
outs = outs[lower:upper]
critic_sal = critic_sal[lower//5:upper//5]
# subset actions that meet threshold
actions1ix = np.where(np.max(outs, axis = 1) > action_thresh)
actions1types = np.argmax(outs[actions1ix], axis=1)
# divide saliency into regions
csaliency1 = critic_sal
csaliency1regions = np.array([csaliency1[:, :ymid, :xmid], csaliency1[:, :ymid, xmid:], csaliency1[:, ymid:, :xmid], csaliency1[:, ymid:, xmid:]])
csaliency1regions_maxs = [region.sum((1,2)).max() for region in csaliency1regions]
# select frames where at least one of the saliency regions meet threshold for that region
csaliency1ix = np.where((csaliency1regions[0].sum((1,2)) > sal_thresh*csaliency1regions_maxs[0]) |
(csaliency1regions[1].sum((1,2)) > sal_thresh*csaliency1regions_maxs[1]) |
(csaliency1regions[2].sum((1,2)) > sal_thresh*csaliency1regions_maxs[2]) |
(csaliency1regions[3].sum((1,2)) > sal_thresh*csaliency1regions_maxs[3]))
# Aggregate and reshape
csaliency1frames = csaliency1regions[:, csaliency1ix]
csaliency1frames = csaliency1frames.sum((3,4))
csaliency1frames = np.squeeze(csaliency1frames)
csaliency1frames = np.swapaxes(csaliency1frames, 0, 1)
# Normalize by region max
for region_ix in range(csaliency1frames.shape[1]):
csaliency1frames[:, region_ix] /= csaliency1regions_maxs[region_ix]
# At this point, csaliency1ix is an array of frame numbers and csaliency1frames is accumulated saliency values by region
# The latter has shape N x 4, where N is the number of frames found that meet threshold and 4 is the number of regions
##### This version selects frames where the *total* saliency is greater than some threshold of the max *total* saliency for the episode
# def chart_data(snapshot):
# history1 = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
# rewards1 = history1['reward'].value
# actions1ix = np.where(np.max(history1['outs'].value, axis = 1) > 0.7)
# actions1types = np.argmax(history1['outs'].value[actions1ix], axis=1)
# csaliency1 = history1['critic_sal'].value
# csaliency1sums = csaliency1.sum((1,2))
# csaliency1max = csaliency1sums.max()
# csaliency1ix = np.where(csaliency1sums > 0.3*csaliency1max)
# csaliency1frames = csaliency1[csaliency1ix]
# csaliency1regions = np.array([[x[:ymid, :xmid].sum(), x[:ymid, xmid:].sum(), x[ymid:, :xmid].sum(), x[ymid:, xmid:].sum()] for x in csaliency1frames])
# csaliency1regions /= csaliency1max
# convert n x n_regions array of saliency values to traces that look like the 2x2 saliency grids along a time line
# example: we pass in n x [1,2,3,4] as values, where 1,2,3,4 are total saliency values of top left, top right, bot left, bot right regions
def plot_region_dots(region_vals, region_ix):
xvals, selPoints = [], []
yvals = np.tile(np.array([1,1,0,0]), len(region_vals))
opacities = ['rgba(200, 68, 68,'+ str(i) + ')' for i in region_vals.flatten()]
width = int(0.024 * len(rewards1)) # scale saliency marks to what user currently sees (to accoutn for zooming)
for i, ix in enumerate(region_ix[0]):
xvals += [ix*5, ix*5 + width, ix*5, ix*5 + width]
# Select points if they were highlighted in frame range specified by dimension 0 of parallel coords plot
# Needed to parse through format of the event callback (was absolutely monstrous)
if parallelSelectedData:
for dim in parallelSelectedData:
if type(dim)==dict:
for k, constraint_objs in dim.items():
if k == 'dimensions[0].constraintrange': # only parse frame data
if type(constraint_objs[0][0]) == list:
constraint_objs = constraint_objs[0]
for c in filter(lambda x: x != None, constraint_objs):
lower = c[0]
upper = c[1]
if ix >= lower//5 and ix <= upper//5:
selPoints += [4*i, 4*i+1, 4*i+2, 4*i+3] # add indices of selected data
# Return data with some styling
return xvals, yvals, dict(color= opacities, size = 14, line = dict(width = 1), symbol = 'square'), selPoints
# Collect data from above function
t1infox, t1infoy, markers, selectedpointsInfo = plot_region_dots(csaliency1frames, csaliency1ix)
# Return all data for 'gantt' plots: rewards, actions, and saliency boxes
return (np.array(t1infox) + lower, t1infoy, markers, selectedpointsInfo), (list(range(lower, upper)), rewards1), (actions1ix[0] + lower, rewards1[actions1ix], actions_to_marker(actions1types))
# Get saliency, rewards, and actions data formatted to put into a chart
trace1sal, trace1rewards, trace1actions = chart_data(snapshot)
# Chart all data
if not trace1sal[3]: # if no selected data
trace1info = go.Scatter(
mode = 'markers',
x = trace1sal[0],
y = trace1sal[1],
marker = trace1sal[2],
cliponaxis= False,
)
else:
trace1info = go.Scatter(
mode = 'markers',
x = trace1sal[0],
y = trace1sal[1],
selectedpoints = trace1sal[3],
marker = trace1sal[2],
cliponaxis= False,
unselected = dict(
marker = dict(color = 'rgba(68, 68, 68,0.1)'),
),
)
# Rewards
trace1 = go.Scatter(
x = trace1rewards[0],
y = trace1rewards[1],
)
# Actions
trace1actions = go.Scatter(
mode = 'markers',
x = trace1actions[0],
y = trace1actions[1],
opacity = 1,
marker = trace1actions[2],
cliponaxis= False
)
# Make subplots skeleton
fig = tools.make_subplots(rows=2, cols=1, vertical_spacing=0.08,
shared_xaxes = True, shared_yaxes = True)
fig.append_trace(trace1info, 1, 1)
fig.append_trace(trace1, 2, 1)
fig.append_trace(trace1actions, 2, 1)
# update layouts to control scale, axis labels
fig['layout'].update(title='Data timeline chart of epoch ep '+str(snapshot),
showlegend=False,
clickmode = 'event+select',
margin = dict(
l = 50,
r = 40,
b = 35,
t = 25,
pad = 4
))
fig['layout']['yaxis1'].update(tickmode='linear',
ticks='outside',
tick0=0,
dtick=2,
showticklabels=False,
range = [0,5],
fixedrange = True)
fig['layout']['yaxis2'].update(title='', range=[0, 7], autorange=False)
return fig
# Control 'gantt' charts (row 2, col 2)
@app.callback(
[Output(component_id='gantt', component_property='figure'),
Output(component_id='gantt2', component_property='figure'),
],
[Input(component_id='gantt-select1', component_property='value'),
Input(component_id='gantt-select2', component_property='value'),
Input(component_id='parallel-sal', component_property='restyleData'),
Input(component_id='gantt', component_property='relayoutData'),
Input(component_id='gantt2', component_property='relayoutData'), # listen for parallel coords linking
]
)
def update_gantts(snapshot1, snapshot2, parallelSelectedData, relayout1, relayout2):
if not snapshot1: # default values on page load
snapshot1 = 90
if not snapshot2:
snapshot2 = 60
range_bounds, range_bounds2 = None, None
# If zoomed, receive a callback to properly scale the 2x2 saliency marks
# fetch x-axis range bounds from each callback
if relayout1 and 'xaxis.range[0]' in relayout1:
range_bounds = [relayout1['xaxis.range[0]'], relayout1['xaxis.range[1]']]
if relayout2 and 'xaxis.range[0]' in relayout2:
range_bounds2 = [relayout2['xaxis.range[0]'], relayout2['xaxis.range[1]']]
fig1 = gantt_figures(snapshot1, parallelSelectedData, range_bounds)
fig2 = gantt_figures(snapshot2, parallelSelectedData, range_bounds2)
return fig1, fig2
# Control parallel coords plot (row 2, col 2)
@app.callback(
Output(component_id='parallel-sal', component_property='figure'),
[Input(component_id='gantt-select1', component_property='value'),
Input(component_id='gantt-select2', component_property='value')],
)
def update_parallel_sal(snapshot1, snapshot2):
ymid, xmid = 80, 80 # pixel lengths to middle of frame
sal_thresh = 0.5 # threshold for selecting saliency as percentage of max saliency in frame
# This version selects frames where the *regional* saliency is greater than some threshold of the max *regional* saliency for the episode, and values are max-normalized by region
# Same as chart_data in gantt functions
def chart_data(snapshot):
history1 = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
rewards1 = history1['reward'].value
csaliency1 = history1['critic_sal'].value
csaliency1regions = np.array([csaliency1[:, :ymid, :xmid], csaliency1[:, :ymid, xmid:], csaliency1[:, ymid:, :xmid], csaliency1[:, ymid:, xmid:]])
csaliency1regions_maxs = [region.sum((1,2)).max() for region in csaliency1regions]
csaliency1ix = np.where((csaliency1regions[0].sum((1,2)) > sal_thresh*csaliency1regions_maxs[0]) |
(csaliency1regions[1].sum((1,2)) > sal_thresh*csaliency1regions_maxs[1]) |
(csaliency1regions[2].sum((1,2)) > sal_thresh*csaliency1regions_maxs[2]) |
(csaliency1regions[3].sum((1,2)) > sal_thresh*csaliency1regions_maxs[3]))
csaliency1frames = csaliency1regions[:, csaliency1ix]
csaliency1frames = csaliency1frames.sum((3,4))
csaliency1frames = np.squeeze(csaliency1frames)
csaliency1frames = np.swapaxes(csaliency1frames, 0, 1)
for region_ix in range(csaliency1frames.shape[1]):
csaliency1frames[:, region_ix] /= csaliency1regions_maxs[region_ix]
return csaliency1ix[0]*5, csaliency1frames
# This version selects frames where the *total* saliency is greater than some threshold of the max *total* saliency for the episode
# def chart_data(snapshot):
# history1 = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
# rewards1 = history1['reward'].value
# csaliency1 = history1['critic_sal'].value
# csaliency1sums = csaliency1.sum((1,2))
# csaliency1max = csaliency1sums.max()
# csaliency1ix = np.where(csaliency1sums > 0.4*csaliency1max)
# csaliency1frames = csaliency1[csaliency1ix]
# csaliency1regions = np.array([[x[:ymid, :xmid].sum(), x[:ymid, xmid:].sum(), x[ymid:, :xmid].sum(), x[ymid:, xmid:].sum()] for x in csaliency1frames])
# csaliency1regions /= csaliency1max
# return csaliency1ix[0]*5, csaliency1regions
# Get data
ep1ix, ep1vals = chart_data(snapshot1)
ep2ix, ep2vals = chart_data(snapshot2)
max_range = max(ep1ix.max(), ep2ix.max())
# Reshape data
ep1data = np.hstack((ep1ix.reshape(-1, 1), ep1vals, np.repeat([int(snapshot1)], ep1ix.shape[0]).reshape(-1,1)))
ep2data = np.hstack((ep2ix.reshape(-1, 1), ep2vals, np.repeat([int(snapshot2)], ep2ix.shape[0]).reshape(-1,1)))
# Put all data into 1 matrix
all_data = np.vstack((ep1data, ep2data))
# Chart data - frame is first dimension and regions follow on next 4 dimensions. Last column in all_data is the epoch number, which determines the color
trace = go.Parcoords(
line = dict(color = all_data[:, -1],
showscale = True),
dimensions = list([
dict(range = [0, max_range],
label = 'Frame', values = all_data[:, 0]),
dict(range = [0,1],
label = 'TopLeft C-Sal', values = all_data[:,1]),
dict(range = [0,1],
label = 'TopRight C-Sal', values = all_data[:,2]),
dict(range = [0,1],
label = 'BotLeft C-Sal', values = all_data[:,3]),
dict(range = [0,1],
label = 'BotRight C-Sal', values = all_data[:,4]),
])
)
data = [ trace]
layout = go.Layout( title = "Frames with by-region saliency > "+str(sal_thresh) + " of max",
margin = dict(
l = 55,
r = 50,
b = 35,
t = 80,
pad = 4
),
)
figure = go.Figure(data = data, layout = layout)
return figure
# Control region subplots (row 2, col 3)
@app.callback(
Output(component_id='regions-subplots', component_property='figure'),
[Input(component_id='snapshot-slider', component_property='value')]
)
def update_regions_plots(snapshot):
ymid, xmid = 80, 80
window_length = 10
# Get data
history = replays['models_model7-02-17-20-41/model.'+str(snapshot)+'.tar/history/0']
actor_frames = history['actor_sal'].value
critic_frames = history['critic_sal'].value
# Sum every frame
actor_tot = actor_frames.sum((1,2))
critic_tot = critic_frames.sum((1,2))
# Divide frames into regions
targets = [(actor_frames[:, :ymid, :xmid], critic_frames[:, :ymid, :xmid]),
(actor_frames[:, :ymid, xmid:], critic_frames[:, :ymid, xmid:]),
(actor_frames[:, ymid:, :xmid], critic_frames[:, ymid:, :xmid]),
(actor_frames[:, ymid:, xmid:], critic_frames[:, ymid:, xmid:])]
# intensity defined by sum of values in frame region divided by sum of total values of full frame
trace_labels = ['TopLeft', 'TopRight', 'BotLeft', 'BotRight']
a_traces = []
c_traces = []
a_ubounds, a_lbounds, c_ubounds, c_lbounds = [],[],[],[]
for i in range(4):
# Change indexing since saliency is only recorded every 5 frames
xrange = list(range(0, actor_frames.shape[0] * 5, 5))
# Get moving averages and moving mins/maxs of window_length (default 10)
data = pd.Series((targets[i][0]).sum((1,2))/actor_tot).rolling(window=window_length)
mavg = data.mean()
lowerbound = data.min()
upperbound = data.max()
ubound = dict(
x = xrange,
y = upperbound,
hoverinfo = 'x+y',
mode='lines',
marker=dict(color="#444"),
line=dict(width=0),
fillcolor='rgba(68, 68, 200, 0.3)',
fill='tonexty'
)
trace = dict(
x = xrange,
y = mavg,
hoverinfo = 'x+y',
mode='lines',
fillcolor='rgba(68, 68, 200, 0.3)',
fill='tonexty',
line = dict(
color = ('rgb(24, 100, 205)'),
width = 2)
)
lbound = dict(
x = xrange,
y = lowerbound,
marker=dict(color="#444"),
line=dict(width=0),
mode='lines')
a_ubounds.append(ubound)
a_lbounds.append(lbound)
a_traces.append(trace)
# Do same as above but fro critic saliency
for i in range(4):
data = pd.Series((targets[i][1]).sum((1,2))/critic_tot).rolling(window=window_length)
mavg = data.mean()