Hello,
I am a beginner in PyTorch and Deep Learning in general. I am trying to train a maskrcnn_resnet50_fpn (https://pytorch.org/docs/stable/torchvision/models.html#torchvision.models.detection.maskrcnn_resnet50_fpn) but I cannot make the model converge even when using 10 Epocs to train a single image. I am basically trying to overfit my model using one training example in order to do a sanity check as there’s no point in training the model on gigabytes of data using a GPU when I can’t even overfit it by training on a single image.
I created a toy example which can be found at Capstone/toyexample/toylearner.py at master · fsafe/Capstone · GitHub
The above scrips uses only one image which I uploaded to this post ( Capstone/toyexample/004408_01_02_088.png at master · fsafe/Capstone · GitHub )
In summary I pass the model an image of a CT scan, labels, a bounding box and a segmentation mask. In the example in the above GitHub you can either use the original image or the cropped version. All you need to do is change the inputs to the model from (inputs, targets) to (inputs_crop, targets_crop). The model outputs 5 loss functions and I sum up all losses in each Epoc to obtain the total training loss. When I train using the uncropped version I get the following:
Epoch 0/9
loss_classifier:tensor(0.6876, grad_fn=)
loss_box_reg:tensor(1.3961e-06, grad_fn=)
loss_mask:tensor(1.2238, grad_fn=)
loss_objectness:tensor(0.6945, grad_fn=)
loss_rpn_box_reg:tensor(0.0133, grad_fn=)
Total Train Loss: 2.6192
Epoch 1/9
loss_classifier:tensor(0.6812, grad_fn=)
loss_box_reg:tensor(0.0225, grad_fn=)
loss_mask:tensor(0.7030, grad_fn=)
loss_objectness:tensor(0.6937, grad_fn=)
loss_rpn_box_reg:tensor(0.0104, grad_fn=)
Total Train Loss: 2.1108
Epoch 2/9
loss_classifier:tensor(0.6727, grad_fn=)
loss_box_reg:tensor(5.6627e-07, grad_fn=)
loss_mask:tensor(0.6804, grad_fn=)
loss_objectness:tensor(0.6932, grad_fn=)
loss_rpn_box_reg:tensor(0.0035, grad_fn=)
Total Train Loss: 2.0498
Epoch 3/9
loss_classifier:tensor(0.6730, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6938, grad_fn=)
loss_objectness:tensor(0.6935, grad_fn=)
loss_rpn_box_reg:tensor(41.7319, grad_fn=)
Total Train Loss: 43.7921
Epoch 4/9
loss_classifier:tensor(0.6734, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6936, grad_fn=)
loss_objectness:tensor(0.6932, grad_fn=)
loss_rpn_box_reg:tensor(79.3949, grad_fn=)
Total Train Loss: 81.4552
Epoch 5/9
loss_classifier:tensor(0.6725, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6934, grad_fn=)
loss_objectness:tensor(0.6927, grad_fn=)
loss_rpn_box_reg:tensor(140.8336, grad_fn=)
Total Train Loss: 142.8922
Epoch 6/9
loss_classifier:tensor(0.6727, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6933, grad_fn=)
loss_objectness:tensor(0.6936, grad_fn=)
loss_rpn_box_reg:tensor(232.2627, grad_fn=)
Total Train Loss: 234.3224
Epoch 7/9
loss_classifier:tensor(0.6723, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6932, grad_fn=)
loss_objectness:tensor(0.6929, grad_fn=)
loss_rpn_box_reg:tensor(362.7281, grad_fn=)
Total Train Loss: 364.7866
Epoch 8/9
loss_classifier:tensor(0.6729, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6932, grad_fn=)
loss_objectness:tensor(0.6926, grad_fn=)
loss_rpn_box_reg:tensor(577.8825, grad_fn=)
Total Train Loss: 579.9412
Epoch 9/9
loss_classifier:tensor(0.6734, grad_fn=)
loss_box_reg:tensor(1.4735e-06, grad_fn=)
loss_mask:tensor(0.6932, grad_fn=)
loss_objectness:tensor(0.6937, grad_fn=)
loss_rpn_box_reg:tensor(932.3105, grad_fn=)
Total Train Loss: 934.3708
Training complete in 4m 50s
Process finished with exit code 0
As you can see the RPN loss increases significantly after the third epoc.
The above run was on model that was NOT pretrained but in the code you can just set pretrained=True but that does not improve matters. Keep in mid some model surgery needs to be done when using a pretrained model ( pretrained on COCO dataset ) and I’m not sure if I am doing this part right but it appears to be right.
The above training (for a single image) was done on a CPU.
Here is the code :
import cv2
import time
import numpy as np
import torch
from torchvision.transforms import functional as TF
from torch.optim import lr_scheduler, SGD
from torchvision.models.detection import maskrcnn_resnet50_fpn
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
from torchvision.models.detection.mask_rcnn import MaskRCNNPredictor
# load image from disk
img = cv2.imread('004408_01_02_088.png', -1)
# subtract 32768from the pixel intensity to obtain the original Hounsfield unit (HU) values
# https://nihcc.app.box.com/v/DeepLesion/file/306056134060
img = img.astype(np.float32, copy=False) - 32768
# intensity windowing with window [-1024, 3071] HU covers the intensity ranges of the lung, soft tissue, and bone.
# (https://arxiv.org/pdf/1806.09648.pdf)
# convert the intensities in a certain range (“window”) to 0-255 for viewing.
img -= -1024
img /= 3071 + 1024
img[img > 1] = 1
img[img < 0] = 0
img *= 255
img = img.astype('uint8')
# convert image to tensor. The output tensor will have range [0,1]
img_T = TF.to_tensor(img)
# create numpy array version of img_T Tensor and adding a blue pseudo_mask
# the addition of the mask and bounding box does not affect original image tensor (img_T)
# on this numpy version by combining 4 quarter sized ellipses. Also add a green bounding box.
# https://arxiv.org/pdf/1901.06359.pdf
img_copy = [img_T.squeeze().numpy()] * 3
img_copy = cv2.merge(img_copy)
bbox = np.array([188.354, 159.003, 223.22, 183.271])
bbox = np.int16(bbox)
cen = np.array([212.17824058, 171.81745919])
semi_axes = np.array([7, 7, 17, 6])
angles = np.array([0.94002174, -179.05997826])
cv2.ellipse(img_copy, tuple(cen.astype(int)), tuple(semi_axes[0:2]), angles[0], 0, 90, 255, -1)
cv2.ellipse(img_copy, tuple(cen.astype(int)), tuple(semi_axes[2:0:-1]), angles[1], -90, 0, 255, -1)
cv2.ellipse(img_copy, tuple(cen.astype(int)), tuple(semi_axes[2:4]), angles[1], 0, 90, 255, -1)
cv2.ellipse(img_copy, tuple(cen.astype(int)), tuple([semi_axes[0], semi_axes[3]]), angles[0], -90, 0, 255, -1)
cv2.rectangle(img_copy, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 255, 0), 1)
# extract pseudo_mask by identifying pixels which are colored blue
pseudo_mask = np.logical_and(img_copy[:, :, 0] == 255, img_copy[:, :, 1] == 0, img_copy[:, :, 2] == 0).astype('uint8')
pseudo_mask_T = torch.from_numpy(pseudo_mask)
# construct inputs to model
inputs = [img_T]
bbox_T = torch.from_numpy(bbox).float()
bboxes = [bbox_T]
bboxes = torch.stack(bboxes)
masks = [pseudo_mask_T]
masks = torch.stack(masks)
label = torch.ones(len(bboxes), dtype=torch.int64)
elem = {'boxes': bboxes, 'masks': masks, 'labels': label}
targets = [elem]
# # uncomment this block to check if inputs to model can be displayed correctly
# for (image, target) in zip(inputs, targets):
# img_display = image.squeeze().numpy()
# images_disp = [img_display] * 3
# images_disp = [im.astype(float) for im in images_disp]
# img_display = cv2.merge(images_disp)
# for (bbox_disp, pseudo_mask_disp) in zip(target["boxes"], target["masks"]):
# bbox_disp = bbox_disp.squeeze().numpy()
# bbox_disp = np.int16(bbox)
# mask_disp = pseudo_mask_disp.squeeze().numpy()
# cv2.rectangle(img_display, (bbox_disp[0], bbox_disp[1]), (bbox_disp[2], bbox_disp[3]), (0, 255, 0), 1)
# msk_idx = np.where(mask_disp == 1)
# img_display[msk_idx[0], msk_idx[1], 0] = 255
# cv2.imshow('original', img_display)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
# crop image by clipping black borders
img_crop = img_T.squeeze().numpy()
u, d, l, r = (115, 430, 0, 511)
img_crop = img_crop[u:d + 1, l:r + 1]
bbox_crop = np.array([0, 0, 0, 0])
bbox_crop[0] = bbox[0] - l
bbox_crop[1] = bbox[1] - u
bbox_crop[2] = bbox[2] - l
bbox_crop[3] = bbox[3] - u
bbox_crop = np.int16(bbox_crop)
pseudo_mask_crop = pseudo_mask[u:d + 1, l:r + 1]
# construct inputs to model
img_crop_T = TF.to_tensor(img_crop)
inputs_crop = [img_crop_T]
bbox_crop_T = torch.from_numpy(bbox_crop).float()
bboxes_crop = [bbox_crop_T]
bboxes_crop = torch.stack(bboxes_crop)
pseudo_mask_crop_T = torch.from_numpy(pseudo_mask_crop)
masks_crop = [pseudo_mask_crop_T]
masks_crop = torch.stack(masks_crop)
label_crop = torch.ones(len(bboxes_crop), dtype=torch.int64)
elem_crop = {'boxes': bboxes_crop, 'masks': masks_crop, 'labels': label_crop}
targets_crop = [elem_crop]
# # uncomment this block to check if inputs to model can be displayed correctly
# for (image, target) in zip(inputs_crop, targets_crop):
# img_display = image.squeeze().numpy()
# images_disp = [img_display] * 3
# images_disp = [im.astype(float) for im in images_disp]
# img_display = cv2.merge(images_disp)
# for (bbox_disp, pseudo_mask_disp) in zip(target["boxes"], target["masks"]):
# bbox_disp = bbox_disp.squeeze().numpy()
# bbox_disp = np.int16(bbox_disp)
# mask_disp = pseudo_mask_disp.squeeze().numpy()
# cv2.rectangle(img_display, (bbox_disp[0], bbox_disp[1]), (bbox_disp[2], bbox_disp[3]), (0, 255, 0), 1)
# msk_idx = np.where(mask_disp == 1)
# img_display[msk_idx[0], msk_idx[1], 0] = 255
# cv2.imshow('cropped', img_display)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
pretrained = False
if pretrained:
model = maskrcnn_resnet50_fpn(pretrained=True)
for param in model.parameters():
param.requires_grad = False
num_classes = 2 # 1 class (lesion) + 0 (background)
# get number of input features for the classifier
in_features = model.roi_heads.box_predictor.cls_score.in_features
# replace the pre-trained head with a new one
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
# now get the number of input features for the mask classifier
in_features_mask = model.roi_heads.mask_predictor.conv5_mask.in_channels
hidden_layer = 64
# and replace the mask predictor with a new one
model.roi_heads.mask_predictor = MaskRCNNPredictor(in_features_mask,
hidden_layer,
num_classes)
params = [p for p in model.parameters() if p.requires_grad]
# Observe that not all parameters are being optimized
optimizer_ft = SGD(params, lr=0.001, momentum=0.9, weight_decay=0.0001)
else:
# Observe that all parameters are being optimized
model = maskrcnn_resnet50_fpn(num_classes=2)
optimizer_ft = SGD(model.parameters(), lr=0.001, momentum=0.9, weight_decay=0.0001)
# don't know how to initialize the weights of the model
# torch.nn.init.kaiming_normal_(model.parameters(), mode='fan_out')
# Decay LR by a factor of 0.1 every 2 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=2, gamma=0.1)
num_epochs = 100
since = time.time()
model.train()
print('Pretrained:' + str(pretrained))
print('momentum:' + str(optimizer_ft.state_dict()['param_groups'][0]['momentum']))
print('weight_decay:' + str(optimizer_ft.state_dict()['param_groups'][0]['weight_decay']))
print('LR decay gamma:' + str(exp_lr_scheduler.state_dict()['gamma']))
print('LR decay step size:' + str(exp_lr_scheduler.state_dict()['step_size']))
for epoch in range(num_epochs):
print('\nEpoch {}/{}'.format(epoch, num_epochs - 1))
print('-' * 10)
print('lr:' + str(optimizer_ft.state_dict()['param_groups'][0]['lr']))
# valid inputs to model are (inputs, targets) or (inputs_crop, targets_crop)
loss_dict = model(inputs, targets)
for (k, i) in loss_dict.items():
print(str(k) + ':' + str(i))
losses = sum(loss for loss in loss_dict.values())
del loss_dict
print('Total Train Loss: {:.4f}'.format(losses.item()))
# zero the parameter gradients
optimizer_ft.zero_grad()
# perform backward propagation, optimization and update model parameters
losses.backward()
optimizer_ft.step()
exp_lr_scheduler.step()
del losses
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
