def test_gbp(self):
model = medcam.inject(self.model, output_dir=os.path.join(self.current_path, 'results/resnet152/test_gbp'), backend='gbp',
evaluate=False, save_scores=False, save_maps=True, save_pickle=False)
model.eval()
data_loader = DataLoader(self.dataset, batch_size=1, shuffle=False)
for i, batch in enumerate(data_loader):
_ = model(batch[0][0])
del model
gc.collect()
torch.cuda.empty_cache()
if CLEAR and os.path.isdir(os.path.join(self.current_path, 'results/resnet152')):
shutil.rmtree(os.path.join(self.current_path, 'results/resnet152'))
def test_gcampp(self):
layer = 'full'
metric = 'wioa'
model = medcam.inject(self.model, output_dir=os.path.join(self.current_path, 'results/unet_seg/gcampp'), backend='gcampp', layer=layer,
evaluate=True, save_scores=False, save_maps=True, save_pickle=False, metric=metric, label=lambda x: 0.5 < x)
model.eval()
data_loader = DataLoader(self.dataset, batch_size=1, shuffle=False)
model.test_run(next(iter(data_loader))["img"])
for i, batch in enumerate(data_loader):
_ = model(batch["img"], mask=batch["gt"])
del model
gc.collect()
torch.cuda.empty_cache()
if CLEAR and os.path.isdir(os.path.join(self.current_path, 'results/unet_seg')):
shutil.rmtree(os.path.join(self.current_path, 'results/unet_seg'))
import torch
from numpy import nper
from medcam import medcam
from c3d import C3D
from custom_dataset import ucf_c3d
from torch.utils.data import Dataset, DataLoader
from vist_util import *
model = C3D(num_classes=101,pretrained=True)
loadfrom = 'c3d46perc.dict'
model.load_state_dict(torch.load(loadfrom))
print(model)
full_dataset = ucf_c3d('video_annotations_resnet','G:\\C3D Data', kernels=['original'])
dataloader = DataLoader(full_dataset, batch_size=1, shuffle=True, num_workers=0, pin_memory = True)
model = medcam.inject(model, output_dir="attention_maps", layer = 'conv5b', backend='gcam', return_attention = True, save_maps=True)
model.eval()
for batch in dataloader:
video_matrix = batch['original']
action_list = batch['action']
score, attention = model(video_matrix)
#print(score.shape)
attention = torch.squeeze(attention).numpy()
video_matrix = torch.squeeze(video_matrix).numpy().astype(np.int)#3,16,112,112
#print(video_matrix)
video_matrix = np.transpose(video_matrix, (1,2,3,0))
#print(attention.shape)
#print(video_matrix.shape)
showbunch(attention)
showbunch(video_matrix)
def training_loop(
training_data,
validation_data,
device,
params,
output_dir,
testing_data=None,
epochs=None,
):
"""
The main training loop.
Args:
training_data (pandas.DataFrame): The data to use for training.
validation_data (pandas.DataFrame): The data to use for validation.
device (str): The device to perform computations on.
params (dict): The parameters dictionary.
output_dir (str): The output directory.
testing_data (pandas.DataFrame): The data to use for testing.
epochs (int): The number of epochs to train; if None, take from params.
"""
# Some autodetermined factors
if epochs is None:
epochs = params["num_epochs"]
params["device"] = device
params["output_dir"] = output_dir
params["training_data"] = training_data
params["validation_data"] = validation_data
params["testing_data"] = testing_data
testingDataDefined = True
if params["testing_data"] is None:
# testing_data = validation_data
testingDataDefined = False
# Defining our model here according to parameters mentioned in the configuration file
print("Number of channels : ", params["model"]["num_channels"])
(
model,
optimizer,
train_dataloader,
val_dataloader,
scheduler,
params,
) = create_pytorch_objects(params, training_data, validation_data, device)
if testingDataDefined:
test_dataloader = get_testing_loader(params)
# Start training time here
start_time = time.time()
if not (os.environ.get("HOSTNAME") is None):
print("Hostname :", os.environ.get("HOSTNAME"))
# datetime object containing current date and time
print("Initializing training at :", get_date_time(), flush=True)
# Setup a few loggers for tracking
train_logger = Logger(
logger_csv_filename=os.path.join(output_dir, "logs_training.csv"),
metrics=params["metrics"],
)
valid_logger = Logger(
logger_csv_filename=os.path.join(output_dir, "logs_validation.csv"),
metrics=params["metrics"],
)
test_logger = Logger(
logger_csv_filename=os.path.join(output_dir, "logs_testing.csv"),
metrics=params["metrics"],
)
train_logger.write_header(mode="train")
valid_logger.write_header(mode="valid")
test_logger.write_header(mode="test")
if "medcam" in params:
model = medcam.inject(
model,
output_dir=os.path.join(
output_dir, "attention_maps", params["medcam"]["backend"]
),
backend=params["medcam"]["backend"],
layer=params["medcam"]["layer"],
save_maps=False,
return_attention=True,
enabled=False,
)
params["medcam_enabled"] = False
# Setup a few variables for tracking
best_loss = 1e7
patience, start_epoch = 0, 0
first_model_saved = False
best_model_path = os.path.join(
output_dir, params["model"]["architecture"] + best_model_path_end
)
# if previous model file is present, load it up
if os.path.exists(best_model_path):
try:
main_dict = load_model(best_model_path, params["device"])
version_check(params["version"], version_to_check=main_dict["version"])
model.load_state_dict(main_dict["model_state_dict"])
start_epoch = main_dict["epoch"]
optimizer.load_state_dict(main_dict["optimizer_state_dict"])
best_loss = main_dict["loss"]
print("Previous model successfully loaded.")
except RuntimeWarning:
RuntimeWarning("Previous model could not be loaded, initializing model")
print("Using device:", device, flush=True)
# Iterate for number of epochs
for epoch in range(start_epoch, epochs):
if params["track_memory_usage"]:
file_to_write_mem = os.path.join(output_dir, "memory_usage.csv")
if os.path.exists(file_to_write_mem):
# append to previously generated file
file_mem = open(file_to_write_mem, "a")
outputToWrite_mem = ""
else:
# if file was absent, write header information
file_mem = open(file_to_write_mem, "w")
outputToWrite_mem = "Epoch,Memory_Total,Memory_Available,Memory_Percent_Free,Memory_Usage," # used to write output
if params["device"] == "cuda":
outputToWrite_mem += "CUDA_active.all.peak,CUDA_active.all.current,CUDA_active.all.allocated"
outputToWrite_mem += "\n"
mem = psutil.virtual_memory()
outputToWrite_mem += (
str(epoch)
+ ","
+ str(mem[0])
+ ","
+ str(mem[1])
+ ","
+ str(mem[2])
+ ","
+ str(mem[3])
)
if params["device"] == "cuda":
mem_cuda = torch.cuda.memory_stats()
outputToWrite_mem += (
","
+ str(mem_cuda["active.all.peak"])
+ ","
+ str(mem_cuda["active.all.current"])
+ ","
+ str(mem_cuda["active.all.allocated"])
)
outputToWrite_mem += ",\n"
file_mem.write(outputToWrite_mem)
file_mem.close()
# Printing times
epoch_start_time = time.time()
print("*" * 20)
print("*" * 20)
print("Starting Epoch : ", epoch)
if params["verbose"]:
print("Epoch start time : ", get_date_time())
params["current_epoch"] = epoch
epoch_train_loss, epoch_train_metric = train_network(
model, train_dataloader, optimizer, params
)
epoch_valid_loss, epoch_valid_metric = validate_network(
model, val_dataloader, scheduler, params, epoch, mode="validation"
)
patience += 1
# Write the losses to a logger
train_logger.write(epoch, epoch_train_loss, epoch_train_metric)
valid_logger.write(epoch, epoch_valid_loss, epoch_valid_metric)
if testingDataDefined:
epoch_test_loss, epoch_test_metric = validate_network(
model, test_dataloader, scheduler, params, epoch, mode="testing"
)
test_logger.write(epoch, epoch_test_loss, epoch_test_metric)
if params["verbose"]:
print("Epoch end time : ", get_date_time())
epoch_end_time = time.time()
print(
"Time taken for epoch : ",
(epoch_end_time - epoch_start_time) / 60,
" mins",
flush=True,
)
# Start to check for loss
if not (first_model_saved) or (epoch_valid_loss <= torch.tensor(best_loss)):
best_loss = epoch_valid_loss
best_train_idx = epoch
patience = 0
model.eval()
save_model(
{
"epoch": best_train_idx,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"loss": best_loss,
},
model,
params,
best_model_path,
onnx_export=False,
)
model.train()
first_model_saved = True
if params["model"]["save_at_every_epoch"]:
save_model(
{
"epoch": epoch,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"loss": epoch_valid_loss,
},
model,
params,
os.path.join(
output_dir,
params["model"]["architecture"]
+ "_epoch_"
+ str(epoch)
+ ".pth.tar",
),
onnx_export=False,
)
model.train()
print("Current Best epoch: ", best_train_idx)
if patience > params["patience"]:
print(
"Performance Metric has not improved for %d epochs, exiting training loop!"
% (patience),
flush=True,
)
break
# End train time
end_time = time.time()
print(
"Total time to finish Training : ",
(end_time - start_time) / 60,
" mins",
flush=True,
)
# once the training is done, optimize the best model
if os.path.exists(best_model_path):
onnx_export = True
if params["model"]["architecture"] in ["sdnet", "brain_age"]:
onnx_export = False
elif (
"onnx_export" in params["model"] and params["model"]["onnx_export"] == False
):
onnx_export = False
if onnx_export:
print("Optimizing best model.")
try:
main_dict = load_model(best_model_path, params["device"])
version_check(params["version"], version_to_check=main_dict["version"])
model.load_state_dict(main_dict["model_state_dict"])
best_epoch = main_dict["epoch"]
optimizer.load_state_dict(main_dict["optimizer_state_dict"])
best_loss = main_dict["loss"]
save_model(
{
"epoch": best_epoch,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"loss": best_loss,
},
model,
params,
best_model_path,
onnx_export,
)
except Exception as e:
print("Best model could not be loaded, error: ", e)