best_epoch_file = get_saved_model_path(best_epoch)
print(
f'\nEpoch: {best_epoch} - New best accuracy! Accuracy: {best_acc}\n\n\n'
)
torch.save(model.state_dict(), best_epoch_file)
# #### Progress bar - uncomment when testing in notebook
# In[15]:
# pbar = ProgressBar(bar_format='')
# pbar.attach(trainer, output_transform=lambda x: {'loss': x})
# #### Train
# In[16]:
print('Training started\n')
trainer.run(loader, max_epochs=8)
# #### Evaluate
# In[ ]:
all_preds, _ = eval_model(model, tloader, best_epoch_file, path_data)
# In[ ]:
apply_plates_leak(all_preds)
os.remove(prev_best_epoch_file)
best_acc = metrics['accuracy']
best_epoch = epoch
best_epoch_file = get_saved_model_path(best_epoch)
print(
f'\nEpoch: {best_epoch} - New best accuracy! Accuracy: {best_acc}\n\n\n'
)
torch.save(model.state_dict(), best_epoch_file)
# #### Progress bar - uncomment when testing in notebook
# In[15]:
# pbar = ProgressBar(bar_format='')
# pbar.attach(trainer, output_transform=lambda x: {'loss': x})
# #### Train
# In[16]:
print('Training started\n')
trainer.run(loader, max_epochs=2)
# #### Evaluate
# In[ ]:
eval_model(model, tloader, best_epoch_file, path_data)
torch.save(model.state_dict(), best_epoch_file)
pbar = ProgressBar(bar_format='')
pbar.attach(trainer, output_transform=lambda x: {'loss': x})
print('Training started\n')
# trainer.run(loader, max_epochs=epochs_per_cell[cell])
trainer.run(loader, max_epochs=1)
tloader = D.DataLoader(cat_test_df,
batch_size=1,
shuffle=False,
num_workers=15)
cell_preds, _ = eval_model(model,
tloader,
best_epoch_file,
path_data,
sub_file=f'submission_{cell}.csv')
all_preds += cell_preds
# In[ ]:
# aggregate submission files
submissions = []
for cell in cells:
submissions += [pd.read_csv(f'submission_{cell}.csv')]
submissions = pd.concat(submissions)
submissions.to_csv(f'submission.csv',
index=False,
columns=['id_code', 'sirna'])