C3D Feature Extraction with https://github.com/axon-research/c3d-keras

feature_extract.py

import better_exceptions
import os
import numpy as np
from tqdm import tqdm

from keras.models import model_from_json
import c3d_model
import sys
import keras.backend as K
dim_ordering = K.image_dim_ordering()
backend = dim_ordering

ROOT_DIR = 'DIR/TO/CONTAINING/VIDEO/FILES/'
EXTENSIONS = ('mp4','avi')
VIDEO_DIR = ROOT_DIR
FV_DIR = os.path.join(ROOT_DIR,'c3d_fv')

BATCH_SIZE = 64
NUM_FRAMES = 16
HEIGHT = 128
WIDTH = 171
FC_LAYER = 'fc6' #'fc7'

video_files = sorted([os.path.join(VIDEO_DIR,f) for f in os.listdir(VIDEO_DIR) if f.endswith(".avi")])

# Helper functions.
import subprocess
def _read_video(filename,h,w) :
    command = ['ffmpeg',
               '-loglevel', 'error',
               #'-ss', str(time),            #from this time
               '-i', filename,
               #'-vframes', str(1),          #get single frame
               '-vf', 'scale=%d/%d'%(w,h),   #given width and height
               '-f', 'image2pipe',
               '-pix_fmt', 'rgb24',
               '-vcodec', 'rawvideo', '-']
    #print(command)
    ffmpeg = subprocess.Popen(command, stderr=subprocess.PIPE ,stdout = subprocess.PIPE )
    out, err = ffmpeg.communicate()
    if(err) : print(err)
    video = np.fromstring(out, dtype='uint8').reshape((-1,h,w,3))
    return video
  
def _aggregate_up_to(batch_size,video):
    Xs = np.zeros([batch_size,NUM_FRAMES]+list(video.shape[1:]))
    for i,s in enumerate(range(0,len(video)-NUM_FRAMES,NUM_FRAMES)):
        Xs[i%batch_size] = video[s:s+NUM_FRAMES]
        if( (i+1)%batch_size == 0 ):
            yield Xs
    if( (i+1)%batch_size != 0 ):
        yield Xs[:i%batch_size+1]

# >>>>>>>>>>>>> Load Models...
model_weight_filename = os.path.join('./models', 'sports1M_weights_tf.h5')
model_json_filename = os.path.join('./models', 'sports1M_weights_tf.json')

model = model_from_json(open(model_json_filename, 'r').read())
model.load_weights(model_weight_filename)
int_model = c3d_model.get_int_model(model=model, layer=FC_LAYER, backend=backend)

# subtract mean
mean_cube = np.load('models/train01_16_128_171_mean.npy')
mean_cube = np.transpose(mean_cube, (1, 2, 3, 0))
# <<<<<<<<<<<<

# >>>>>>>>>>>> Extract Features!
for _,v in enumerate(tqdm(video_files)):
    video = _read_video(v,HEIGHT,WIDTH)
    fvs_name = os.path.join(FV_DIR,os.path.basename(os.path.splitext(v)[0]+'.npy'))

    fvs = np.zeros((len(video)//NUM_FRAMES,4096))
    for i,(Xs) in enumerate(tqdm(_aggregate_up_to(BATCH_SIZE,video),total=len(video)//NUM_FRAMES//BATCH_SIZE)):
        fv = int_model.predict_on_batch((Xs-mean_cube)[:,:,8:120,30:142,:]) # Crop Center
        fvs[i*BATCH_SIZE:i*BATCH_SIZE+len(fv)] = fv
    np.save(open(fvs_name,'w'),fvs)
# <<<<<<<<<<<<

from where can i get filles : sports1M_weights_tf.h5 , sports1M_weights_tf.json and train01_16_128_171_mean.npy.