tokenize

Fill in a module description here

The autoreload extension is already loaded. To reload it, use:
  %reload_ext autoreload

EXTERNAL_PARENT_DIR = Path('/Users/al/Dropbox/2-Areas/2-Sketchbooks')
EXTERNAL_SKETCHBOOKS_DIR = EXTERNAL_PARENT_DIR / 'sketchbooks'
EXTERNAL_DATA_HOME = EXTERNAL_PARENT_DIR / 'singleline_data'

EPOCH = SketchbookEpoch(epoch='20240104', data_home=EXTERNAL_DATA_HOME)

no env var SINGLELINE_DATA_HOME, defaulting to: None

THRESHOLDED_OUTPUT_DIR = EPOCH.dir_08_THRESHOLDED()
THESHOLDED_TSV_FNAME = EPOCH.tsv_08_THRESHOLDED()
fjdf = pd.read_csv(THESHOLDED_TSV_FNAME, delimiter="\t", index_col="idx")
fjdf.head()

	new_idx	group_fname	group_id	group_num_strokes	group_num_points	indiv_fnamebboxsep_	orig_fname	labelbboxsep_	handlabeled_path	is_new	...	err_msg	indiv_fname	label	emb_csv	knn_dist	knn_clusterid	knn_label	cluster_label	rdp_num_points	rdp_epsilon
idx
0	0	0_drawings/sb55p003g00.png	0	4	1510	sb55p003.jpg	art/sb55p003.jpg	art	0_drawings/sb55p003.jpg	False	...	NaN	sb55p003g00.png	0_drawings	0.04728329,-0.5714155,-0.73929733,-0.35399058,0.22400337,-0.65341413,0.11710799,-0.4896673,-0.7537334,-0.27084306,-0.43680793,-0.72562873,0.26104796,-0.6266761,-0.68991846,-0.57515705,-0.5467274,-0.6520286,2.2132683,-0.63401383,-0.64108604,0.0068377256,-0.5446661,-0.6779002,-0.40022308,-0.6690378,-0.25018138,-0.5982413,-0.73986673,1.0058696,-0.68509877,-0.67007643,0.5276551,-0.4592305,-0.600525,-0.3091834,-0.73933,-0.57338566,0.84570634,-0.6179389,-0.52985966,0.5803701,0.10534239,-0.6682474,-0.1908136,-0.59257406,-0.6157326,-0.6001429,1.3485551,1.1220317,0.10932493,-0.66552746,0.8113811,1....	196.44476	14	NaN	good	1510	0.0
1	1	0_drawings/sb73p061g00.png	0	2	128	sb73p061.jpg	art/sb73p061.jpg	art	0_drawings/sb73p061.jpg	False	...	NaN	sb73p061g00.png	0_drawings	-0.6198182,-0.5714155,-0.73929733,-0.6553513,-0.6254566,-0.65341413,-0.7267234,-0.46724933,-0.7537334,0.9455167,-0.62038505,-0.72562873,0.18925023,-0.6266761,0.33996195,-0.15316445,-0.5467274,0.03829068,1.0683823,0.23568732,-0.64108604,-0.55193967,-0.5446661,-0.022136033,1.4963098,-0.4510746,-0.4518578,-0.5982413,-0.28755826,0.1980134,-0.68509877,-0.67007643,-0.52171326,-0.19965711,-0.0069928765,-0.46161294,-0.73933,-0.50798476,0.03861904,0.68692446,-0.5830822,0.46990138,1.1344135,-0.6682474,0.6321523,-0.59257406,0.156174,-0.6001429,2.36085,0.16527855,-0.13570255,-0.63657653,0.61075664,0.7...	119.71115	10	NaN	good	128	0.0
5	5	0_drawings/sb73p041g00.png	0	3	702	sb73p041.jpg	art/sb73p041.jpg	art	0_drawings/sb73p041.jpg	False	...	NaN	sb73p041g00.png	0_drawings	-0.6198182,-0.5714155,-0.73929733,-0.6553513,-0.4395254,-0.65341413,-0.7267234,-0.30100727,-0.7537334,0.6979399,0.5823597,-0.72562873,-0.662768,0.702158,-0.6232817,-0.57515705,-0.22156492,-0.07576728,-0.25915077,0.7949797,-0.5030652,-0.55193967,-0.5446661,-0.6779002,-0.57759243,-0.6690378,-0.4518578,0.8875855,-0.73986673,0.2878165,-0.68509877,-0.67007643,-0.52171326,-0.20293939,-0.600525,0.28690326,-0.73933,0.13321829,-0.65925777,-0.6179389,-0.61049384,-0.58021957,-0.6189488,-0.6682474,-0.75794554,-0.59257406,0.2206502,-0.6001429,-0.3428496,-0.37840372,-0.59497476,-0.66552746,-0.6482407,0....	285.93152	8	NaN	good	702	0.0
6	6	0_drawings/sb73p041g01.png	1	1	205	sb73p041.jpg	art/sb73p041.jpg	art	0_drawings/sb73p041.jpg	False	...	NaN	sb73p041g01.png	0_drawings	-0.6198182,-0.5714155,-0.73929733,-0.6553513,0.058876395,-0.65341413,-0.7267234,-0.42173523,-0.7537334,0.7433976,0.19966531,-0.72562873,0.62771916,-0.6266761,-0.68991846,-0.57515705,-0.5467274,1.5417054,2.402433,-0.6497359,-0.64108604,-0.4386818,-0.5446661,-0.6779002,-0.57759243,-0.6690378,-0.4518578,0.11702931,-0.73986673,-0.45211998,-0.68509877,-0.67007643,-0.52171326,0.34430972,-0.48163074,-0.6186755,-0.73933,0.040438354,-0.65925777,0.7474549,-0.61049384,0.30740798,1.8930382,-0.6682474,-0.75794554,-0.59257406,-0.3586682,0.89375025,-0.651804,0.08712256,1.415978,-0.66552746,-0.6482407,-0....	179.57721	8	NaN	good	205	0.0
10	10	0_drawings/sb55p080g00.png	0	4	1462	sb55p080.jpg	art/sb55p080.jpg	art	0_drawings/sb55p080.jpg	False	...	NaN	sb55p080g00.png	0_drawings	-0.6198182,-0.5714155,-0.12864059,-0.6553513,0.03367281,-0.3521561,-0.7267234,-0.47485968,-0.7537334,-0.42715576,1.136474,-0.72562873,2.6017675,-0.6266761,-0.68991846,-0.57515705,-0.5467274,1.0752423,-0.02930379,1.118658,0.39956325,-0.55193967,-0.5446661,-0.6779002,-0.57759243,-0.6690378,-0.4518578,0.16788304,-0.42450532,-0.45211998,-0.68509877,-0.67007643,-0.23189741,0.37339267,1.0403941,0.11935049,-0.73933,-0.19533288,1.3258636,-0.53331745,-0.61049384,1.4947255,0.25517285,-0.59559596,-0.75794554,-0.59257406,0.2976374,-0.6001429,-0.651804,0.09863156,1.5771475,-0.66552746,-0.15108436,0.443...	143.85107	5	NaN	good	1462	0.0

5 rows × 27 columns

THRESHOLDED_NPZ_ALL = THRESHOLDED_OUTPUT_DIR / 'furtherfiltered_all.npz'
ds = np.load(THRESHOLDED_NPZ_ALL, encoding='latin1', allow_pickle=True)['arr_0']
print(f"num drawings in dataset: {ds.shape[0]}")

num drawings in dataset: 1550

# stack all drawings, to get one big array of all the stroke3 steps in the dataset
all_steps = np.vstack(ds).astype(np.float32)
print(f"num individual steps in all strokes in dataset: {all_steps.shape[0]}")

num individual steps in all strokes in dataset: 315396

zero_indices = []
one_indices = []
for idx, is_zero in enumerate(all_steps[:,2]==0):
    if is_zero:
        zero_indices.append(idx)
    else:
        one_indices.append(idx)

zero_steps = all_steps[zero_indices]
one_steps = all_steps[one_indices]

print(f"num steps with lift_pen=0: {zero_steps.shape[0]}")
print(f"num steps with lift_pen=1: {one_steps.shape[0]}")

num steps with lift_pen=0: 312557
num steps with lift_pen=1: 2839

kmeans_zero = train_kmeans(zero_steps, ncentroids=960, nredo=10, seed=42)
kmeans_one = train_kmeans(one_steps, ncentroids=40, nredo=10, seed=42)

print(kmeans_zero.centroids.shape, kmeans_one.centroids.shape)

#kmeans_zero.centroids

# kmeans_one.centroids

source

Stroke3Tokenizer

 Stroke3Tokenizer (centroids_zero, centroids_one)

Initialize self. See help(type(self)) for accurate signature.

# t = Stroke3Tokenizer(kmeans_zero.centroids, kmeans_one.centroids)
# t.save(THRESHOLDED_OUTPUT_DIR)

tokenizer = Stroke3Tokenizer.load(THRESHOLDED_OUTPUT_DIR)

(2048, 3)

test_toks = tokenizer.encode(ds[2])

test_toks[0:5]

[1635, 1251, 1393, 1013, 671]

test_decoded = tokenizer.decode(test_toks)
test_decoded[:5]

array([[ 0.3155888, 37.36093  ,  0.       ],
       [ 7.065908 , -8.754673 ,  0.       ],
       [-1.8780638,  6.4204984,  0.       ],
       [ 4.526787 ,  4.7838955,  0.       ],
       [ 4.9135633, 11.580178 ,  0.       ]], dtype=float32)

plot_strokes(deltas_to_strokes(test_decoded))

tokenizer.centroids_zero

array([[ 0.6209417 , -0.33721057,  0.        ],
       [-4.1855707 , -3.1003277 ,  0.        ],
       [ 5.7066207 ,  2.6960166 ,  0.        ],
       ...,
       [-1.1269675 , -0.28028095,  0.        ],
       [-3.6861207 , -4.5447764 ,  0.        ],
       [-1.6025403 ,  0.5354872 ,  0.        ]], dtype=float32)

plot_strokes(deltas_to_strokes(ds[2]))

deltas_tmp = ds[2]

deltas_tmp.shape

(242, 3)

deltas_tmp[:5]

array([[ 0.3771549 , 38.41915779,  0.        ],
       [ 6.84299279, -8.26956107,  0.        ],
       [-1.94953761,  6.4101698 ,  0.        ],
       [ 4.55352531,  4.95268558,  0.        ],
       [ 4.70517464, 11.38796166,  0.        ]])

D0, I0 = kmeans_zero.index.search(deltas_tmp, 1)
D1, I1 = kmeans_one.index.search(deltas_tmp, 1)

S0 = kmeans_zero.centroids[I0]
S1 = kmeans_one.centroids[I1]

plot_strokes(deltas_to_strokes(deltas_tmp))

tokenized_steps = []
for idx in range(deltas_tmp.shape[0]):
    row = deltas_tmp[idx]
    if row[2] == 0:
        tokenized_steps.append(S0[idx][0])
    elif row[2] == 1:
        tokenized_steps.append(S1[idx][0])
    else:
        raise Exception ('didnt find a 0 or 1 in the lift_pen column')

new_deltas = np.array(tokenized_steps)

new_deltas[0:5]

array([[-3.3815517, 40.04465  ,  0.       ],
       [ 6.9500613, -8.3709345,  0.       ],
       [-1.9793537,  6.5232167,  0.       ],
       [ 4.097289 ,  4.915239 ,  0.       ],
       [ 4.9076104, 11.547217 ,  0.       ]], dtype=float32)

len(new_deltas)

len(deltas_tmp)