Problems with EfficientNet

[MyLtYkRiTiK]

Hello! Have you tested this code with EfficientNet (https://github.com/lukemelas/EfficientNet-PyTorch)? I tried to do that, but I have very unrealistic visualizations with huge attention in one place, nearly identical for every image. And I do not understand what the problem.

[kazuto1011]

I think I did it :)

bull mastiff	tiger cat

code:

import matplotlib.pyplot as plt
import matplotlib.cm as cm
import numpy as np
import torch
from efficientnet_pytorch import EfficientNet
from PIL import Image
from torchvision import transforms

from grad_cam import GradCAM

if __name__ == "__main__":

    image_path = "cat_dog.png"
    device = "cpu"

    # Model from torchvision
    target_layer = "_blocks.15"
    model = EfficientNet.from_pretrained("efficientnet-b0")
    model.to(device)
    model.eval()

    # Images
    image = Image.open(image_path)
    raw_image = np.asarray(image)
    image = transforms.Compose(
        [
            transforms.Resize(224),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ]
    )(image)
    images = torch.stack([image]).to(device)

    gcam = GradCAM(model=model)
    _, sorted_ids = gcam.forward(images)
    ids_1st = sorted_ids[:, [0]]  # 'bull_mastiff'
    ids_6th = sorted_ids[:, [5]]  # 'tiger_cat'

    # 1st round for the dog

    gcam.backward(ids=ids_1st)
    heatmap = gcam.generate(target_layer=target_layer)
    heatmap = heatmap.cpu().numpy().squeeze()
    heatmap = cm.turbo(heatmap)[..., :3] * 255.0
    heatmap = (heatmap.astype(np.float) + raw_image.astype(np.float)) / 2
    plt.imshow(np.uint8(heatmap))
    plt.show()

    # 2nd round for the cat

    gcam.backward(ids=ids_6th)
    heatmap = gcam.generate(target_layer=target_layer)
    heatmap = heatmap.cpu().numpy().squeeze()
    heatmap = cm.turbo(heatmap)[..., :3] * 255.0
    heatmap = (heatmap.astype(np.float) + raw_image.astype(np.float)) / 2
    plt.imshow(np.uint8(heatmap))
    plt.show()

[MyLtYkRiTiK]

Hello! Thank you for your code! I tried it with my models and pretrained models and find the same strange thing that i see when tried to run your code on my own.

For example, all block layer for tench image, which is first class at imagenet: tench image You can see strange red area in left top corner. Not all images have the same area, but lots of them have. Problem is the biggest for b4 model, slightly for others. Do you know what can it be? Is it came from your visualization code or what?

For my trained model problem much worse and area is bigger :(

[MyLtYkRiTiK]

Code for reproducing:

import matplotlib.pyplot as plt
import matplotlib.cm as cm
import numpy as np
import torch
from efficientnet_pytorch import EfficientNet
from torchvision import transforms
from PIL import Image
import urllib.request
import io

from grad_cam import GradCAM

def softmax(x):
    """ applies softmax to an input x"""
    e_x = np.exp(x)
    return e_x / e_x.sum()

URL = 'https://cdn.shopify.com/s/files/1/0923/2396/files/1093844_10153046864430043_480555795_o_large.jpg?9023175636926334502'
with urllib.request.urlopen(URL) as url:
    f = io.BytesIO(url.read())

# Images
img = Image.open(f)

device='cpu'
model_name = "efficientnet-b4"
model = EfficientNet.from_pretrained(model_name)
model.to(device)
model.eval()


params_dict = {
    # (width_coefficient, depth_coefficient, resolution, dropout_rate)
    'efficientnet-b0': (1.0, 1.0, 224, 0.2),
    'efficientnet-b1': (1.0, 1.1, 240, 0.2),
    'efficientnet-b2': (1.1, 1.2, 260, 0.3),
    'efficientnet-b3': (1.2, 1.4, 300, 0.3),
    'efficientnet-b4': (1.4, 1.8, 380, 0.4),
    'efficientnet-b5': (1.6, 2.2, 456, 0.4),
    'efficientnet-b6': (1.8, 2.6, 528, 0.5),
    'efficientnet-b7': (2.0, 3.1, 600, 0.5),
    'efficientnet-b8': (2.2, 3.6, 672, 0.5),
    'efficientnet-l2': (4.3, 5.3, 800, 0.5),
}


raw_image = np.asarray(
    transforms.Compose(
    [
       transforms.Resize(int(params_dict[model_name][2] * 1.05)),
            transforms.CenterCrop(params_dict[model_name][2])])(img))
image = transforms.Compose(
    [
        transforms.Resize(int(params_dict[model_name][2] * 1.05)),
            transforms.CenterCrop(params_dict[model_name][2]),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)(img)
images = torch.stack([image]).to(device)

predicted_class = np.argmax(softmax(model(image[None, ...]).detach().numpy()))
print(f'Predicted class {predicted_class}')

gcam = GradCAM(model=model)
_, sorted_ids = gcam.forward(images)
ids_predicted = sorted_ids[:, [predicted_class]]

n_rows = len(model._blocks)//4 if len(model._blocks)%4==0 else len(model._blocks)//4+1
n_cols = 4

fig, ax = plt.subplots(n_rows, n_cols, figsize=(40,40))
gcam.backward(ids=ids_predicted)
layer = 0 
for row in range(n_rows):
    for col in range(n_cols):
        target_layer = f"_blocks.{layer}"
        heatmap = gcam.generate(target_layer=target_layer)
        heatmap = heatmap.cpu().numpy().squeeze()
        heatmap = cm.turbo(heatmap)[..., :3] * 255.0
        heatmap = (heatmap.astype(np.float) + raw_image.astype(np.float)) / 2
        ax[row,col].imshow(np.uint8(heatmap))
        ax[row,col].set_title(target_layer)
        if layer == len(model._blocks)-1:
            break
        else:
            layer += 1

plt.show()

[xiaohuaibaoguigui]

Code for reproducing:

import matplotlib.pyplot as plt
import matplotlib.cm as cm
import numpy as np
import torch
from efficientnet_pytorch import EfficientNet
from torchvision import transforms
from PIL import Image
import urllib.request
import io

from grad_cam import GradCAM

def softmax(x):
    """ applies softmax to an input x"""
    e_x = np.exp(x)
    return e_x / e_x.sum()

URL = 'https://cdn.shopify.com/s/files/1/0923/2396/files/1093844_10153046864430043_480555795_o_large.jpg?9023175636926334502'
with urllib.request.urlopen(URL) as url:
    f = io.BytesIO(url.read())

# Images
img = Image.open(f)

device='cpu'
model_name = "efficientnet-b4"
model = EfficientNet.from_pretrained(model_name)
model.to(device)
model.eval()


params_dict = {
    # (width_coefficient, depth_coefficient, resolution, dropout_rate)
    'efficientnet-b0': (1.0, 1.0, 224, 0.2),
    'efficientnet-b1': (1.0, 1.1, 240, 0.2),
    'efficientnet-b2': (1.1, 1.2, 260, 0.3),
    'efficientnet-b3': (1.2, 1.4, 300, 0.3),
    'efficientnet-b4': (1.4, 1.8, 380, 0.4),
    'efficientnet-b5': (1.6, 2.2, 456, 0.4),
    'efficientnet-b6': (1.8, 2.6, 528, 0.5),
    'efficientnet-b7': (2.0, 3.1, 600, 0.5),
    'efficientnet-b8': (2.2, 3.6, 672, 0.5),
    'efficientnet-l2': (4.3, 5.3, 800, 0.5),
}


raw_image = np.asarray(
    transforms.Compose(
    [
       transforms.Resize(int(params_dict[model_name][2] * 1.05)),
            transforms.CenterCrop(params_dict[model_name][2])])(img))
image = transforms.Compose(
    [
        transforms.Resize(int(params_dict[model_name][2] * 1.05)),
            transforms.CenterCrop(params_dict[model_name][2]),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)(img)
images = torch.stack([image]).to(device)

predicted_class = np.argmax(softmax(model(image[None, ...]).detach().numpy()))
print(f'Predicted class {predicted_class}')

gcam = GradCAM(model=model)
_, sorted_ids = gcam.forward(images)
ids_predicted = sorted_ids[:, [predicted_class]]

n_rows = len(model._blocks)//4 if len(model._blocks)%4==0 else len(model._blocks)//4+1
n_cols = 4

fig, ax = plt.subplots(n_rows, n_cols, figsize=(40,40))
gcam.backward(ids=ids_predicted)
layer = 0 
for row in range(n_rows):
    for col in range(n_cols):
        target_layer = f"_blocks.{layer}"
        heatmap = gcam.generate(target_layer=target_layer)
        heatmap = heatmap.cpu().numpy().squeeze()
        heatmap = cm.turbo(heatmap)[..., :3] * 255.0
        heatmap = (heatmap.astype(np.float) + raw_image.astype(np.float)) / 2
        ax[row,col].imshow(np.uint8(heatmap))
        ax[row,col].set_title(target_layer)
        if layer == len(model._blocks)-1:
            break
        else:
            layer += 1

plt.show()

there are some error ,when i run with it . File "effcietn_cam.py", line 80, in heatmap = cm.turbo(heatmap)[..., :3] * 255.0 AttributeError: module 'matplotlib.cm' has no attribute 'turbo'

[kazuto1011]

The turbo colormap is available from matplotlib 3.3.0. Please upgrade it or use jet instead. https://matplotlib.org/3.3.3/users/whats_new.html#turbo-colormap