Issue with LR find / fit_one_cycle

[chromade]

Hello,

Not sure if this an error or - more likely - a mistake by me as I am quite new to fastai/tsai

Following tutorial 1 using frequency data, lr_find yields a straight line and fit_one_cycle does not work either as shown below.
I also have the assertion error highlighted by geoHell in the Issues section.

Thanks!

Full code here:

import tsai
from tsai.all import *
print('tsai       :', tsai.__version__)
print('fastai     :', fastai.__version__)
print('fastcore   :', fastcore.__version__)
print('torch      :', torch.__version__)

tsai       : 0.2.15
fastai     : 2.2.7
fastcore   : 1.3.19
torch      : 1.7.1

import pandas as pd
import numpy 
import sys, os 

all_data = pd.read_csv('/home/msms-dl/DbgDir01/data_set_for_time_series_modelling.csv')

	direct_num	sample_num	y_list	0	1	2	3	4	5	6	...	90	91	92	93	94	95	96	97	98	99
0	0	0	0.224494	0.364835	0.520462	0.783266	0.645734	0.512022	0.632783	...	0.168986	0.156310	0.023165	0.214470	0.265909	0.416136	0.750068	0.491910	0.137570	0.094487
0	0	0	0.206774	0.352226	0.482965	0.759031	0.602404	0.457867	0.607926	...	0.161466	0.157124	0.040742	0.208043	0.285894	0.458393	0.774518	0.486450	0.115466	0.111570
0	0	0	0.186787	0.337114	0.440572	0.730284	0.555950	0.402950	0.579868	...	0.151666	0.157832	0.059492	0.199419	0.303068	0.501200	0.794325	0.479213	0.091583	0.129218
0	0	0	0.165137	0.319422	0.394515	0.697630	0.508092	0.349335	0.548963	...	0.139652	0.158393	0.077990	0.188847	0.317104	0.543089	0.809362	0.470215	0.066559	0.146691
0	0	0	0.142494	0.299314	0.345786	0.662003	0.460838	0.298788	0.515857	...	0.125158	0.159286	0.095373	0.176511	0.327910	0.582447	0.819847	0.459235	0.041058	0.163729
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
0	167	0	0.322240	0.539831	0.505063	0.167877	0.633645	0.554848	0.353178	...	0.373876	0.270555	0.167381	0.181751	0.142647	0.125436	0.207500	0.308182	0.373462	0.184427
0	167	0	0.315026	0.525238	0.516878	0.162293	0.629577	0.510102	0.360265	...	0.348680	0.249225	0.190153	0.168258	0.133828	0.112630	0.229045	0.324270	0.380510	0.205917
0	167	0	0.301370	0.509864	0.529479	0.179382	0.621629	0.461042	0.366502	...	0.324255	0.228143	0.210581	0.153149	0.126947	0.102250	0.248849	0.337763	0.383555	0.228272
0	167	0	0.281651	0.493954	0.542676	0.216512	0.609205	0.409628	0.370260	...	0.301928	0.208219	0.228157	0.136695	0.122028	0.096530	0.266542	0.348133	0.382708	0.250095
0	167	0	0.255783	0.477968	0.555399	0.267130	0.592272	0.357991	0.369419	...	0.283566	0.190531	0.242686	0.119003	0.118727	0.097081	0.281786	0.355272	0.377765	0.270722

X, y = df2xy(all_data, feat_col='sample_num', target_col='y_list', data_cols=all_data.columns[3:103])

X.shape, y.shape

((600, 168, 100), (600,))

tfms  = [None, [Categorize()]]
dsets = TSDatasets(X, y, tfms=tfms, splits=splits)
dsets


(#600) [(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0)),(TSTensor(vars:168, len:100), TensorCategory(0))...]

model = InceptionTime(dls.vars, dls.c)
learn = Learner(dls, model, metrics=accuracy)

learn.lr_find(suggestions=False)

learn.fit_one_cycle(25, lr_max=1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.000000	0.000000	1.000000	00:00
1	0.000000	0.000000	1.000000	00:00
2	0.000000	0.000000	1.000000	00:00
3	0.000000	0.000000	1.000000	00:00
4	0.000000	0.000000	1.000000	00:00
5	0.000000	0.000000	1.000000	00:00
6	0.000000	0.000000	1.000000	00:00
7	0.000000	0.000000	1.000000	00:00
8	0.000000	0.000000	1.000000	00:00
9	0.000000	0.000000	1.000000	00:00
10	0.000000	0.000000	1.000000	00:00
11	0.000000	0.000000	1.000000	00:00
12	0.000000	0.000000	1.000000	00:00
13	0.000000	0.000000	1.000000	00:00
14	0.000000	0.000000	1.000000	00:00
15	0.000000	0.000000	1.000000	00:00
16	0.000000	0.000000	1.000000	00:00
17	0.000000	0.000000	1.000000	00:00
18	0.000000	0.000000	1.000000	00:00
19	0.000000	0.000000	1.000000	00:00
20	0.000000	0.000000	1.000000	00:00
21	0.000000	0.000000	1.000000	00:00
22	0.000000	0.000000	1.000000	00:00
23	0.000000	0.000000	1.000000	00:00
24	0.000000	0.000000	1.000000	00:00

[oguiza]

Hi @chromade,
Could you please indicate how you create the splits? I can’t see that in the code.

[chromade]

Thank you for the response!

oops, I define splits between df2xy() and tfrms = as:

splits = get_splits(y, valid_size=.2, stratify=True, random_state=23, shuffle=True) splits

((#480) [566,413,338,482,422,484,238,388,318,493...],
 (#120) [142,517,533,513,12,42,249,306,247,360...])

[oguiza]

Thanks @chromade.
It strage that the train loss and valid loss are both 0, and accuracy is 100%.
It looks to me that the issue is in the label. All labels seem to be set to 0, so there's nothing to learn.
The y_list data is of type float. this means this is a regression problem, instead of a classification one. For regression problems you need to use this:

ts = [None, TSRegression()]

instead of

ts = [None, Categorize()] # or [None, TSClassification()] they are both the same.

Please, try it and let me know if it works.

[chromade]

Hi again,

So the table above is copied wrong so I have uploaded the notebook to github. Here you can see the 3 categories im trying to classify (with 100 variables over 150 timepoints).
I followed your very clear tutorial again but the dls still only recognises 1 label when there are 3 (which Ive highlighted in the above attempt).

Thanks again for your input.

[oguiza]

Could you please run this:

check_data(X, y, splits)

and share the results?
I still think that for some reason all your y's are set to 0.

[chromade]

Hi again,

You're right -- all the y's are set to zero and check_data(X, y, splits) has the error:

X      - shape: [672 samples x 100 features x 150 timesteps]  type: ndarray  dtype:float64  isnan: 0

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-36-5c89bd81473c> in <module>
----> 1 test = check_data(X, y, splits)
      2 test

~/miniconda3/envs/timeseriesAI/lib/python3.8/site-packages/tsai/data/external.py in check_data(X, y, splits, show_plot)
    217         if isinstance(y[0], (Integral, str)):
    218             n_classes = f'{len(np.unique(y))} ({len(y)//len(np.unique(y))} samples per class) {L(np.unique(y).tolist())}'
--> 219             y_is_nan = 'nan' in [c.lower() for c in np.unique(y)]
    220             print(f'y      - shape: {y.shape}  type: {cls_name(y)}  dtype:{y.dtype}  n_classes: {n_classes}  isnan: {y_is_nan}')
    221         else:

~/miniconda3/envs/timeseriesAI/lib/python3.8/site-packages/tsai/data/external.py in <listcomp>(.0)
    217         if isinstance(y[0], (Integral, str)):
    218             n_classes = f'{len(np.unique(y))} ({len(y)//len(np.unique(y))} samples per class) {L(np.unique(y).tolist())}'
--> 219             y_is_nan = 'nan' in [c.lower() for c in np.unique(y)]
    220             print(f'y      - shape: {y.shape}  type: {cls_name(y)}  dtype:{y.dtype}  n_classes: {n_classes}  isnan: {y_is_nan}')
    221         else:

AttributeError: 'numpy.int64' object has no attribute 'lower'

I also tried converting the y_list in dataframe from np.int64 into int and then reconstructing the dataframe. But the y coming out of get_splits is still all 0's.

Thanks

[oguiza]

Hi @chromade, I've released an updated df2xy in the GitHub repo. It'd be good if you try it again.