optimizer,
criterion,
lr_scheduler=scheduler,
metrics={
"acc2": MoseiAcc2(exclude_neutral=True),
"acc2_zero": MoseiAcc2(exclude_neutral=False),
"acc5": MoseiAcc5(),
"acc7": MoseiAcc7(),
"mae": torchmetrics.MeanAbsoluteError(),
},
)
trainer = make_trainer(
EXPERIMENT_NAME,
max_epochs=100,
patience=10,
gpus=1,
save_top_k=1,
)
# watch_model(trainer, model)
trainer.fit(lm, datamodule=ldm)
# trainer.test(ckpt_path="best", test_dataloaders=ldm.test_dataloader())
best_model = RnnPLModule.load_from_checkpoint(
trainer.checkpoint_callback.best_model_path,
model=model,
optimizer=optimizer,
criterion=criterion,
lr_scheduler=scheduler,
)
lm = MultimodalTransformerClassificationPLModule(
model,
optimizer,
criterion,
metrics={
"acc2": MoseiAcc2(exclude_neutral=True),
"acc2_zero": MoseiAcc2(exclude_neutral=False),
"acc5": MoseiAcc5(),
"acc7": MoseiAcc7(),
"f1": MoseiF1(exclude_neutral=True),
"f1_zero": MoseiF1(exclude_neutral=False),
"mae": torchmetrics.MeanAbsoluteError(),
},
)
trainer = make_trainer(**config.trainer)
watch_model(trainer, model)
trainer.fit(lm, datamodule=ldm)
results = trainer.test(ckpt_path="best",
test_dataloaders=ldm.test_dataloader())
import csv
import os
csv_folder_path = os.path.join(config.trainer.experiments_folder,
config.trainer.experiment_name,
"results_csv")
csv_name = os.path.join(csv_folder_path, "results.csv")
if __name__ == "__main__":
EXPERIMENT_NAME = "mnist-autoencoder"
configure_logging(f"logs/{EXPERIMENT_NAME}")
train, test = get_data()
ldm = PLDataModuleFromDatasets(train,
test=test,
batch_size=128,
batch_size_eval=256)
model = Net()
optimizer = Adam(model.parameters(), lr=1e-3)
criterion = nn.MSELoss()
lm = AutoEncoderPLModule(model, optimizer, criterion)
trainer = make_trainer(EXPERIMENT_NAME,
max_epochs=5,
gpus=1,
wandb_project="testpl")
watch_model(trainer, model)
trainer.fit(lm, datamodule=ldm)
trainer.test(ckpt_path="best", test_dataloaders=ldm.test_dataloader())
num_layers=2,
num_heads=2,
dropout=0.3,
hidden_size=100,
inner_size=200,
prenorm=True,
scalenorm=True,
)
optimizer = AdamW([p for p in model.parameters() if p.requires_grad],
lr=1e-4)
criterion = nn.MSELoss()
lm = MultimodalTransformerClassificationPLModule(
model,
optimizer,
criterion,
# metrics={"acc": FromLogits(pl.metrics.classification.Accuracy())},
)
trainer = make_trainer(EXPERIMENT_NAME,
max_epochs=100,
gpus=1,
save_top_k=1,
gradient_clip_val=1.0)
watch_model(trainer, model)
trainer.fit(lm, datamodule=ldm)
trainer.test(ckpt_path="best", test_dataloaders=ldm.test_dataloader())