File size: 5,257 Bytes
9079f0c | 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 | import torch
from torch import nn
from transformers import PreTrainedModel
from transformers.modeling_outputs import SequenceClassifierOutput
from .configuration_yield import YieldConfig
from .yield_transformer import YieldTransformer
class YieldForSequenceClassification(PreTrainedModel):
config_class = YieldConfig
base_model_prefix = "yield_model"
def __init__(self, config: YieldConfig):
super().__init__(config)
self.yield_model = YieldTransformer(
w_dim=config.W,
soil_dim=config.S,
d_model=config.d_model,
nhead=config.nhead,
num_layers=config.num_layers,
dim_ff=config.dim_ff,
dropout=config.dropout,
use_crop=config.use_crop,
crop_emb_dim=config.crop_emb_dim,
max_weeks=max(52, config.K),
pool=config.pool,
)
self.post_init()
def forward(
self,
weather,
soil,
crop_id,
horizon_idx=None,
labels=None,
**kwargs,
):
# ==================================================
# Shape checks
# ==================================================
if weather.ndim != 3:
raise ValueError(
f"weather must have shape [B,K,W], "
f"received {tuple(weather.shape)}"
)
if soil.ndim != 2:
raise ValueError(
f"soil must have shape [B,S], "
f"received {tuple(soil.shape)}"
)
# ==================================================
# Training normalization statistics
# ==================================================
w_mean = torch.tensor(
self.config.w_mean,
device=weather.device,
dtype=weather.dtype,
)
w_std = torch.tensor(
self.config.w_std,
device=weather.device,
dtype=weather.dtype,
)
s_mean = torch.tensor(
self.config.s_mean,
device=soil.device,
dtype=soil.dtype,
)
s_std = torch.tensor(
self.config.s_std,
device=soil.device,
dtype=soil.dtype,
)
# NaN -> training mean
weather = torch.where(
torch.isnan(weather),
w_mean.view(1, 1, -1),
weather,
)
soil = torch.where(
torch.isnan(soil),
s_mean.view(1, -1),
soil,
)
# normalize
weather = (
weather - w_mean.view(1, 1, -1)
) / w_std.view(1, 1, -1)
soil = (
soil - s_mean.view(1, -1)
) / s_std.view(1, -1)
# ==================================================
# Cutoff
# ==================================================
if horizon_idx is None:
horizon_idx = torch.full(
(weather.shape[0],),
weather.shape[1],
dtype=torch.long,
device=weather.device,
)
if not torch.is_tensor(horizon_idx):
horizon_idx = torch.tensor(
horizon_idx,
dtype=torch.long,
device=weather.device,
)
horizon_idx = horizon_idx.to(
weather.device
).long()
if horizon_idx.ndim == 0:
horizon_idx = horizon_idx.unsqueeze(0)
# FlexServe requests are normally one sample.
# Ensure one common temporal length for a batch.
unique_cutoffs = torch.unique(
horizon_idx
)
if len(unique_cutoffs) != 1:
raise ValueError(
"All samples in one batch must use the same cutoff."
)
t_eff = int(
unique_cutoffs[0].item()
)
weather = weather[
:,
:t_eff,
:
]
# ==================================================
# Existing trained model
# ==================================================
logits_norm = self.yield_model(
weather,
soil,
crop_id,
horizon_idx=horizon_idx,
causal=True,
return_sequence=False,
)
# Restore yield to bu/acre.
y_mean = torch.tensor(
self.config.y_mean,
device=logits_norm.device,
dtype=logits_norm.dtype,
)
y_std = torch.tensor(
self.config.y_std,
device=logits_norm.device,
dtype=logits_norm.dtype,
)
predicted_yield = (
logits_norm * y_std
+ y_mean
)
# TextClassificationPipeline expects [B, num_labels].
logits = predicted_yield.unsqueeze(-1)
loss = None
if labels is not None:
labels = labels.to(
logits.dtype
).view(-1)
loss = nn.functional.mse_loss(
predicted_yield,
labels,
)
return SequenceClassifierOutput(
loss=loss,
logits=logits,
) |