Model utilities

Utility functions used to build PyTorch timeseries models.

---

source

apply_idxs

 apply_idxs (o, idxs)

Function to apply indices to zarr, dask and numpy arrays

---

source

SeqTokenizer

 SeqTokenizer (c_in, embed_dim, token_size=60, norm=False)

Generates non-overlapping tokens from sub-sequences within a sequence by applying a sliding window

---

source

get_embed_size

 get_embed_size (n_cat, rule='log2')

test_eq(get_embed_size(35), 6)

---

source

has_weight_or_bias

 has_weight_or_bias (l)

---

source

has_weight

 has_weight (l)

---

source

has_bias

 has_bias (l)

---

source

is_conv

 is_conv (l)

---

source

is_affine_layer

 is_affine_layer (l)

---

source

is_conv_linear

 is_conv_linear (l)

---

source

is_bn

 is_bn (l)

---

source

is_linear

 is_linear (l)

---

source

is_layer

 is_layer (*args)

---

source

get_layers

 get_layers (model, cond=<function noop>, full=True)

---

source

check_weight

 check_weight (m, cond=<function noop>, verbose=False)

---

source

check_bias

 check_bias (m, cond=<function noop>, verbose=False)

---

source

get_nf

 get_nf (m)

Get nf from model’s first linear layer in head

---

source

ts_splitter

 ts_splitter (m)

Split of a model between body and head

---

source

transfer_weights

 transfer_weights (model, weights_path:pathlib.Path,
                   device:torch.device=None, exclude_head:bool=True)

Utility function that allows to easily transfer weights between models. Taken from the great self-supervised repository created by Kerem Turgutlu. https://github.com/KeremTurgutlu/self_supervised/blob/d87ebd9b4961c7da0efd6073c42782bbc61aaa2e/self_supervised/utils.py

---

source

build_ts_model

 build_ts_model (arch, c_in=None, c_out=None, seq_len=None, d=None,
                 dls=None, device=None, verbose=False, s_cat_idxs=None,
                 s_cat_embeddings=None, s_cat_embedding_dims=None,
                 s_cont_idxs=None, o_cat_idxs=None, o_cat_embeddings=None,
                 o_cat_embedding_dims=None, o_cont_idxs=None,
                 patch_len=None, patch_stride=None, fusion_layers=128,
                 fusion_act='relu', fusion_dropout=0.0,
                 fusion_use_bn=True, pretrained=False, weights_path=None,
                 exclude_head=True, cut=-1, init=None, arch_config={},
                 **kwargs)

from tsai.data.core import get_ts_dls, TSClassification
from tsai.models.TSiTPlus import TSiTPlus
from fastai.losses import CrossEntropyLossFlat

X = np.random.rand(16, 3, 128)
y = np.random.randint(0, 2, (16, 3))
tfms = [None, [TSClassification()]]
dls = get_ts_dls(X, y, splits=RandomSplitter()(range_of(X)), tfms=tfms)
model = build_ts_model(TSiTPlus, dls=dls, pretrained=False, verbose=True)
xb, yb = dls.one_batch()
output = model(xb)
print(output.shape)
loss = CrossEntropyLossFlat()(output, yb)
print(loss)
assert output.shape == (dls.bs, dls.d, dls.c)

arch: TSiTPlus(c_in=3 c_out=2 seq_len=128 arch_config={} kwargs={'custom_head': functools.partial(<class 'tsai.models.layers.lin_nd_head'>, d=3)})
torch.Size([13, 3, 2])
TensorBase(0.8796, grad_fn=<AliasBackward0>)

---

source

count_parameters

 count_parameters (model, trainable=True)

---

source

build_tsimage_model

 build_tsimage_model (arch, c_in=None, c_out=None, dls=None,
                      pretrained=False, device=None, verbose=False,
                      init=None, arch_config={}, **kwargs)

---

source

build_tabular_model

 build_tabular_model (arch, dls, layers=None, emb_szs=None, n_out=None,
                      y_range=None, device=None, arch_config={}, **kwargs)

from tsai.data.external import get_UCR_data
from tsai.data.core import TSCategorize, get_ts_dls
from tsai.data.preprocessing import TSStandardize
from tsai.models.InceptionTime import *

X, y, splits = get_UCR_data('NATOPS', split_data=False)
tfms = [None, TSCategorize()]
batch_tfms = TSStandardize()
dls = get_ts_dls(X, y, splits, tfms=tfms, batch_tfms=batch_tfms)
model = build_ts_model(InceptionTime, dls=dls)
test_eq(count_parameters(model), 460038)

---

source

get_clones

 get_clones (module, N)

m = nn.Conv1d(3,4,3)
get_clones(m, 3)

ModuleList(
  (0-2): 3 x Conv1d(3, 4, kernel_size=(3,), stride=(1,))
)

---

source

split_model

 split_model (m)

---

source

output_size_calculator

 output_size_calculator (mod, c_in, seq_len=None)

c_in = 3
seq_len = 30
m = nn.Conv1d(3, 12, kernel_size=3, stride=2)
new_c_in, new_seq_len = output_size_calculator(m, c_in, seq_len)
test_eq((new_c_in, new_seq_len), (12, 14))

---

source

change_model_head

 change_model_head (model, custom_head, **kwargs)

Replaces a model’s head by a custom head as long as the model has a head, head_nf, c_out and seq_len attributes

---

source

true_forecaster

 true_forecaster (o, split, horizon=1)

---

source

naive_forecaster

 naive_forecaster (o, split, horizon=1)

a = np.random.rand(20).cumsum()
split = np.arange(10, 20)
a, naive_forecaster(a, split, 1), true_forecaster(a, split, 1)

(array([0.99029138, 1.68463991, 2.21744589, 2.65448222, 2.85159354,
        3.26171729, 3.67986707, 4.04343956, 4.3077458 , 4.44585435,
        4.76876866, 4.85844441, 4.93256093, 5.52415845, 6.10704489,
        6.74848957, 7.31920741, 8.20198208, 8.78954283, 9.0402    ]),
 array([4.44585435, 4.76876866, 4.85844441, 4.93256093, 5.52415845,
        6.10704489, 6.74848957, 7.31920741, 8.20198208, 8.78954283]),
 array([4.76876866, 4.85844441, 4.93256093, 5.52415845, 6.10704489,
        6.74848957, 7.31920741, 8.20198208, 8.78954283, 9.0402    ]))