app.py

import gradio as gr
from matplotlib import cm 
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import make_axes_locatable
import numpy as np
# import onnxruntime as ort
from PIL import Image
from scipy import special
import sys
from types import SimpleNamespace
from transformers import AutoModel
import torch

# sys.path.insert(1, "../")
# from utils import model_utils, train_utils, data_utils, run_utils
# from model_utils import jason_regnet_maker, jason_efficientnet_maker

model_path = 'chlab/'
# model_path = './models/'

# plotting a prameters
labels = 20
ticks = 14
legends = 14
text = 14
titles = 22
lw = 3
ps = 200
cmap = 'magma'

effnet_61_hparams = {
    "num_classes": 2,
    "gamma": 0.032606396652426956,
    "lr": 0.008692971067922545,
    "weight_decay": 0.00008348389688708425,
    "batch_size": 23,
    "num_channels": 61,
    "stochastic_depth_prob": 0.003581930052432713,
    "dropout": 0.027804120950575217,
    "width_mult": 1.060782511229692,
    "depth_mult": 0.7752918857163054,
}
effnet_61_config = SimpleNamespace(**effnet_61_hparams)

# which layers to look at
activation_indices = {'efficientnet': [0, 3]}


def normalize_array(x: list):

    '''Makes array between 0 and 1'''
    
    x = np.array(x)
    
    return (x - np.min(x)) / np.max(x - np.min(x))

# def load_model(model: str, activation: bool=True):
    
#     if activation:
#         model += '_w_activation'
    
#     # set options for onnx runtime
#     options = ort.SessionOptions()
#     options.intra_op_num_threads = 1
#     options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
#     provider = "CPUExecutionProvider"
    
#     # start session
#     ort_session = ort.InferenceSession(model_path + '%s.onnx' % (model), options, providers=[provider])
#     # ort_session = ORTModel.load_model(model_path + '%s.onnx' % (model))
    
#     return ort_session

def get_activations(model, image: list, model_name: str,
                    layer=None, vmax=2.5, sub_mean=True,
                    channel: int=0):
    
    '''Gets activations for a given input image'''
    
    # run model
    # input_name = intermediate_model.get_inputs()[0].name
    # outputs = intermediate_model.run(None, {input_name: image})
    
    
    layer_outputs = {}
    for i in range(len(model.model.features)):
        image = model.model.features[i](image)
        layer_outputs[i] = image
        print(i, layer_outputs[i].shape)
    output = model.model(image).detach().cpu().numpy()
    output_1 = activation_indices[model_name].detach().cpu().numpy()
    output_2 = activation_indices[model_name].detach().cpu().numpy()
    
    # get activations
    # output_1 = outputs[1]
    # output_2 = outputs[2]
    
    # get prediction
    # output = outputs[0][0]
    output = special.softmax(output)
    
    # sum over velocity channels
    if channel == 0:
        in_image = np.sum(image[0, :, :, :], axis=0)
    else:
        image[0, int(channel-1), :, :]
    in_image = normalize_array(in_image)

    if layer is None:
        # sum over all velocity channels
        activation_1 = np.sum(output_1[0, :, :, :], axis=0)
        activation_2 = np.sum(output_2[0, :, :, :], axis=0)
    else:
        # select a single channel
        activation_1 = output_1[0, layer, :, :]
        activation_2 = output_2[0, layer, :, :]
    
    if sub_mean:
        # y = |x - <x>|
        activation_1 -= np.mean(activation_1)
        activation_1 = np.abs(activation_1)
        
        activation_2 -= np.mean(activation_2)
        activation_2 = np.abs(activation_2)
    
    return output, in_image, activation_1, activation_2

def plot_input(input_image: list, origin='lower'):
    
    ##### make the figure for the input image #####
    plt.rcParams['xtick.labelsize'] = ticks
    plt.rcParams['ytick.labelsize'] = ticks
    
    input_fig, ax = plt.subplots(nrows=1, ncols=1, figsize=(18, 8))
    
    im0 = ax.imshow(input_image, cmap=cmap,
                     origin=origin)    
    
    divider = make_axes_locatable(ax)
    cax = divider.append_axes('right', size='5%', pad=0.05)
    input_fig.colorbar(im0, cax=cax, orientation='vertical')
        
    ax.set_title('Input', fontsize=titles)
    
    return input_fig

def plot_activations(activation_1: list, activation_2: list, origin='lower'):
    
    
     ##### Make the activation figure ######
    plt.rcParams['xtick.labelsize'] = ticks
    plt.rcParams['ytick.labelsize'] = ticks
    
    fig, axs = plt.subplots(nrows=1, ncols=2, figsize=(27, 12))
    
    ax1, ax2 = axs[0], axs[1]
    
    im1 = ax1.imshow(activation_1, cmap=cmap,
                     origin=origin)
    im2 = ax2.imshow(activation_2, cmap=cmap, 
                     origin=origin) 
    
    ims = [im1, im2]
    
    for (i, ax) in enumerate(axs):
        divider = make_axes_locatable(ax)
        cax = divider.append_axes('right', size='5%', pad=0.05)
        fig.colorbar(ims[i], cax=cax, orientation='vertical')
        
    # ax0.set_title('Input', fontsize=titles)
    ax1.set_title('Early Activation', fontsize=titles)
    ax2.set_title('Late Activation', fontsize=titles)
    
    return fig


def predict_and_analyze(model_name, num_channels, dim, input_channel, image):
    
    '''
    Loads a model with activations, passes through image and shows activations
    
    The image must be a numpy array of shape (C, W, W) or (1, C, W, W) 
    '''
    
    model_name = model_name.lower()
    num_channels = int(num_channels)
    W = int(dim)

    print("Loading data")
    print(image)
    image = np.load(image.name, allow_pickle=True)
    image = image.astype(np.float32)
    
    if len(image.shape) != 4:
        image = image[np.newaxis, :, :, :]
        
    image = torch.from_numpy(image)
        
    assert image.shape == (1, num_channels, W, W), "Data is the wrong shape"
    
    # pipeline = pipeline(task="image-classification", model=model_path + "%s_%i_.pyt" % (model_name, num_channels))
    
    # model_name += '_%i' % (num_channels)
    
    model_loading_name =  model_path + "%s_%i_planet_detection" % (model_name, num_channels)
    
    print("Loading model")
    # model = load_model(model_name, activation=True)
    model = AutoModel.from_pretrained(model_loading_name)
    print("Model loaded")
    
    print("Looking at activations")
    output, input_image, activation_1, activation_2 = get_activations(model, image, model_name, 
                                                                      channel=input_channel,
                                                                      sub_mean=True)
    print("Activations and predictions finished")
    
    if output[0] < output[1]:
        output = 'Planet predicted with %.3f percent confidence' % (100*output[1])
    else:
        output = 'No planet predicted with %.3f percent confidence' % (100*output[0])
        
    input_image = normalize_array(input_image)
    activation_1 = normalize_array(activation_1)
    activation_2 = normalize_array(activation_2)
    
    # convert input image to RGB (unused for now since not outputting actual image)
    # input_pil_image = Image.fromarray(np.uint8(cm.magma(input_image)*255))
    
    print("Plotting")
    
    origin = 'upper'
    
    # plot input image
    input_fig = plot_input(input_image, origin=origin)
    
    # plot mean subtracted activations
    fig1 = plot_activations(activation_1, activation_2, model_name, origin=origin)
    
    # plot raw activations
    _, _, activation_1, activation_2 = get_activations(model, image, model_name, 
                                                       channel=input_channel,
                                                       sub_mean=False)
    activation_1 = normalize_array(activation_1)
    activation_2 = normalize_array(activation_2)
    fig2 = plot_activations(activation_1, activation_2, model_name, origin=origin)
    
    print("Sending to Hugging Face")
    
    return output, input_fig, fig1, fig2


if __name__ == "__main__":

    demo = gr.Interface(
        fn=predict_and_analyze,
        inputs=[gr.Dropdown(["EfficientNet", "RegNet"], 
                            value="EfficientNet",
                            label="Model Selection",
                            show_label=True), 
                gr.Dropdown(["47", "61", "75"], 
                            value="61",
                            label="Number of Velocity Channels",
                            show_label=True), 
                gr.Dropdown(["600"], 
                            value="600",
                            label="Image Dimensions",
                            show_label=True), 
                gr.Number(value=0.,
                            label="Input Channel to show (0 = sum over all)",
                            show_label=True), 
                gr.File(label="Input Data", show_label=True)],
        outputs=[gr.Textbox(lines=1, label="Prediction", show_label=True), 
                # gr.Image(label="Input Image", show_label=True), 
                gr.Plot(label="Input Image", show_label=True), 
                gr.Plot(label="Mean-Subtracted Activations", show_label=True), 
                gr.Plot(label="Raw Activations", show_label=True) 
                ],
        title="Kinematic Planet Detector"
    )
    demo.launch()